Reasonable
Doubts

Fit to Serve

Corinne Alvarez has sixty days to tell her board what AI should mean for a legal-aid organization already stretched beyond capacity.

Reasonable Doubts · 18 chapters · ~2 min each

Chapter One (~3 minutes)

The Sixty Days

The item took four minutes of a ninety-minute meeting. Ellery Pike had circulated an article about a firm in Chicago. Nathan Ross asked what it would cost. Dolores Vance, who had sat in the intake chairs herself, eleven years and one eviction ago, and now sat on the board, asked what it would do for the people who would be out there on Thursday. Nobody answered her. Corinne Alvarez noticed that nobody answered her, and that she had not answered her either.

Harriet Bloom stayed behind while the chairs were being stacked.

"I'm going to ask you for something, and you're going to try to negotiate it," Harriet said. "Sixty days. Bring us a vision for how this office uses artificial intelligence, or declines to, and why."

"I can have you a report in three weeks."

"I don't want a report. A report tells me what exists." Harriet found her coat on the back of a chair. "The case statement for the campaign goes to the printer the week after you present. Whatever you put in front of us decides what we ask donors for and what we promise them, and it holds for three years. Pike will fund a vision. Ross will vote against a hunch. Vance will want to know who it protects. You are the one who has to stand up and say it."

"I don't understand the thing," Corinne said.

"I know. That's what the sixty days are for."

In the parking lot she did the arithmetic she had been avoiding for a year. At home she was an enthusiastic amateur. She used it to plan a week of meals, to argue with a contractor's estimate, to draft her uncle's eulogy and then throw away everything it wrote except one line. At work she had never once opened it, and had never written down why not. Twenty staff. Four hundred and six open files. A director who could not state her own position in a sentence a treasurer would accept.

What made it worse was that the thing would not hold still to be learned. In the eighteen months she had spent not deciding, the tools had gone from a box you typed one question into, to assistants that carried a conversation across weeks, to systems that searched, ran code, read a PDF and a recording in the same breath, and took steps on their own behalf.1 Whatever she carried into that boardroom would be, on the morning she carried it, already a little out of date.

And the rules had moved under her too, which she had somehow expected less. In a few short years, competence in these tools had gone from something a lawyer might sensibly acquire to something a lawyer was presumed to have, written into formal opinions and state guidance that took no interest in whether an office was funded.2 Her staff attorneys carried that duty one by one. She carried it for all of them at once, and for the volunteers, and for the intake supervisor who was not a lawyer at all.

She tried to say to herself what understanding would actually require, and found she could get most of the way. Not a syllabus. Knowing each tool's capabilities and its limits well enough to say which tasks were safe to hand it, which needed checking, and how hard.3 That was a sentence she could defend. She had no idea yet whether it was a sentence she could deliver.

At home there was mail. Simone Adeyemi's foundation had added a line to the renewal form: Describe your organization's approach to artificial intelligence, including safeguards. Not whether. How. The blank was four inches long and worth two hundred and sixty thousand dollars a year, and whatever she wrote in it, whether enthusiasm or refusal or an honest account of a governed middle, would be read by strangers as a statement about whether this office was still fit to be trusted with the work.4

She did not feel excited. That was the part she kept turning over. She felt the way she felt about the answering machine, which showed forty-one messages when she let herself back in at seven the next morning: not inspired, just responsible.

Plate One

Topics in the chapter

1The Evolution of AI SystemsModule 3.1 · Topic 4
2Evolving Standards of CareModule 7.1 · Topic 3
3AI Literacy for Legal ProfessionalsModule 3.1 · Topic 5
4Building Strategic PositionModule 7.3 · Topic 4

Further reading

1Change Management and Team UpskillingModule 6.3 · Topic 4
2Risk Management StrategiesModule 7.1 · Topic 5
3Long-Term Implementation and PlanningModule 7.2 · Topic 4
4Advising on AI Compliance in Legal ContextsModule 6.3 · Topic 5
5What Is Generative AI?Module 3.1 · Topic 1
6Staying Current for CompetenceModule 7.3 · Topic 3

Chapter Two (~2 minutes)

The Unmet Need

On the fourth day she sat behind the intake desk for a full morning, which she had not done in three years, because it seemed dishonest to ask what a machine could do before she could say plainly what the office could not.

Dev Ramachandran gave her the phone log without comment. Eleven hundred and eighty calls in March. Four hundred and two returned. Ninety-six households screened, sixty-one files opened. Median wait for a housing screening, twenty-six days.

"What happens to the other seven hundred and seventy-eight?" she asked.

"They call somebody else, or they don't," Dev said. "Mostly they don't. We don't have a category for them, so they aren't in any number you'll show the board."

At ten past ten a woman named Alma Pryor came to the window with a summons in a grocery bag and an answer due Friday. Dev offered her the first screening slot, which was three weeks out, and a self-help packet in nine-point type that opened with the words responsive pleading. Corinne watched her read the first line twice and put it in the bag with the summons.

That was the morning's education, and it had nothing to do with technology. The office's failure was not that its lawyers were slow. It was that between a person at the window and a lawyer's judgment stood a great deal of typing.

She started listing the typing. Somebody read the grocery bag: sixty pages of leases, notices, money orders, a repair log kept on the backs of envelopes, and county printouts scanned so badly that half the dates were guesses. Somebody keyed the parties and the dates into the case system. Somebody wrote the denial letters, four hundred a year, and somebody, usually Priscilla, spent forty minutes turning one of them into English a person could act on, which meant most of them went out in the language of the statute instead.1

Dev walked her through what he thought the machine could take. Reading the bag and pulling out the dates and the parties, so the human read a summary and checked it against the paper instead of building it from nothing.2 Classifying what arrived and routing it: housing here, benefits there, anything unclear into a review queue rather than a guess.3 Assembling a matter, tracking the deadlines across four hundred files, drafting the update letters, doing a first pass at research that a lawyer would then check line by line.4

"And the gates," Corinne said. "Every one of those has to be checked by somebody."

"Every one." Dev did not look up. "You can automate the typing. You cannot automate Priscilla."

That was the sentence she wrote on the back of the phone log, because it named both halves at once. The gain was real and it was large: if the reading and the keying and the letters ran faster and truer, the twenty-six days could come down, and Alma Pryor's Friday might have a lawyer in it. And every hour of that gain was purchased with a review that only the scarcest people in the building could perform.

She did not write down what AI could do. She wrote down who the office had failed in March, by name where she had one, and she made Dev help her put a number on the ones there was no name for. Seven hundred and seventy-eight households. It was the first page of anything she had that Dolores Vance would recognize as an answer.

Plate Two

Topics in the chapter

1Communication & ExpressionModule 3.2 · Topic 2
2Gaining Document Processing PowerModule 1.1 · Topic 2
3Legal Workflow AutomationModule 4.2 · Topic 5
4Legal Practice ApplicationsModule 3.2 · Topic 5

Further reading

1Workflow Automation PatternsModule 4.2 · Topic 2
2Creation & IdeationModule 3.2 · Topic 1
3Analysis & EvaluationModule 3.2 · Topic 3
4Multimodal AI CapabilitiesModule 3.3 · Topic 2
5Tool Ecosystems and IntegrationsModule 4.1 · Topic 2
6Gaining Confidence Through First InteractionsModule 1.1 · Topic 1

Chapter Three (~2 minutes)

Two on the Payroll

She found the first one by accident, on the ninth day, because the monthly expense reconciliation had started arriving on Tuesday instead of Friday.

"You changed something," Corinne said.

Yolanda Reyes turned her monitor around with the frank pleasure of a person who has been waiting eight months for someone to ask. At home, she said, she had described what she wanted in plain sentences and let the machine write the code, then tested it, told it what broke, and let it fix that, three rounds of it, no programming, a working thing by Sunday.1 The receipts came into a mailbox and went out to a spreadsheet through a chain of blocks she had wired together on screen, with a step that retried when the bank's site was down.2 And she kept a standing assistant with fixed written instructions that formatted her mother's medical appointment letters the same way every single time, because her mother's clinic changed the form every quarter and Yolanda had stopped caring.3 What she had learned in the first tool she had carried whole into the second and the third; the skill, she said, was not the tool.4

"How much of this have you used here?"

"None of it."

"Why not?"

Yolanda looked at her as though the question were slightly cruel. "Because nobody ever said I could."

Marcus Feld said something close to the same thing that afternoon, in different clothes. Six weeks into the job, out of a four-hundred-person insurance brokerage, he had a binder: a tiered list of approved tools, low to high; data guardrails naming what could never go into an outside system; mandatory training with a signed attestation before anyone got the higher tiers; a one-page approval form that became the record of due diligence if anything went wrong.5 It was, Corinne thought, a genuinely good binder. It had been written for a workplace with no privilege, no duty of competence, and customers who could absorb a mistake.

She tested it the only way she could think of. She described Alma Pryor's grocery bag and asked what each of them would do.

"Scan it, upload it, ask for a timeline and the deadlines," Yolanda said promptly. "Ten minutes."

"Whose file is that?"

"Ours."

"It's hers," Corinne said. "And most of what's in that bag she can never take back once it leaves this building."

Yolanda absorbed it without defensiveness, which Corinne liked, and then asked what a privileged file was, which she asked seriously, and which nobody in the building had ever taught her, because nobody had ever imagined she would need to know.

So the ninth day gave her two things and took one away. It gave her real capability, already on the payroll, already paid for, sitting idle nine feet from the accounting printer. It gave her a rollout playbook by someone who had actually done a rollout. And it took away the comfortable idea that her problem was money, because the thing standing between these two people and the queue was not a budget line. It was a document she had never written.

The cost surfaced when she asked Marcus what the brokerage's version had taken. A program manager, full-time, six months. A security review he had not had to pay for out of his own budget. "And we were allowed to be wrong for a quarter," he said. "That was the part I didn't notice at the time."

Plate Three

Topics in the chapter

1Vibe Coding for Non-DevelopersModule 5.2 · Topic 2
2No-Code Workflow BuildingModule 6.2 · Topic 4
3Building Custom AssistantsModule 1.2 · Topic 5
4Expanding Your AI ToolkitModule 1.2 · Topic 2
5Organizational AI PoliciesModule 6.3 · Topic 2

Further reading

1The Text-to-Action ParadigmModule 5.2 · Topic 1
2Gaining Professional Application SkillsModule 1.2 · Topic 4
3Scaling AI Across TeamsModule 4.2 · Topic 4
4Foundational Prompt CraftModule 2.1 · Topic 1
5Change Management and Team UpskillingModule 6.3 · Topic 4

Chapter Four (~2 minutes)

Judgment Without a Keyboard

Priscilla Oyelaran had sixty-one open files and a reputation, entirely earned, for being the last person in the building who would sign something she had not read twice. On the twelfth day Corinne sat down across from her and asked her to explain her objection.

"I don't have an objection," Priscilla said. "I have a list of things I would have to be able to check, and I can't check any of them, because I can't work the tools."

"That's not a small admission."

"It's not a small problem." She had never got past a sign-in screen. She called all of them the chatbot. But when Corinne asked what she thought the danger actually was, the answer came out in the shape of a closing argument. "Hand it a lease and ask what's in it and it is reading; everything it needs is on the page in front of it. Ask it what our notice period is and it isn't reading, it's remembering, and it hasn't got anything to remember with. It fills the gap the same way it fills everything else, with whatever fits the pattern, and it says it in exactly the voice it uses when it's right".1

Corinne wrote that down verbatim. It was a better statement of the failure mode than anything she had managed for herself.

"Then there's which one you point at the question," Priscilla went on. "Marcus showed me four of them last week. One is quick and cheap. One takes longer and shows you its working. One goes and looks things up and hands you the sources. Choosing wrong is not a technology decision, it's a competence decision, and it lands on whoever signs the letter".2

"Who chooses?"

"That's my question. It won't be me, I can't tell them apart. And it will be my name on the answer."

They spent an hour on what checking would have to look like, and Priscilla, who had never used one of these things, described the discipline better than the binder did: not trust or refusal, but rigor scaled to what a mistake would cost: a glance at low stakes, a full read at middle, and for anything that left the building with the office's name on it, a second pair of eyes and a written record of what had been confirmed.3 She was already doing it, she pointed out, for the volunteers.

"And the rules haven't changed," she added, "which is the part people keep hoping about. If I hand work to somebody I supervise, the work is still mine. The duty doesn't move because the assistant is software instead of a law student".4

Walking back to her office, Corinne understood the shape of her sixty days for the first time. The capability was in accounting and human resources, held by two people who did not know what a privileged file was. The judgment was in this office, held by a woman who could not open the application. Neither half could be hired; she could not afford either. The vision would have to marry them, and the marriage had a price she now had to say out loud: every gate she built spent Priscilla, and Priscilla was the scarcest thing in the building.

"If we do this," Priscilla said, when Corinne came back to say it, "how many of my Thursdays does it take?"

"I don't know yet."

"Find out before you promise anybody anything."

Plate Four

Topics in the chapter

1Why AI Gets Things WrongModule 3.1 · Topic 3
2Model Selection for Legal Research and AnalysisModule 2.3 · Topic 5
3Trust Calibration in PracticeModule 4.3 · Topic 4
4Professional Responsibility for Agent-Assisted WorkModule 5.3 · Topic 5

Further reading

1Prompting for Legal Analysis and ResearchModule 2.1 · Topic 5
2Evolving Standards of CareModule 7.1 · Topic 3
3Understanding AI Reliability & ErrorsModule 4.3 · Topic 1
4Gaining Understanding of AI BoundariesModule 1.1 · Topic 3
5AI Literacy for Legal ProfessionalsModule 3.1 · Topic 5
6Protecting Quality and Avoiding ErrorsModule 1.2 · Topic 3

Chapter Five (~2 minutes)

Shadow Work

It came out on a Tuesday, sixteen days in, and it came out because Dev Ramachandran chose to say it rather than because anybody caught him.

"Before you write whatever you're writing," he said, standing in her doorway with his coat already on, "you should know I've been using one. Since about May."

"Using it how?"

"Intake notes. I talk to somebody for forty minutes and I have six lines to put in the system before the next call, so I paste the note in and ask it to compress it. And the letters. The ones where we tell people no." He said it evenly. "I never thought of it as a decision. It was Tuesday afternoon and there were nine people in the queue."

Corinne's first feeling was not anger. It was recognition, which was worse. She had spent two weeks assembling a case for acting, and here was the office already acting, without a policy, without a tool anyone had vetted, without a single record of what had gone where.

They pulled the free account's terms up on his monitor. The consumer tier retained what was typed into it, used it to improve the system, and had changed its wording twice since May. Names, addresses, dates of birth, case numbers, a protective order, one client's immigration status, the ages of children, the whole substance of an intake, keyed in by a man who had never been told there was a difference between a tool that kept your words and one that promised not to, and never taught the discipline of stripping a name to an initial before it left his hands.1

She made herself name the exposure precisely rather than dramatically. Client information had gone to an outside service without protection, which was the plainest kind of breach; some of it was privileged, and privilege is not a thing an office holds on a client's behalf and can spend. Worse was the arithmetic: the same habit, the same tool, on hundreds of matters, so any single problem was not one problem but the same problem multiplied by every file it touched, and the only way to know the count was to look backwards through eleven months.2

"How many?" she asked.

"Three hundred, give or take."

They built the response that night, because there was a shape for it and Marcus knew the shape. Stop the practice. Classify what had happened by how badly a person could be hurt by it, which put the protective-order file and the immigration file in a different tier from a summarized housing note. Assess the impact on the people actually affected, not on the organization. Decide who had to be told.3 Then the part that mattered more than any of it: an amnesty, announced to all twenty staff and every volunteer, seven days to say what you had been doing with no consequence, and a written rule that bound everyone the day it closed.4

Four people came forward in the seven days. One was a volunteer whose disclosure would matter enormously later.

The audit cost two weeks of the sixty. Priscilla and Marcus split three hundred intake notes between them, and the housing screenings slipped by nine days, which meant real people waited longer because of a thing Corinne had not written.

"I keep wanting to apologize," Dev said.

"Don't. You had an unmanageable queue and no rule. That's mine." She was writing as she said it. "I told myself I hadn't decided anything yet. I had. I decided we'd do this without governing it, and that turns out to be the worst of the available answers."

Plate Five

Topics in the chapter

1Confidentiality and Data Governance in AI ContextModule 2.2 · Topic 5
2The Liability LandscapeModule 7.1 · Topic 1
3Risk Classification and Impact AssessmentModule 6.3 · Topic 3
4Organizational AI PoliciesModule 6.3 · Topic 2

Further reading

1Model Access and DeploymentModule 3.3 · Topic 4
2Coverage Gaps and Emerging ExclusionsModule 7.2 · Topic 2
3Tools and Data ConnectionsModule 6.2 · Topic 3
4Risk Management StrategiesModule 7.1 · Topic 5
5Gaining Professional Application SkillsModule 1.2 · Topic 4
6Managing Risk for ProtectionModule 7.3 · Topic 2

Chapter Six (~2 minutes)

The Crash Course

They met Thursday evenings in the conference room with the bad chairs: Corinne, Yolanda, Marcus, Priscilla, and a bar-association course that nobody had opened in the two years the office had been paying for it.

The course had a temperament, which surprised her. It refused to be read about. You were meant to learn the thing by working an assistant while it explained itself, treating it as a training partner rather than a work tool, and to move through the material in whatever order your ignorance demanded rather than front to back.1 Priscilla, who had come prepared to take notes, was made to type within eleven minutes.

The first evening went to what the thing actually was. Not a database and not a program in the sense any of them meant it: a system that produces text by predicting what plausibly comes next, trained on an enormous quantity of writing, and therefore probabilistic where ordinary software is fixed, so the same question can yield two different answers and neither is a malfunction.2 Marcus, who had rolled one of these out to four hundred people, admitted he had not known that. Yolanda had known it and had never once said it out loud.

The second evening opened the hood as far as any of them needed. Words become tokens, tokens become numbers, the numbers move through layers of arithmetic that find relationships, and at the end the system produces a distribution over what might come next and picks from it, one piece at a time, then feeds its own choice back in and does it again.3

"So when it doesn't know," Priscilla said slowly, "it doesn't stop."

"It can't stop," Yolanda said. "There's no place in there where stopping lives."

Priscilla asked it a question about a subsection of their own state's landlord-tenant code. It answered in three confident paragraphs, with a subsection number that did not exist, and when she asked it to try again it produced a different number with the same composure. It was, everyone agreed afterward, the most useful four minutes of the six weeks: a system that fabricates where it is thinnest, in precisely the register it uses when it is right, and does so most reliably on the questions that require it to recall rather than to read.4

Yolanda's turn came on the third evening, and it changed Priscilla's posture more than any warning had. She showed them the hidden instructions sitting underneath every exchange, a standing set of rules the operator writes and the user never sees, arranged in a chain of command that the system will not break even when politely asked to. She asked it to skip a house rule just this once; it declined, gracefully, and then obeyed the rule anyway.5

"It's a constrained thing," Priscilla said. "Not an open field."

"You can map it," Yolanda said. "That's all I've been doing for two years."

The argument arrived on the fourth evening, on schedule. Marcus wanted a pilot by Friday: pick a low-risk task, put fifteen people on it, learn by doing what no amount of Thursday would teach. Priscilla wanted every failure mode written down first, because the fifteen people were carrying other people's housing. Yolanda was right about the mechanics and wrong about the stakes, and said so herself before anyone had to.

Corinne went home that night and understood that the first draft of her vision was not a slide. It was this argument, held honestly, with all three positions still standing, which was going to be a difficult thing to present to a room that wanted a picture.

The evenings cost twelve staff-hours a week for six weeks. She converted it, because Ross would: about nineteen screenings not done.

Plate Six

Topics in the chapter

1The LawQi ApproachModule 1.0 · Topic 1
2What Is Generative AI?Module 3.1 · Topic 1
3How Generative AI Actually WorksModule 3.1 · Topic 2
4Why AI Gets Things WrongModule 3.1 · Topic 3
5Gaining Understanding of AI BoundariesModule 1.1 · Topic 3

Further reading

1What Context Is and Why It MattersModule 2.2 · Topic 1
2Understanding Context WindowsModule 2.2 · Topic 2
3The Prompt Stack and HierarchiesModule 4.1 · Topic 1
4System Prompts and Instruction LayersModule 6.1 · Topic 2
5Structured and Multi-Turn PromptingModule 2.1 · Topic 2

Chapter Seven (~2 minutes)

What an Upload Waives

On the twenty-fourth day Priscilla did something she had refused to do for three years, which was teach.

She chose a file and put it on the conference room screen with the name blacked out: a woman whose address was confidential by court order, whose employer appeared in an exhibit, and who was one bad week, a missed shift or a landlord's phone call, from losing the apartment and the order both.

"Show me what you'd do," she said to Yolanda.

Yolanda had the pages scanned in under two minutes and asked for a chronology, and it was, Corinne had to admit, extremely good: dates in order, the notice and the response side by side, the two figures that did not reconcile flagged without being asked. Then Yolanda closed the file and asked the same question with nothing attached, and the answer went instantly vague and general, which was the point of the demonstration.1 What it knew was what you handed it. Nothing more, and nothing less.

"Now," Priscilla said. "Where did the pages go?"

Nobody could answer, and that was the lesson. They spent the next hour on it: whether the service kept what was typed into it and for how long, whether it used what it kept to improve itself, who at the vendor could read it, which other companies it passed through on the way. The redaction discipline Priscilla had been practicing on paper for twenty years turned out to translate exactly: initials for names, a range instead of an exact figure, identifiers proxied or left out entirely, with one hard limit she made them all repeat: redaction is a thing you do before, because after the pages have gone, there is nothing left to cure.2

"And the privilege isn't ours to spend," she said. "It's hers. She's not in the room. She will never know the room existed."

Marcus asked the practical question, which was whether any of this was fixable, and the answer was that most of it was a matter of which door you used rather than whether you used the thing at all. The same system could be reached through a public website that kept everything; through an organizational account with retention switched off and a contract behind it; through an application on a machine; through a browser extension quietly reading whatever tab was open, which was the one that made Priscilla put down her pen; or, for work that justified it, on hardware where nothing left the building at all.3

"So the question isn't whether we use it," Corinne said. "It's which channel, on which file, by whom."

"That's four questions," Priscilla said, "and we haven't answered one of them in writing."

Yolanda was quiet for a while and then said the thing that Corinne kept for the board: "I've been thinking about it as my file. It's not my file."

They agreed the sequence: learn on invented facts and public material until the habits were automatic, then move to real matters only through a vetted channel, and verify everything that came out either way, because the two duties were separate and both applied.4

The cost went straight onto the sheet. Organizational seats with retention off, at thirty dollars a person, came to just under eight thousand dollars a year for a staff of twenty-two, more than the office spent on postage and less than it spent on the copier. Keeping the most exposed work on hardware in the building would take a server and someone to run it, which was a person, which was a number she could not make fit. And redaction ran about four minutes a document across four hundred files, every year, forever.

Plate Seven

Topics in the chapter

1Gaining Document Processing PowerModule 1.1 · Topic 2
2Confidentiality and Data Governance in AI ContextModule 2.2 · Topic 5
3Model Access and DeploymentModule 3.3 · Topic 4
4Gaining Professional Application SkillsModule 1.2 · Topic 4

Further reading

1The Liability LandscapeModule 7.1 · Topic 1
2Evaluating AI Tool Stacks in Legal EnvironmentsModule 6.1 · Topic 5
3Designing Effective ContextModule 2.2 · Topic 3
4Tools and Data ConnectionsModule 6.2 · Topic 3
5Legal Tool EvaluationModule 3.3 · Topic 5
6Memory, Persistence, and ContinuityModule 2.2 · Topic 4

Chapter Eight (~2 minutes)

One Eviction From the Street

Alma Pryor came back. She had lost the Friday, been served again in June, and this time the screening slot fell on the right side of the calendar, and she was assigned to Hollis Pratt, a retired tax lawyer who gave the office two mornings a week and had never once been late.

He had come forward during the amnesty, which is why they knew. He used one at home, on his own laptop, to get himself oriented in areas that were not his.

On the twenty-ninth day, at twenty to seven in the evening, Priscilla read his draft answer because she read everything that left the building with the office's name on it. She read it twice and then walked down the hall with it in her hand and her reading glasses still on.

"He's got a repair-and-deduct provision in here that isn't ours," she said. "Section number, case name, the lot. It's from a different state."

"Fabricated?"

"The section exists somewhere. Not here. And the case doesn't exist anywhere." She turned the page. "That's not what nearly did it, though. Look at what isn't in it."

The demand for a jury had a five-day window from service and the draft did not mention it. That was the true near-miss: not the invented rule, which was loud, but the silence, which was quiet. The system had produced something fluent, complete-looking, and confident, with nothing in its construction that could tell it what it had failed to say.1

Hollis was mortified in the way of a careful man who has been careless once. Corinne made a point of sitting with him rather than sending a memo.

"I asked it a fair question," he said.

"You asked it a question with no state in it." Priscilla had brought a legal pad. She could not run the tools, but she could write the question, and she wrote it in front of him: name the state, name the statute, name the court; ask for holdings out of the judgment text rather than a summary; require a source for every proposition; and tell it that saying I am not sure is a better answer than a confident guess.2

Then she made him verify it the way she verified a junior's memo. Does the authority exist. Does it say what the draft claims. Is it still good law. Does it actually reach the point it is cited for. Each one written down, with the date, in the file.3 And the habit underneath all of it, which cost nothing: ask it what it used, ask it what it could not check, ask how sure it is, and then do not take its answer about its own reliability as reliable either.4

What they built out of the evening was a scale rather than a rule, because a rule that says check everything is a rule that gets ignored by Thursday. A quick read for anything internal. A detailed check where the mistake would cost a client rework or worry. And for anything that went to a court or told a person what to do about their housing, a lawyer's full review, timed and recorded, with no exception for a deadline.5 Priscilla contributed the tells: a citation with no year or court, a number too precise to be real, a paragraph that is confident and says nothing.

Alma Pryor kept her apartment. Corinne wrote the address on the sheet, because Vance would ask and because she wanted to stop saying a wrong answer is costly and start saying what it costs.

"The guardrail worked," Marcus said the next morning, meaning it kindly.

"The guardrail is a person," Corinne said, "reading at twenty to seven at night, unpaid past five. If I put that in the vision as a control, I'm lying. If I put her in as a budget line, nobody funds it. That's the whole problem in one sentence."

Plate Eight

Topics in the chapter

1Understanding AI Reliability & ErrorsModule 4.3 · Topic 1
2Prompting for Legal Analysis and ResearchModule 2.1 · Topic 5
3Legal Verification StandardsModule 4.3 · Topic 5
4Protecting Quality and Avoiding ErrorsModule 1.2 · Topic 3
5Verification and Quality ControlModule 4.1 · Topic 4

Further reading

1Court Cases & SanctionsModule 7.1 · Topic 2
2Verification FrameworksModule 4.3 · Topic 2
3Multi-Model and Consensus StrategiesModule 4.2 · Topic 3
4Quality and Risk in AI-Generated ArtifactsModule 5.2 · Topic 4
5Intermediate AI for Legal PracticeModule 4.1 · Topic 5

Chapter Nine (~2 minutes)

The Flagship

Gordon Trask's conference room was on the thirty-first floor and the coffee came in a pot somebody else carried in. On the thirty-third day he slid a single sheet across the table and let her read it before he said anything, which was courteous of him.

Four hundred thousand dollars over two years. Kiosks in the courthouse lobby and a public website where an unrepresented tenant types a question and gets an answer. The Trask & Whitlow AI Legal Clinic. He would bring Simone Adeyemi's foundation in for a match, and Ellery Pike would move heaven to be photographed at the ribbon.

Corinne wanted it. That was worth being honest about. It was six years of eviction-defense night clinics from a man who had never once asked for a plaque then, and it was more unrestricted-adjacent money than the campaign had raised in either of the last two years.

"Say the objection," Trask said. "You've got your objection face."

"It's not an objection yet. Let me classify it in front of you." She turned the sheet over and drew three boxes. A tool that informs somebody's work, easily overridden, was one thing. A tool that shapes operations was another. But a machine that tells an unrepresented person what to do about the roof over their head touches that person's rights directly and is very hard to override, because there is no lawyer standing between the screen and the decision. That tier does not get built without an impact assessment, testing on the people who would actually use it rather than on people like the two of them, a human override, an audit trail, and a route for someone to say this was wrong about me.1

"That sounds like a lot of paper for a kiosk."

"It's a lot of paper because of whose name is on the answer." She said it carefully. "If that machine tells a tenant she has ten days and she has five, the liability doesn't attach to the machine, and it doesn't attach to the vendor. It attaches to whoever delivered the work. That would be us. And it would be your name over the door while we did it".2

Trask sat back. "You're the first person in this process to tell me that."

"Then here's the rest. Whatever we build, we say plainly what it is, in writing, to every person who touches it, saying what the machine did and what a human did and where the human step is, and we hold ourselves to the strictest standard anybody imposes rather than the loosest one we could get away with, because the moment a tenant discovers we were coy about it, we have spent something the office can't rebuy".3

He was quiet, and then he asked what she would do with the money instead, which was the question she had come for.

The counter-offer fit on a napkin. The public thing does not answer; it routes and it translates. It takes what a person types in their own words, tells them which door and which deadline category applies, in their language and at a reading level a frightened person can hold, and every single path ends with a human being. Two hundred and sixty thousand for that. One hundred and forty for the unglamorous half: the policy, the vetted channel, the training, the attestations, and the attorney hours that make the checking real. And his name goes on the second half, not the first.

"You're asking me to fund plumbing and call it a flagship."

"I'm asking you to put your name on the part that's true. It'll differentiate this office more than the kiosk will, and the funders are already asking which of us can show our governance, not which of us has the shiniest lobby".4

He said he would think about it, which from Trask meant something between a yes and a coin-flip, and Corinne rode down thirty-one floors knowing she might come back to Harriet Bloom on the sixtieth day with a smaller number and no ribbon.

Plate Nine

Topics in the chapter

1Risk Classification and Impact AssessmentModule 6.3 · Topic 3
2Intellectual Property and Liability in AI-Generated ContentModule 5.2 · Topic 5
3Maintaining Transparency for TrustModule 7.3 · Topic 1
4Building Strategic PositionModule 7.3 · Topic 4

Further reading

1Legal Practice ApplicationsModule 3.2 · Topic 5
2Disclosure RequirementsModule 7.1 · Topic 4
3Multimodal Creation WorkflowsModule 5.2 · Topic 3
4Legal Workflow AutomationModule 4.2 · Topic 5
5Research & StrategyModule 3.2 · Topic 4
6Professional Liability Insurance in the AI EraModule 7.2 · Topic 1

Chapter Ten (~2 minutes)

What No One Will Fund

Simone Adeyemi took the meeting at a bakery two blocks from the foundation, because the foundation's rooms were full of people waiting to ask for money and she did not want Corinne sitting in that particular queue. It was a kindness. Corinne had become, over thirty-one days, a woman who noticed queues everywhere.

The ask had been rewritten four times and had five lines on it. A triage pilot on the housing docket. A written policy. A half-time person to run oversight. Training for twenty staff and sixty volunteers. A quarterly review.

Simone read it twice, which was more than most funders did. "I can fund line one," she said. "Generously, probably. I can't fund two through five."

"Two through five are what make line one safe."

"I know. My board funds what changes a client's life. They don't fund what protects an institution; they've been told for thirty years that's overhead, and they believe it." She turned the page over, wrote evaluation on the back, and slid it across. "I can fund evaluation. Build me a real evaluation component and some of your oversight can live inside it. That's the trick everyone uses and I'm not proud of teaching it to you."

Corinne took the paper. Then she said the thing she had been carrying since Tuesday, when the broker's renewal packet arrived. "Here's my problem. The unfundable half isn't a nicety. It's the condition of being insured at all."

She had read the packet twice with a highlighter, the way she read leases. Their policy still covered errors in the delivery of legal services, and an error made with a machine's help was, in principle, an error like any other.1 But the renewal came with three pages that had not been there the year before, and the three pages were specific: no coverage for claims arising from AI use not disclosed to the client, none for errors from output nobody independently verified, and a carve-out for fabrications that read as though somebody had been paying attention to the sanctions docket.2 The questionnaire asked for her written policy. It asked for training records. It asked, in a small box that took Corinne twenty minutes to stop staring at, whether the office maintained a log of AI errors caught before delivery.

"They don't care what I know," she said. "They care what I do, and whether I wrote it down. Policy, verification procedures, training attendance, a map of what client data goes where, an incident log".3

Simone was quiet. "That is a genuinely good argument," she said, "and it will not move my board, because it's an argument about your survival rather than your clients'."

"It's the same argument. If we lose coverage, there's no clinic to fund."

"Then make it the same argument on the page." Simone tapped the sheet. "And make it proportionate. If you come to me saying every use of this technology needs the full apparatus, I'll believe you're gold-plating. If you come to me with tiers, so that this touches a benefits determination and gets an impact assessment and a human override and an audit trail, while this drafts a volunteer newsletter and gets a spot check, then I can defend you".4

Corinne walked back with two things she had not had at breakfast: a way to smuggle a quarter of the governance into an evaluation line, and the certainty that the other three quarters would have to come out of the general fund, which meant out of case handling, which meant out of the queue. Twenty-nine days left. The vision now had to carry a cost she could not raise, and she was not going to pretend otherwise in front of the board.

Plate Ten

Topics in the chapter

1Professional Liability Insurance in the AI EraModule 7.2 · Topic 1
2Coverage Gaps and Emerging ExclusionsModule 7.2 · Topic 2
3Best Practices for Maintaining CoverageModule 7.2 · Topic 3
4Risk Classification and Impact AssessmentModule 6.3 · Topic 3

Further reading

1Risk Management StrategiesModule 7.1 · Topic 5
2No-Code Workflow BuildingModule 6.2 · Topic 4
3Workflow Automation PatternsModule 4.2 · Topic 2
4Managing Risk for ProtectionModule 7.3 · Topic 2
5Legal-Specific AI GovernanceModule 6.2 · Topic 5
6Long-Term Implementation and PlanningModule 7.2 · Topic 4

Chapter Eleven (~2 minutes)

The Scar Tissue

Ray Okonjo's office was two states and four hours away, in a converted bank with a vault they used for closed files. He gave Corinne the numbers first, because he said everybody wanted the numbers first and it was better to get the seduction over with.

Callbacks within forty-eight hours: thirty-one percent before, seventy-eight percent after. Applications completed at first contact: up by half. For the first time in nine years, his waiting list had gone down in a quarter and stayed down the next.

"Now the rest," he said.

The rest took two hours. He had put the machinery where the work was repetitive and the stakes were survivable: organizing case files, indexing discovery by custodian and date, running database searches, assembling standard-form documents from templates.1 Every authority it surfaced was pulled and read by a human before anyone cited it, a rule he had written after the third month and should have written before the first.

"Where did it hurt?"

"Four places." He counted them off. The system that over-tagged, filing a document under six issues instead of two, which nobody minded, until a paralegal started trusting the tags and stopped opening the documents. The assembly workflow that produced a flawless motion on the wrong county's form for eleven days. The compliance monitor that flagged every deadline change in the state, so his staff stopped reading its alerts inside a month, which meant the one alert that mattered went by unread. And a supervising attorney who quit, not over an error but over the feeling that she was now signing her name to work she had not watched being done.

"The failures were boring, is the thing," Ray said. "Confident wrong numbers. Eight of ten items when you asked for ten. Prose when you asked for a table. Once you can name the pattern you stop arguing with the machine and start diagnosing it: fix the instruction, or fix what you fed it, or admit the task was never a fit".2 He had rebuilt the workflows to absorb the failures rather than prevent them: gates before anything moved downstream, staging so a bad run could be rolled back, and a rule that the system stop and ask rather than guess.

His hardest problem had nothing to do with tools. "I had one person who was good at it, and for five months every piece of work in the building routed through her. That's not adoption, that's a bottleneck with a nice attitude." What broke it open was writing down what she did: templates for the six things they actually needed, so that people who would never invent a workflow could run a proven one, and a few of them eventually learned to change it.3

"And the woman who quit?"

"That's the part I got wrong and can't get back." He said two-thirds of the trouble in his first year was people, not software: what they feared, what they weren't taught, what nobody bothered to explain. He had run it as a technology project of announce, deploy and measure savings, when it was a change project: a few weeks of telling people plainly what was happening and why, then months of training pitched differently for the person who used the tool, the person who supervised its use, and the person who only needed to recognize when to escalate.4 "And measure from the first day. Not the savings. Who's actually using it, and who's quietly refusing. I started measuring at month seven and by then the refusal was a personality."

Driving home, Corinne understood that she had come looking for a verdict and been handed a map instead. Twenty-five days left. The vision now had to name where an office like hers gets hurt, and say, in advance, who would notice.

Plate Eleven

Topics in the chapter

1Agentic AI in Legal WorkflowsModule 5.1 · Topic 5
2Error Recovery and ResilienceModule 5.3 · Topic 4
3Scaling AI Across TeamsModule 4.2 · Topic 4
4Change Management and Team UpskillingModule 6.3 · Topic 4

Further reading

1What AI Agents AreModule 5.1 · Topic 1
2Trust Calibration for Agent WorkModule 5.3 · Topic 3
3Building Reusable AI AssetsModule 4.2 · Topic 1
4Multi-Agent PatternsModule 6.2 · Topic 2
5The Human-Agent Collaboration ModelModule 5.1 · Topic 4
6Building Quality Evaluation CriteriaModule 4.3 · Topic 3

Chapter Twelve (~2 minutes)

Four Minutes Ahead

Tessa Vaughn arrived early, plugged in her own adapter, and gave a demonstration that was, in every respect that mattered to her, honest. That was the disorienting part. Corinne had braced for a salesperson and got a believer.

The product screened intake, drafted a plain-language summary of the client's problem, pulled the relevant self-help materials, and assembled a standard answer to an eviction complaint. It did all of this on screen in eleven minutes. Priscilla sat with her arms folded. Yolanda leaned forward. Dev watched the intake screen the way a man watches someone else drive his car.

"What's underneath it?" Yolanda asked, before Corinne could. "Is the accuracy piece a database it searches first, or is it an instruction telling it to be careful?"

Tessa said it was their proprietary legal accuracy engine, and Corinne felt the small click of a question landing where it should. Six weeks ago she would have written the phrase down as an answer. Now she knew that a tool of this kind is layers: hidden instructions that shape behavior, a retrieval step that grounds an answer in real documents or doesn't, a routing layer that turns text into an action in somebody's system, filters on the way out. "Proprietary engine" named none of them.1 Whether the citations were real depended entirely on which of those layers existed.

Marcus asked whose model it ran on and what happened when that provider shipped an update. Tessa named a vendor; she did not know the knowledge cutoff and said she would find out. Corinne asked how many pages it could take in one pass, because a lease with exhibits ran sixty and an eviction file could run two hundred.2 That answer was in the deck, in a footnote, and it was smaller than sixty.

Priscilla asked the question that changed the room. "Your demonstration answered in two seconds. What is answering?"

Tessa explained that the live product used the fast tier for responsiveness. Corinne had spent an evening on precisely this: the quick, fluent tier that is cheap and confident and fine for a first draft; the slower tier that deliberates and shows its steps, which is what you want when someone will be held to the answer; the research tier that decomposes a question and gathers across sources, which costs more and takes minutes.3 The demonstration had been built on the tier you would never let near a housing file, and the price sheet assumed it.

Then the four minutes. Corinne had started keeping a tally on her legal pad: a mark each time a capability was described in the present tense and located, on follow-up, in the spring release. Four marks. Redaction before upload. The audit trail. Deletion on request. Integration with their case management system.

"None of those are lies," she said afterward, to a conference room that had emptied of everyone but the four of them. "They're a roadmap read out in the present tense. She believes every word of it. That's why you have to write the questions down before she gets here."

The questions were now a page: where the data physically sat and under whose law; encryption and access control; retention, deletion, and whether anything was used to improve the product; certifications with dates on them, because expired ones are worse than none.4 Priscilla had added her own column, which asked what happens if we leave. Could they export their own matters in a format anything else could read; how much of the office's daily practice would be rebuilt around this one company's way of doing things; and what the switching cost would be in year three.5

Twenty-one days left. The vision had not gained a vendor. It had gained a yardstick, and two of the four things Tessa promised were things the office genuinely needed, fourteen months out, at a price that would eat the pilot.

Plate Twelve

Topics in the chapter

1What Scaffolding Is and Why It MattersModule 6.1 · Topic 1
2Foundation Models and ProvidersModule 3.3 · Topic 1
3The Model SpectrumModule 2.3 · Topic 1
4Legal Tool EvaluationModule 3.3 · Topic 5
5Evaluating AI Tool Stacks in Legal EnvironmentsModule 6.1 · Topic 5

Further reading

1Specialized and Fine-Tuned ModelsModule 3.3 · Topic 3
2Retrieval-Augmented Generation (RAG)Module 4.1 · Topic 3
3How Different Models Process InputsModule 2.3 · Topic 2
4Types of AI AgentsModule 5.1 · Topic 2
5Multi-Model StrategiesModule 2.3 · Topic 4

Chapter Thirteen (~2 minutes)

Who May Press Enter

They wrote it on a Thursday, in the small conference room, on a whiteboard that had held the same fundraising thermometer since March. Marcus brought a template with three tiers on it. Priscilla brought a legal pad and a mood.

"Tiers by tool is how it's done," Marcus said. "Low risk, no approval: summarizing a memo, cleaning up a newsletter. Middle tier, manager approval. Top tier, executive sign-off, anything customer-facing."

"Client," said Priscilla. "And your tiers are attached to the wrong noun. In my building the risk doesn't live in the tool, it lives in the file."

That was the argument that made the policy. What came out of it had tiers, guardrails on what data could touch what, a training requirement with an attestation nobody could skip, and an approval route that generated a piece of paper, because a decision no one recorded was a decision the office could not later prove it had made.1 It also had an amnesty clause, three lines long, that closed the months Dev had spent pasting client details into a free chatbot and opened a channel he was actually permitted to use. He read it twice and did not say anything, which from Dev was a great deal.

Yolanda got a real permission for the first time: the approved tier on funder reports, the volunteer roster, the grant budget, anything with no client in it. She had asked for more. Corinne said not yet, and said why: not because Yolanda was careless, but because Yolanda still could not reliably tell which of the office's documents were privileged, and the policy would not run on anybody's good intentions.

Priscilla wrote the delegation rules herself, and they were shorter and harder than Marcus expected. Nothing went to a machine as "look into this." Each task got a stated objective, the form of the answer, the sources it was allowed to use, and an explicit definition of done, because a vague instruction produced a confident, unusable answer and then somebody's afternoon went into deciding whether to trust it.2 Work of any size got broken into steps with a person between them. "Comprehensive research" was struck out wherever it appeared and replaced with criteria a supervisor could actually check: every authority pulled and read, every date confirmed against the docket.3

Then the harder column: what never leaves a human. Anything requiring judgment. Anything you could not undo: a filing, a service, a notice, a client told they did not qualify. Anything where being wrong cost more than the time it saved.4 Marcus, to his credit, wrote the line about rubber stamps himself: that approving work you have not read is not supervision, it is theater, and it is worse than no supervision because it manufactures a record of care that was never taken.

The connector question took the last hour and belonged to Yolanda, who understood it fastest. If anything ever reached into the case management system, it would reach into exactly the fields it needed and no others: matter type, county, deadline. Not the narrative. Not the client's address. Not billing. Credentials held by one named person, stored properly, never shared; every flow of data written down so that if something went wrong they could trace where it went and who had touched it.5

And nothing at all, without exception, ran on a live matter until it had been tested on closed files, documented, approved by both her and Priscilla, and given a log that recorded what it did and who reviewed it.6

Eleven hours across three evenings, plus Priscilla's entire Thursday, which was three intake appointments that did not happen. Seventeen days left. The vision now had a permission structure, and an owner: Marcus at a fifth of his week, Priscilla at an hour she did not have.

Plate Thirteen

Topics in the chapter

1Organizational AI PoliciesModule 6.3 · Topic 2
2Effective Task DelegationModule 5.3 · Topic 1
3Supervision and Oversight StrategiesModule 5.3 · Topic 2
4The Human-Agent Collaboration ModelModule 5.1 · Topic 4
5Tools and Data ConnectionsModule 6.2 · Topic 3
6Legal-Specific AI GovernanceModule 6.2 · Topic 5

Further reading

1Professional Responsibility for Agent-Assisted WorkModule 5.3 · Topic 5
2Trust Calibration for Agent WorkModule 5.3 · Topic 3
3Connectors and the Model Context ProtocolModule 6.1 · Topic 4
4How Computer Use WorksModule 5.1 · Topic 3

Chapter Fourteen (~2 minutes)

Bringing the House Along

The all-staff session cost twenty people two hours, which was a day of intake, which was somewhere between nine and fourteen families who called and got the machine. Corinne made herself put that number on the first slide, because if the price of learning was going to come out of the queue she was not going to hide it from the people standing in it.

She had planned to lecture. Marcus talked her out of it in the parking lot. What the office needed was not a briefing but a practice session, people using the thing while somebody experienced stood nearby, in a place where a mistake cost nothing, because understanding acquired by contact survives and understanding acquired by slideshow does not.1 So they opened laptops instead. The first exercise was three minutes long: ask it what it is and how it comes to know what it says, then read what came back for what it admitted it could not do.2

Then Corinne did the thing she had been most afraid of, and asked questions with no technology in them at all. How does your phone guess the next word you're typing. What does a person do when they are terrified of saying I don't know. If you sent someone to the store and said only buy food, what comes home. How would you work with an assistant who forgets everything overnight.3

Priscilla, who would not touch the keyboard, answered three of them out loud and was the best in the room. Corinne watched that land on the paralegals. It was the single most useful ninety seconds of the sixty days: the office's most careful lawyer demonstrating that the competence in question was judgment, and that judgment was already in the building.

The fear came in the second hour, from Rosalind at the front desk, who said what everyone under forty was too polished to say: is this what replaces me. Corinne did not reassure her. She said the honest thing, which was that the office had a waiting list, not a labor surplus, and that nothing in the plan reduced a position; that if that ever changed she would say so in a room like this one before it happened, not after. Two people remained visibly angry. She wrote both names down, because the ones who say it out loud are the ones you can still work with.

Marcus laid out the shape of it: a few weeks of telling people plainly what was coming, then training pitched by role: literacy for everyone, hands-on work for the people who would actually use a tool, a different session for supervisors on how to catch drift, and something advanced for the two or three who would build things.4 He wanted six months. Corinne had nineteen days to the vote and eleven months to the end of the first phase, and said so.

The volunteer evening was harder and stranger. Sixty people on a video call, half of them retired attorneys, one of them Dolores Vance, who had been a client of this office in 2009 and now sat on its board. The volunteers wanted permission to use tools they were already using at their day jobs. Corinne gave them a boundary instead of a lecture: nothing on a client matter, ever, and here is why, and here is what you may use for the newsletter and the clinic flyer.

What she took from both rooms was Ray's warning made local. Their capability was concentrated in Yolanda and Marcus, and if she did nothing else, every piece of work in the building would route through two people who were not lawyers. The answer was to write down what they did: templates for the six or seven things the office actually needed, so somebody who would never invent a workflow could run a proven one, and a few would eventually learn to change it.5

Nineteen days. The vision now had a readiness plan, and a cost measured in intake hours she could name to the dollar.

Plate Fourteen

Topics in the chapter

1The LawQi ApproachModule 1.0 · Topic 1
2Gaining Confidence Through First InteractionsModule 1.1 · Topic 1
3Skill DevelopmentModule 1.0 · Topic 3
4Change Management and Team UpskillingModule 6.3 · Topic 4
5Scaling AI Across TeamsModule 4.2 · Topic 4

Further reading

1Core Mechanics & InteractionModule 1.0 · Topic 2
2Staying Current for CompetenceModule 7.3 · Topic 3
3Foundational Prompt CraftModule 2.1 · Topic 1
4Advanced Prompting StrategiesModule 2.1 · Topic 3
5Gaining Customization AdvantagesModule 1.2 · Topic 1

Chapter Fifteen (~2 minutes)

The Vision, With Guardrails

She assembled it on a Saturday at her kitchen table with the good coffee and forty-one days of notes, and discovered that the assembly was mostly subtraction.

She had used the machinery itself to get there, which struck her as either fitting or cheap and she had stopped caring which. She had asked it to map the plan's downside case, to lay out what a skeptical treasurer would attack first, to sketch the trade-offs between doing the housing docket first and doing intake first. It was fast and it was structured and roughly a fifth of what it told her about legal aid funding was wrong in ways she could catch only because she had run this office for nine years.1 She kept the shape and threw out the facts. That, she thought, was the whole relationship in one Saturday.

What survived was three phases across three years and a page of things the office would not do.

The spine was not clever. It was four things that only work together: written rules about who may use what and on which files; training for everyone who touches it; a verification step that catches errors before they reach a client or a court; and deliberate choices about which tools are permitted and how they are configured.2 Any three of the four failed. Rules without training produced a binder nobody read. Training without verification produced well-informed people making undetected mistakes. She wrote that sentence twice and kept the second one.

Underneath sat the tiering that made it affordable. Anything touching a client's eligibility, a filing, or a housing outcome was top tier: assessed before it ran, human override, an audit trail, reassessed every year. Anything operational, whether a funder report or a volunteer schedule or a plain-language summary of a self-help guide, got a spot check and a named owner. Everything got written down somewhere a carrier and a funder could both read it.3 She borrowed the four-word spine that governance frameworks kept converging on, govern and map and measure and manage, not because the office was subject to any of them, but because a broker and a program officer would both recognize the shape, and because she would rather adopt a discipline than invent one.4

The hardest page was quality. Priscilla had insisted that "accurate" was not a standard, and she was right: an internal brainstorm needed only to be conceptually sound, a plain-language client letter needed every fact checked, and anything going to a court or determining whether a family kept a home needed verification against the source, disclosed limits, and a named human signing for it.5 Three thresholds, three intensities of checking, written so they survived Priscilla leaving.

And then the page that Corinne knew would cost her the room's enthusiasm. What the office declined: no generated text in any document filed with a court, in any phase. No client-identifying information in any tool outside the approved channel. Nothing autonomous on an eviction-defense file, at all, in year one, and in year two only behind the verification standard, and only if the first year's incident log said it had earned it. Gordon Trask's clinic, which would have opened in the fall with his name over the door, became phase two, contingent, and smaller.

She sat with that for a while. A vision carrying only the ambition would be reckless with people who could not absorb a single avoidable error. A vision carrying only the caution would abandon the people the office was already failing, which was most of them, which was the whole reason any of this started. Both halves were the same duty wearing different clothes, and if she could get the board to feel that in one evening the rest was arithmetic.

Fourteen days. The vision was smaller than the flagship, steadier, and hers.

Plate Fifteen

Topics in the chapter

1Research & StrategyModule 3.2 · Topic 4
2Risk Management StrategiesModule 7.1 · Topic 5
3Risk Classification and Impact AssessmentModule 6.3 · Topic 3
4Global Governance FrameworksModule 6.3 · Topic 1
5Building Quality Evaluation CriteriaModule 4.3 · Topic 3

Further reading

1Managing Risk for ProtectionModule 7.3 · Topic 2
2From Prompting to Context EngineeringModule 2.1 · Topic 4
3Building Reusable AI AssetsModule 4.2 · Topic 1
4Creating Reusable SkillsModule 6.2 · Topic 1
5Skills, Plugins, and Capability BundlesModule 6.1 · Topic 3

Chapter Sixteen (~2 minutes)

The Board Is Not One Mind

Harriet Bloom's advice, delivered over a bad phone connection, was the most useful sentence of the week. "Don't write a pitch that survives all four of them. Write one that tells each of them the truth. They talk to each other, Corinne. They will compare notes in the parking lot."

So she took four meetings in five days.

Ellery Pike gave her ninety minutes and wanted the flagship. He was generous, he was impatient, and he was not stupid. "You've built a compliance department for a problem you haven't had."

"We've had it," Corinne said. "We just haven't been sued for it yet." She had run her own case through the machinery first, as a reviewer, asking it for the strongest argument against her plan, the gaps, what a skeptic would attack, and then treated every flag it raised as a hypothesis to test on an actual human rather than a finding to accept.1 Ellery raised two objections the machine had predicted and one it had not, which she thought was about the right ratio.

Nathan Ross wanted numbers and got a taxonomy. She walked the treasurer through what could actually go wrong and what each thing was called: an error nobody caught and a client relied on; analysis reaching a client without a lawyer standing behind it; confidential facts leaving the building through a tool nobody vetted; a court's standing order ignored; privilege waived because somebody used the free version of something on a live file.2 Then the part that had kept her up: at a firm, one bad output harms one client. In an office running a hundred and forty housing files off the same template, one systematic error is a hundred and forty families at once, discovered in month four.

Nathan wrote that down. "That's the sentence for the board," he said. "Not the savings. That."

Dolores Vance asked the only question nobody else thought to ask. She had been a client here in 2009, on a benefits overpayment that took eleven months, and she asked it plainly: "When I call, will I know?"

Which is how disclosure stopped being a compliance line and became an intake script. Corinne had already mapped what the courts they appeared in required: two of the three had issued something, one wanted a certification in the filing itself, and a statute in the state next door would reach some of their clients before the year was out.3 She had decided to hold the office to the most demanding of them everywhere rather than sort it court by court, and to say it to clients in plain words at the first call rather than bury it in a retainer nobody reads: here is what we use, here is what a person checks, here is what we will never hand to a machine.4

"Read it to me," Dolores said. Corinne read it. Dolores made her change two words, and the two words were better.

Harriet, last, counted votes out loud like a woman doing inventory. Ellery: yes, unhappily, and he would say so. Nathan: yes if the liability page was in the packet, no if it was in an appendix. Dolores: yes, and she would speak, and her speaking would move two others. The rest would follow Harriet, who would not tell Corinne which way she was going.

"You know what you have?" Harriet said. "Four people who agree the office has to understand this, and four people who would spend the money differently. Don't try to close that. It won't close, and if you paper it over you'll get a mandate that dissolves the first time something goes wrong."

Six days. The vision now had four hearings and one text.

Plate Sixteen

Topics in the chapter

1Analysis & EvaluationModule 3.2 · Topic 3
2The Liability LandscapeModule 7.1 · Topic 1
3Disclosure RequirementsModule 7.1 · Topic 4
4Maintaining Transparency for TrustModule 7.3 · Topic 1

Further reading

1Communication & ExpressionModule 3.2 · Topic 2
2Matching Tasks to ModelsModule 2.3 · Topic 3
3Advising on AI Compliance in Legal ContextsModule 6.3 · Topic 5
4Creation & IdeationModule 3.2 · Topic 1
5Professional Liability Insurance in the AI EraModule 7.2 · Topic 1
6Trust Calibration in PracticeModule 4.3 · Topic 4

Chapter Seventeen (~2 minutes)

The Night Before

At eleven on the fifty-ninth night the building was hers and the radiator, and the queue was still on the wall where she had pinned it in the first week: seven hundred and seventy-eight calls never returned, printed small enough to fit on one sheet.

The doubt arrived on schedule and was not stage fright. That afternoon a provider had shipped something that made a paragraph of her plan sound quaint. The tools no longer forgot you between conversations, they reached out to fetch what they did not know, they read a photograph of a lease as readily as the text of one, and the demonstrations that week showed them working unattended for hours at a stretch.1 Her phasing assumed a technology that would sit still for eleven months. It would not. It had not sat still for the sixty days.

Priscilla came back for her coat at half past eleven and stayed, which she did roughly once a year.

"It'll be obsolete," Corinne said. "Some of it by Thursday."

"Then say that tomorrow."

So she did the arithmetic that survives obsolescence. Three things in the plan did not depend on any particular tool: telling people what the office used, even when nobody required it; matching the checking to the stakes rather than applying one standard everywhere; and writing down what was done, because a practice with no record is indistinguishable from a practice that never happened. Everything else was implementation and would be replaced.2 She moved those three to the front and stopped calling them controls.

Then competence, which she had been treating as a project with an end date and which was not one. She budgeted the hours, an afternoon a month for herself, more for Marcus and Priscilla, an annual refresh for everyone else, and named who tracked what: Priscilla on the courts' standing orders and the bar's guidance, Marcus on tool behavior and vendor changes, Corinne on the statutes and the carrier. Quarterly, on the calendar, with dates.3 Anything less was a plan to be surprised.

She also cut a slide. It had claimed the office would "align with recognized governance standards," which was true in the sense that she had borrowed their spine and false in the sense a board member would hear it. She rewrote it as what it actually was: a discipline she had adopted because it was better than the one she would have invented, and a single page tracking which binding rules might one day reach an office of this size.4

Then she cut the three slides that would have won.

The first was the rendering of the clinic with Trask's name in the stone. The second was Marcus's projection of forty percent faster intake by month nine, which was a real model built on somebody else's office and which she could not defend under a question from Nathan. The third was the sentence about being the first legal aid office in the state to do any of this, which was true, and vain, and had nothing to do with anyone on the wall.

In their place she put the amnesty clause, the incident log, and a page headed What I do not know: whether the pilot would hold at volume, what the tools would be able to do by the time phase two opened, whether ninety-one thousand dollars was the right number or half of it, and whether she was the right person to run this at all.

"That last one will cost you," Priscilla said.

"It'll cost me Ellery." Corinne looked at the sheet on the wall. "It would cost me more to stand up there and pretend."

Priscilla put her coat on properly, for the second time. "For what it's worth," she said from the door, "I still wouldn't let it near my files. And I'll vote for every word of this."

"I know," Corinne said. "That's the version I want."

Plate Seventeen

Topics in the chapter

1The Evolution of AI SystemsModule 3.1 · Topic 4
2Long-Term Implementation and PlanningModule 7.2 · Topic 4
3Staying Current for CompetenceModule 7.3 · Topic 3
4Global Governance FrameworksModule 6.3 · Topic 1

Further reading

1Evolving Standards of CareModule 7.1 · Topic 3
2Advising on AI Compliance in Legal ContextsModule 6.3 · Topic 5
3Building Quality Evaluation CriteriaModule 4.3 · Topic 3
4Research & StrategyModule 3.2 · Topic 4
5Risk Management StrategiesModule 7.1 · Topic 5
6Change Management and Team UpskillingModule 6.3 · Topic 4

Chapter Eighteen (~3 minutes)

The Vote

Eleven board members, a plate of supermarket cookies, and the sixtieth day.

She spoke for nineteen minutes. She began with the wall, the seven hundred and seventy-eight calls never returned and the applicants turned away at the door on a Tuesday in March, because the case for acting had never been a picture of the future and she was not going to dress it as one now. Then the plan: three phases, the tiers, the four supports that only work together, the page of things the office would not do. Then the page headed What I do not know, which she read out slowly, and after which the room was quieter than she had expected.

The resolution adopted a policy before it adopted a program: approved tools by tier, guardrails on what data may touch what, training with an attestation, an approval route that leaves a record, an amnesty for use that predated the rules, and an annual review with a date on it.1 Nothing in the top tier would deploy without an assessment, a human override, and a log, reviewed every year by a person named in the minutes.2 Three years, built on the three habits that would outlast every tool in the plan, which were to say what you use, match the checking to the stakes, and write it down, with a full reassessment scheduled for month thirty.3 And the campaign case, which Corinne had rewritten four times before she believed it: that an office able to show a funder, a court, and a client exactly how it governs its tools is in a stronger position than one that merely uses them, and that this was worth saying out loud because it was true and not because it sounded good.4

Ellery Pike moved to double the first phase and pull the clinic forward a year. It failed, three to eight. He argued for it well.

Nathan Ross moved to hold half the technology allocation in reserve until the first annual audit. It failed, four to seven. Dolores Vance voted for it, and said why: "Because I was on the other end of this phone in 2009, and I would rather you were slow than sorry." Then she said the other thing, which Corinne did not expect and did not forget. "And I would rather you were quick than nothing. The office was nothing for me for four months."

The main motion carried eleven to nothing.

Harriet Bloom did not let it end there. "Both amendments go in the minutes," she said, "with their movers, and the counts. Not withdrawn, not noted. Recorded." Someone objected that a unanimous vote looked better in a campaign packet. "It looks tidier," Harriet said. "In three years I want to be able to read who thought we were too slow and who thought we were too fast, and find out which of them was right. That is not embarrassment. That is a board doing its job."

So the record shows a vision the room aligned on, and two members who would, given the same money, have bet it differently, and their names.

Afterwards Corinne stood in the hallway with the lights half off. Nobody had been converted. Ellery still thought the guardrails were timidity and had said so on the record and would say so again at the spring gala. Priscilla, who had voted nothing because she was not on the board and had stayed anyway, would still not touch a keyboard, and had written half the policy that governed everyone who would. Dev's amnesty had taken effect at the moment of the vote; he was, as of tonight, the first person in the building permitted to do what he had already been doing for months, in a channel that would not waive anybody's privilege.

The queue would be there in the morning. Nine hundred and six, and the phone starting at eight. What was different was small and not nothing: there was now a permission with a name on every line of it, a log with nothing yet in it, and a page listing what the office would not do, which she had come to understand was not the opposite of the plan but the half of it that made the other half defensible.

Sixty days. She turned off the last light and went home.

Plate Eighteen

Topics in the chapter

1Organizational AI PoliciesModule 6.3 · Topic 2
2Risk Classification and Impact AssessmentModule 6.3 · Topic 3
3Long-Term Implementation and PlanningModule 7.2 · Topic 4
4Building Strategic PositionModule 7.3 · Topic 4

Further reading

1Risk Management StrategiesModule 7.1 · Topic 5
2Change Management and Team UpskillingModule 6.3 · Topic 4
3Maintaining Transparency for TrustModule 7.3 · Topic 1
4Scaling AI Across TeamsModule 4.2 · Topic 4
5Legal-Specific AI GovernanceModule 6.2 · Topic 5
6Managing Risk for ProtectionModule 7.3 · Topic 2

End · Reasonable Doubts

Contents

Cast

—Corinne AlvarezExecutive director and protagonistRuns a regional legal-aid office with a waiting list that never shrinks. She is an enthusiastic amateur AI user at home and a candid laggard at work, more willing to name what she does not understand than to pretend certainty. A sixty-day board deadline forces her to build enough understanding to lead an organizational decision that balances service, risk, governance, fundraising, and mission.
—Yolanda ReyesJunior accounting clerkAutomates her personal life enthusiastically and is technically fluent, fast, and eager. She has never used AI at the office because no policy permits it and initially does not understand legal concepts such as privilege that radically change the consequences of ordinary tool use. Her capability becomes valuable once it is combined with legal context and governance.
—Marcus FeldSenior HR specialistRecently arrived from the corporate world and has already run an organizational AI rollout. He brings useful implementation experience, but his prior playbook comes from an environment without legal privilege, professional competence duties, or the vulnerability of legal-aid clients, forcing him to adapt rather than import it wholesale.
—Priscilla OyelaranStaff attorneyThe office's most careful lawyer. She possesses excellent professional judgment and can identify the risks better than almost anyone, but cannot work the AI tools herself. Her knowledge complements Yolanda's tool fluency and Marcus's rollout experience, creating an internal group in which no one person's competence is sufficient on its own.
—Dev RamachandranIntake supervisorOverwhelmed by an endless queue and has quietly been pasting client details into a free chatbot for months to keep up. His conduct embodies shadow adoption driven by operational pressure rather than simple recklessness, and makes the cost of organizational inaction impossible to treat as theoretical.
—Harriet BloomBoard chairGives Corinne the sixty-day deadline and insists on a genuine organizational vision rather than another research report. She keeps the process oriented toward a board decision, a fundraising campaign, and a governable mandate.
—Gordon TraskLarge-firm pro bono partnerPrepared to finance a highly visible flagship AI clinic and wants his name attached to it. He brings resources and enthusiasm while also introducing incentives that may not align perfectly with the legal-aid office's actual needs or governance capacity.
—Simone AdeyemiFoundation program officerRepresents a common funding tension: willingness to pay for visible innovation but much less enthusiasm for financing the governance, controls, and organizational capacity that make innovation safe.
—Ray OkonjoPeer legal-aid executive directorRuns an office two states away that adopted earlier. He can show both improved intake results and the operational scar tissue left by moving quickly, giving Corinne evidence that neither enthusiasm nor caution captures the whole experience.
—Tessa VaughnVendor account executiveA polished technology seller whose demonstration stays slightly ahead of what the product can actually deliver. Her claims give Corinne's team a practical test of whether their growing literacy is sufficient to interrogate vendors rather than be led by them.
—Ellery PikeBoard memberA technology-optimistic donor and the board's strongest advocate for moving quickly. He gives the final decision a credible pro-adoption position that must be answered rather than caricatured.
—Nathan RossBoard member and treasurerThe board's cautious financial voice. He evaluates ambition against cost, sustainability, governance capacity, and fiduciary responsibility.
—Dolores VanceBoard memberA client-community representative and former client of the office. She keeps the debate anchored in the people the organization exists to serve, especially those who cannot easily absorb a confident error or experimental failure.