No Obvious Owner

For twenty-six years Glenda Pruitt has handled the problems nobody else quite owns. Then AI lands on her desk with one difference: nobody can tell her what question she is supposed to answer.

Reasonable Doubts · 18 chapters · ~2 min each

Cast →

Chapter One (~3 minutes)

The Ask

Aaron Quist stood at the edge of her desk in the posture of a man who has been sent. Glenda had watched twenty-six years of that posture, summer associates and bookkeepers and once a bank manager with a lease he did not want to sign, and she knew before he spoke that whatever he was carrying did not belong to him.

"Warren wondered whether you might look into AI," he said. "For the firm."

She turned the retention schedule face down. "Look into what about it?"

"That's..." He stopped, and she thought better of him for stopping. "I don't know. I was told to ask you. I wasn't told what for."

"What would he like to have at the end of it?"

"He said, see what you think."

Glenda wrote the date on a fresh folder tab. She had opened files this way for the professional-liability renewal, the retention schedule, the move to the fourth floor, and the associate two summers ago who would only be unhappy at her. Anything that arrived without a natural home arrived here, and her method never varied: read everything, put it in order, recommend, and let the partners decide direction and money. She did not argue for her recommendations. Twice in twenty-six years one had been set aside entirely, and she had not minded either time, because a recommendation had never been the same thing as herself.

"Give me the names of everyone who's raised it with you," she said.

He gave her four, then a fifth, then the honest thing. "It's everywhere, Glenda. It's in the kitchen. It's in the elevator."

That much she knew. Since about February the building had acquired a weather. Howard Pell had come back from a bar function saying the firms two states over were three years ahead of them. Renata Cruz had told a client, within Glenda's hearing, that the firm was using AI on the review, which was either true or it wasn't, and Glenda could not have said which. What struck her was that all of it agreed and none of it could be traced. Nobody could say where they had heard the thing they were certain of. Twenty lawyers and eleven staff held two propositions without evidence: that they ought to know something they didn't, and that they ought to be doing something they weren't.

That night she did what she always did first, which was find out whether somebody had already written the answer down.

What she found was not vague. The professional standard had moved, and recently, and not by degrees.¹ A lawyer using one of these systems was now expected to understand what it could and could not do, and to keep that understanding current as the thing changed underneath them. That was not a technology question. That was competence, a subject on which this firm held firm views, generally about other firms.

The rest followed and got worse as it went. The duty landed on the individual who used the tool, not on the department that bought it, and it came with an arithmetic nobody at Salter Nunn had ever performed: the hours a year it takes to stay current with something that will not hold still.² Not one of those hours was in anybody's budget. And the anxiety that had been walking the halls since February turned out to have a literature of its own, thick with surveys reporting how many clients now simply asked their lawyers the question outright.³ So the hallway was not wrong. The hallway was unsourced, which is a different problem and a harder one, because you cannot answer a feeling.

The one thing she could already state plainly she wrote out twice, to be sure of it. Understanding meant knowing a tool's capabilities and its limits well enough to say which work was safe to hand it, which had to be checked, and how hard.⁴

At eleven she put down her pen. She had a folder with a date on it, five names, three pages of notes, and no question.

Every other file she had ever opened at this desk had arrived with its question already inside. Renew or don't. Seven years or ten. Fourth floor or sixth. This one had arrived with nothing in it at all.

She wrote What are we being asked? on the inside cover, underlined it, and went to bed.

Chapter Two (~3 minutes)

Terms of Reference

She gave herself a week and one question, and asked it the same way each time. If I bring something back, what would you want it to answer?

Warren Nunn said, "See what you think," which she had expected, and then, because he was a kind man who knew he had given her nothing, "You'll know it when you've got it."

Cecily Bourne wanted to know whether the firm was doing it. Miles Ferraro wanted to know whether the firm was doing enough of it. Howard Pell wanted to know what the big-city firms had. Tanya Osei wanted to know whether it was going to get somebody sanctioned. June Ellery wanted to know what it cost. Renata Cruz wanted to know who would be in charge of it, and asked twice. Dot Ackroyd, nineteen years a paralegal, said, "Ask me a narrower question and I'll answer it," which was the most useful sentence Glenda collected all week and the only one she wrote down word for word.

By Friday she had fifteen questions from eleven people. Laid side by side they were not variations on a theme. They were fifteen different files, and several of them had opposite answers.

She asked one follow-up of everybody: where did you hear that? Pell had it from a man at a bar function who had it from his firm's consultant. Cruz had it from a webinar she could not name. Ferraro said, reasonably, that he had not heard it anywhere, he had simply noticed. Nobody could produce a source for the thing they were all certain of, and the certainty did not weaken when its source failed to appear. It behaved, she thought, like every rumor she had ever managed, except that this one might be true.

The reading she did that weekend was written for people instructing machines, and it kept describing, with a precision she found faintly insulting, the anatomy of an instruction that actually works: what is to be produced, the boundaries it must respect, what "good" looks like, and when to stop.¹ Four things. Her mandate had none of them.

It went on. Clarity and specificity determine what comes back; supply the background, name the format, and where the conclusion matters, ask to see the reasoning.² And then the passage that made her set the page down: vague tasks fail. Not occasionally. The scope has to match what the thing being asked can actually do, and anything large has to be broken into units with a defined handoff, or it collapses under itself.³

Last, a worked specification: a name that is a verb phrase, what goes in, the steps, what makes the output acceptable, what comes out, and a round of testing before anybody relies on it.⁴ The example it gave of a bad one was a single paragraph describing roughly what somebody ought to do.

Glenda thought about see what you think, and did not smile.

She had never resented the partners for deciding. That was the arrangement, and it had suited her for twenty-six years: they carried the risk and the clients, she carried the order of things. What she noticed now was narrower and would not go away. She was not being asked to answer a question. She was being asked to write one, and the writing of it was the part with judgment in it, and it was the part nobody would sign.

So she wrote it. Three-quarters of a page, headed Terms of Reference: the objective she proposed, the boundaries, what a usable answer would look like, and the line she was proudest of and least sure of, which was what she would not be doing. She left the last line blank because she did not know how to fill it in, and took it to Warren.

He read it standing up, which is how he read everything.

"That's about right," he said, and handed it back unmarked.

No signature, no scope, no budget. What she had that she had not had on Monday was a document the firm had declined to disagree with. She knew exactly what that was worth. It was worth proceeding.

Chapter Three (~2 minutes)

The Browser Tab

The professional-liability renewal came round every June, and every May Glenda asked each practice group to inventory the software it used. It was the dullest recurring task she had and had never once taken more than an afternoon.

She was doing it early, and for another reason, when Miles Ferraro turned his screen around.

"You'll want this on there," he said, pleased with himself and with her for asking. "We've been doing first passes on term sheets with it since about September."

She came round the desk and looked. A browser tab, an ordinary one, with the Harrow Valley supply agreement in the box at the bottom and a competent summary of its unusual clauses above.

The work was good. She could see that it was good. Ferraro's group had turned a term sheet in a day and a half faster than the year before, and the client had noticed and said so. He was not hiding anything. He had told the associates to use it and had shown two of them how.

"Which one is it?" she asked.

"The free one."

"Does the firm have an account?"

"I assume somebody does."

"What happens to what you paste in?"

There was a pause of the kind she had learned to wait through. "It gives it back to you," he said.

That evening she learned the thing that reorganized the whole file. The same underlying system reaches you through several different doors: a public web page, a firm account, a panel inside the mail program, a version running on machinery the firm controls. The doors are not interchangeable. What differs between them is where the words go, who keeps them, and for how long.¹ Ferraro had chosen a door. He did not know he had chosen anything.

The exposures had names, too, and there were five of them, and she read the headings twice before deciding which one they were standing in.² It was the confidentiality one, with privilege standing directly behind it. Client terms, not yet public, typed into a service the firm had never examined, on eleven matters, since September.

She sat with that for a while, and then noticed something that changed its color. Ferraro had broken no rule. There was no rule. The gap the material described, between what a firm's leadership assumes its people are doing and what they are actually doing, was not a gap here, because nobody had ever formed an assumption to be wrong about.³ Fewer than half of organizations had anything written down at all, according to the same page. Salter Nunn was in the larger half, and had been all along, and had not known it.

The plainest sentence she found was in the most elementary material, which was where she usually found the plainest sentences: establish whether the tool keeps what you give it, and keep client-confidential material out of anything that falls short of your own standards.⁴ Nobody had established anything, because establishing was not anyone's job.

She had been sent to find out whether the firm should begin.

The firm had begun in September. It had begun on a Tuesday, in a browser tab, on a good matter, and the reason it had never been discussed was that no one had ever been asked to decide it. Non-use had not been the firm's position. The firm had never taken a position. That was not the same thing, and the difference was going to cost somebody a great deal of work.

By Thursday the inventory ran to two pages.

Chapter Four (~2 minutes)

Faster Than the Lawyers

Dot Ackroyd's row in the inventory was blank because Glenda had left it blank, and she had left it blank because she wanted an hour and not five minutes.

She got the hour on a Wednesday, in the file room, which is where Dot preferred to be.

"You're going to ask whether I use it," Dot said.

"I'm going to ask what you do with it."

Dot went to a drawer and came back with a coil notebook. It was not a notebook about the tools. It was a notebook of what worked. Fourteen pages, nineteen years of handwriting compressed into eighteen months. The wording that got a usable chronology out of four hundred pages of medical records. The wording that hadn't, struck through, dated. A note that the thing invents dates where the record is illegible, and invents them plausibly, so the illegible ones get checked twice. And a rule of her own devising, underlined: never ask it a question the file cannot answer.

Glenda read the notebook twice. The second time her stomach went cold, because she was holding something the firm owned without knowing it owned it, and would lose on a Friday afternoon whenever Dot decided to go.

That was the shape the material described, when she got to it. A firm typically has one to three people genuinely fluent and everybody else unaware or wary, and the fluent ones quietly become a bottleneck, because work routes to them as the ones who know how. The move that matters is not turning everyone into an expert; it is writing down what the expert already found and letting the rest execute it.¹ A discovery like Dot's is not supposed to live in a personal notebook. It is supposed to be documented, tagged, versioned, and reachable, so that the tenth person to need it does not reinvent it badly.²

The checking had a literature too, and reading it was the uncomfortable part. Verification was meant to be systematic rather than instinctive: go to the original rather than the paraphrase, confirm the passage actually says what is claimed of it, check whether it is still current, look for the authority that cuts the other way, and write down what you checked, because an undocumented check is indistinguishable from no check at all.³

Dot did the first four by reflex. She did none of the fifth. There was no record anywhere in the building that she had ever caught anything, and she had caught, by her own count, eleven things.

Then Glenda found the sentence that turned the chapter of her file over. Everything Dot produced this way, the chronologies and the extractions and the summaries that went into memos, was meant to be reviewed by a lawyer before anybody acted on it, and the responsibility for that review sat with the lawyer and not with Dot.⁴

She checked. It hadn't happened. Not once, not because anyone had decided against it, but because the work came back looking finished and Dot's work had always been reliable, and reliability is the thing that makes supervision feel unnecessary right up until the morning it isn't.

"Has anybody ever gone behind you?" she asked.

Dot thought about it honestly, which was one of the reasons the firm was lucky to have her. "Tanya read one right through, in February. She said it was fine."

"Did she say what she'd checked?"

"No."

Walking back to her desk, Glenda tried to name what she had actually found, and it took her the length of the corridor. The most careful reader of these outputs in the building was not a lawyer. She had built, unpaid and unasked, the only quality control the firm had. And nobody had ever read her.

Chapter Five (~2 minutes)

Two Who Want to Be Asked

She saw them separately and on purpose, and both of them arrived having prepared.

Renata Cruz came with a demonstration. She had the firm's standard shareholders' agreement open in one window and a paragraph of clean, unremarkable prose in the other, and she said the thing Glenda would hear from her twice more that spring: it's all the same underneath, you just pick a bigger or smaller one depending on how much you want to spend.

Simon Vasco came with a warning. He had read about the sanctions. His position was that the tools were fine for tidying language and catastrophic for anything factual, and that the firm should say so in writing, ideally in a memorandum bearing his name.

Both of them were partly right, which was the difficulty. Neither of them could tell her which part.

The spectrum was not a matter of size. A fast model answers in seconds and is priced for volume; a reasoning model spends thirty or sixty seconds deliberating and will show you the steps it took, which is the whole point of it where the conclusion has to be defended to somebody; a deep research tool breaks a question into sub-questions, goes looking, and comes back minutes later with sources, at five or ten times the cost.¹ Renata's bigger-or-smaller was not a simplification. It was a different claim, and it was wrong.

Simon's rule failed in the other direction, and more interestingly. The risk was not evenly spread across "facts." It tracked a line between asking the thing to work with what you had given it and asking it to produce something out of itself. Summarizing the agreement in the window in front of it was the safe side. Recalling a holding it had not been shown was the dangerous one, because when the pattern runs out the model does not stop; it produces the most plausible continuation, and a plausible continuation is indistinguishable from a real one from the inside.²

So the answer to may we use it was not yes and not no. It was: for what.

What stayed with her longer was a passage about how to tell a reliable answer from a confident one. Vague sourcing. Absolute statements where a qualified one would do. No caveats anywhere. A conclusion arriving before the evidence.³ She read the signals twice and thought, without pleasure, that she had spent an hour that morning with two people who matched them, and that three weeks ago she had matched them herself.

She tested it. She asked Renata what happened to a document after it was pasted in. Renata said it was encrypted, with the confidence of somebody who had heard the word. She asked Simon which of the sanctioned lawyers had used which tool. He could not say, and had the grace to look at the table.

The last thing she read that night was a short piece on checking what these systems produce: don't try to verify everything, that isn't possible; pick the five or ten claims that would do the most damage if wrong, verify those, and if one fails, go back and read the whole thing again for the same fault.⁴ It also said to look for the document contradicting itself between one section and another, which struck her as a good test for a firm as well.

She wrote her rule for the file in one line, and it was not about technology at all.

Do not take an answer about these systems from anyone who has no way of finding out they were wrong.

Then, because she was fair, she added underneath: including me, until June.

Chapter Six (~2 minutes)

Nothing Passed Through Anything

Cecily Bourne read the inventory without touching it, hands flat on the table either side, the way she read everything.

"You want to know whether my group is going to use it."

"I want to know what your answer would be if the firm said yes."

"My answer would still be no. Do you want the reason or the sentence?"

"Both, in that order."

The reason took twenty minutes and Glenda did not interrupt any of it. A family-law affidavit is a person's account of the worst period of their life, in their own words, sworn. What the tools were being sold as good at, namely smoothing tone, shortening sentences, and turning the language of a frightened person into something an audience finds easier to receive,¹ was not an improvement to that document. It was the removal of the only thing the document had.

"I read them aloud to clients," Cecily said. "I've had a woman stop me at a line and say, that's not how it was, and we've changed it, and the changed version was the true one. I can't do that with a paragraph neither of us wrote."

Then the harder half. If the firm used one of these systems materially on a client's matter, the client would have to be told, and told properly, in advance, in terms a person could actually weigh, not buried in a retainer letter.² Cecily's objection was not that the conversation would be embarrassing. It was that she could not have it honestly, because she could not answer the questions it would raise: where the affidavit goes, who at the vendor can see it, how long it is kept, whether it trains anything, which other companies stand behind the one the firm signed with.³ And behind all of that, the possibility that putting a privileged document through the wrong service is itself the act that costs the client the privilege.

"I can't get consent to a risk I can't describe," she said. "So I don't ask."

Glenda wrote that down verbatim. It was the first sentence anyone had given her all spring that would survive being read out at a partners' meeting.

Then Cecily did the thing that made her, in Glenda's private estimation, the best lawyer in the building.

"Don't write me down as against it."

"You've just spent twenty minutes..."

"Against it on affidavits. On parenting plans. On anything a client signs believing it's theirs." She turned the inventory round and put a finger on the third row. "Retainer letters, I don't care. Closing letters, I don't care. If it will read a bank statement disclosure and tell me which months are missing, I'd take that tomorrow, and so would every family lawyer in this state."

Reading afterwards, Glenda found Cecily's distinction already written out by other people, in a table that sorted work by type and said plainly where a machine could draft and where a human had to, and which said in every row that the person who hands the thing over is the one who answers for it.⁴ The firm did not have that table. Cecily had been carrying it in her head, unwritten, and had assumed, as everyone in the building assumed about their own practice, that it was obvious.

Walking back, Glenda understood that she had been asking the wrong question for six weeks. There was no answer to may the firm use it, because the firm was not the unit that used anything. Desks used things. Matters used things.

She went back to her office and drew a line down a fresh page.

Chapter Seven (~2 minutes)

What the Big Firms Have

Howard Pell had come back from the estates and trusts conference with a grievance he could not specify, and he had been carrying it since March.

"They're all doing it," he said. "The big shops. Somebody in the session had a thing on his laptop and everyone at the table had seen one."

"What did it do?"

"It knew their documents."

That was the whole of it. Four months of certainty compressed into three words, and to Howard's credit they turned out to be the most accurate description anyone at Salter Nunn had produced all spring.

It took her two evenings to work out what he had probably seen. Not a smarter machine. A machine pointed at somebody's own filing cabinet. The technique had a name and an unglamorous mechanism: documents are ingested, converted into something searchable, and when a question comes in the system finds the relevant ones and hands them over before it answers, so the answer comes out of the firm's material rather than out of the general soup.¹ The literature was candid about how it fails, too, and the failures were the quiet kind. The retrieval can miss and the system will not tell you it missed; a question about one thing can pull a document about a neighboring thing that reads plausibly; and where there are thousands of documents only a handful come back per question, so the decisive one can simply never surface.

Which meant that what Howard wanted was not a purchase. It was a curated shelf: the firm's own precedents, weeded of superseded versions and duplicates and the eleven variants of the same will that everybody knew to avoid. The weeding was the product. Nobody sells you that. You do it yourself, in hours, and the hours belong to the people who know which version is the good one.

She kept reading, because Howard's grievance had opened onto the whole catalog of what these things were actually being used for in practices like his: reading a document against a checklist of unusual terms, preliminary research to be verified afterwards rather than relied on, client correspondence, checking a filing against local requirements before it goes.² Some of that Salter Nunn already did badly by hand. Some of it Dot already did well.

The tempting middle turned out to be smaller than it looked. There were tools trained specifically on legal material, and they did outperform general ones on legal tasks, and they cost more and were worse at everything else; there was also the option of training a general one on the firm's own precedents, which needed fifty-odd good examples and a maintenance commitment the firm would forget about by winter.³ For a twenty-lawyer full-service shop, where the same person does an estate on Monday and a lease on Tuesday, the trade ran against specialization and Howard would not like hearing it.

What was within reach was duller and better. You could write a standing set of instructions for a recurring task, covering the format the output must take, the fields it must fill, and the rule that it must never add, infer or assume anything the document does not say, and save it, so that the tenth judgment reviewed comes back in the same shape as the first.⁴ That was a thing Dot could build in an afternoon and the firm could actually own.

And the honest description of the work these things were good at was narrower than the conference table implied: finding patterns across a stack of documents, reviewing a draft for gaps and weak points, putting two versions side by side and showing the differences, with every pattern treated as a hypothesis to be checked against the source rather than a finding.⁵

She wrote Howard's want down in one line and then wrote its price underneath, and the price was not the money.

Chapter Eight (~2 minutes)

The Certification

Tanya Osei did not lecture. She put a folder on the table, sat down opposite, and let Glenda open it herself.

Printouts. A lawyer in New York whose brief cited cases that did not exist, and who said, when it came apart, that he had not understood the tool could invent them. A patent firm sanctioned twelve thousand dollars for filing citations nobody had checked. An appellate court writing guidance into an opinion because it had seen enough of them. A firm large enough to be ranked by headcount, sanctioned for an appellate brief with fabrications in it, and covered in the national press for a week.¹

"Read the third one properly," Tanya said.

The third one was not a fabrication. The cases existed. The lawyers had relied on the machine's summaries of them without opening the cases themselves, and the summaries had got the holdings wrong. That was the one that frightened Glenda, because it did not require anybody to be careless in a way she would have recognized as careless.

"There's a shape to them," Tanya said. "It's always one of three. You cite something that isn't there. You accept a summary of something that is there without reading it. Or you don't tell anyone you used it when the rules said you had to."

The checking regime that answered this was not complicated, which somehow made its absence worse. Confirm the authority exists. Read it and confirm it says what the summary claims. Confirm it hasn't been reversed or overruled since. Confirm it actually supports the point it is being cited for. Then write down that you did all four, because a check nobody recorded is a check nobody can prove.²

There was a discipline on the front end too, and it was mostly about being specific: name the jurisdiction rather than letting the thing choose one, ask it to identify its sources and say where its confidence is low, and take an "uncertain" over a confident guess every time.³ Glenda copied out the figure that came with it, that studies of legal systems of this kind had found something in the order of one hallucination in six answers, and looked at it for a while, because it was the first number in the whole file.

Then Tanya said the thing that went into the file underlined.

"Understand what the court is actually asking for. It isn't asking whether a machine helped. It's asking whether the person who signed it stands behind every line. That's not a new rule. It's the old rule, and it's mine, because it's my name."

"Some of the courts want it certified expressly."

"Some do. More will."

Glenda asked the question she had been carrying since February. "In February you read one of Dot's chronologies through. You said it was fine. What did you check?"

There was a long pause, and Tanya did not decorate it. "I read it. It read right." She turned her hand over. "I couldn't tell you now what I checked, and that means I didn't."

That was the whole book, Glenda thought later, in eleven words from the most careful litigator in the firm.

The material called for scaling the checking to what was at stake, light for a brainstorm and heavy for anything going to a court, and for writing down, per task, who was responsible for verifying what.⁴ Salter Nunn had no such record for any task. It had eleven matters in the commercial group, a drawer full of chronologies, and one lawyer's honest admission.

On the way back she wrote two words at the top of a clean page and underlined them twice. Who signs?

Chapter Nine (~2 minutes)

Two New Pages

Neil Brackett had been placing the firm's professional-liability cover for eleven years and had never once come up in person to deliver a renewal package.

He came up in person.

"There's two new pages," he said, and put it on her desk the way you put down something you would rather not have carried. "Nine questions. They're not optional and the answers go on the application."

Glenda read them standing. Does the firm maintain a written policy governing the use of artificial-intelligence tools? Are staff trained, and is the training documented? Describe your verification procedure for AI-assisted work product. What client data may be processed by such tools, and under what agreement with the vendor? Has the firm disclosed such use to clients? List the tools in use. And at the foot, in bold, the date: twenty-six days.

Every honest answer was no.

What she had not understood until that evening was that this was not a form. It was the carrier moving the risk. The cover had always answered for errors in the delivery of legal services, and an error made with one of these systems was still, in principle, an error like any other, but the principle now came with conditions, and insurers were pricing the firm not on whether an error happened but on what the firm did to prevent it: the written policy, the training records, the verification workflow, the controls on what client material goes where, and whether the use had been disclosed at all.¹

The exclusions were where her stomach turned over. Carriers were writing out coverage for undisclosed use, for reliance on output nobody verified, for confidentiality breaches arising through these tools, and, the one she read three times, for fabrications, on the theory that a system known to fabricate is a risk the firm chose rather than an accident that befell it.² Ferraro's eleven matters sat inside two of those. Dot's chronologies sat inside one.

The remedy was almost insultingly plain, and she recognized the shape of it because it was the shape of every renewal she had ever prepared. The carrier did not care what the firm knew about any of this. It cared what the firm could produce: a policy, verification procedures written down, training records with dates and names, disclosure recorded in engagement letters, a map of what data goes to which vendor under what agreement, and a log of the errors the firm's own checking had caught.³ An action problem, not a knowledge problem. Salter Nunn's knowledge had improved enormously in four months and its evidence was a folder in her drawer.

And it would keep moving. The direction of travel was legible in the material, in narrower and more specific exclusions, prescriptive training, standardized disclosure language and formal agreements with vendors, and what survived all of it was three habits rather than any particular answer: tell people, match the effort to the stakes, and write down what you did.⁴

She took the two pages to Warren, who read them standing up, and then sat down, which she had seen him do perhaps four times.

"Twenty-six days," he said.

"Twenty-six."

"What happens if we answer honestly?"

"They price it. Or they add conditions. Or they decline the AI part and we carry it ourselves without knowing we're carrying it."

"And if we answer the way it should be answered?"

"Then we have to be a firm that can say those things by the twelfth."

Warren looked at the form for a while. Then he gave her the first instruction of the whole business with a boundary inside it.

"Answer it the way we can defend."

She wrote that down verbatim too. It was not a scope and it was not command, but it had an edge to it, and for four months everything she had been handed had been round.

Chapter Ten (~2 minutes)

The Thing That Talks Back

The bar association had the course. It had had the course since the autumn, in the members' area, behind the same login Glenda used to renew nine practicing certificates every December, and in five months nobody at Salter Nunn had mentioned it, because nobody at Salter Nunn had looked.

She started on a Sunday and gave it three weeks of evenings.

It was not built the way she expected. There was no sequence to complete and no certificate at the end; you were meant to move around it as your questions moved, and the questions were meant to be put to an assistant sitting inside the material, so that you learned the thing by working it rather than by reading about it.¹ The first exercise asked her to open it and ask it who it was and what it was for, and then to ask how it knew what it knew.² She felt ridiculous doing it. She did it anyway, at her kitchen table, at ten past nine on a Sunday, and something in the twenty-six years of her working life turned slightly on its axis: for the first time, the subject of her research answered her back, and then let her ask a worse question.

By the end of the first week she had the vocabulary, and the vocabulary did most of the work. There was a hierarchy: the broad idea of machines that mimic intelligence, then the narrower one of systems that learn patterns from data, then the narrowest and the one everyone in the hallway meant, which produces new text by predicting what comes next.³ A token was a chunk of text roughly the size of a word. A knowledge cutoff was the date its training stopped, past which it knew nothing unless something fetched it. And a hallucination was not a malfunction. It was the ordinary operation of the thing running past what it had patterns for.

The second week was the machinery, and it was shorter than she expected and worse for the firm. Three stages: learn the patterns; turn the words into numbers and push them through layers of arithmetic; then produce a probability for what should come next, choose one, and do it again, and again, until the answer is finished.⁴ Ask it the same thing twice and you may get two different answers, and that is not a fault in the product. It is the product.

She put her pen down at that.

Because it meant the thing at the center of five months of certainty was not a machine that knew things. It was a machine that continued sentences. Everything she had been told since February arranged itself behind that sentence like filings behind a magnet. Ferraro's it gives it back to you. Renata's it's all the same, just bigger. Simon's blanket prohibition, which was too broad in the same way Renata's confidence was too broad. Howard's it knew their documents, which had been right, and right for a reason he could not have supplied: because somebody had put the documents in front of it, and it had continued sentences out of those instead of out of everything.

At the end of the third week she wrote one sentence at the top of a clean page and did not change a word of it afterwards.

It does not know things; it predicts what should come next, and it is very good at that and indifferent to whether the result is true.

Then she sat in her kitchen with the light off, feeling something she had no practice at. In twenty-six years she had never once finished a piece of research knowing more about the firm's exposure than the people who owned the firm. It was not satisfaction. It was closer to vertigo. Competence, she was discovering, is only comfortable when somebody above you has it too.

Chapter Eleven (~2 minutes)

Our Size, Sold Back to Us

Sloane Vickery had done her homework, which Glenda respected, and had done it on the firm rather than on the problem, which she did not.

"Twenty lawyers," Sloane said, to Warren, to Howard, to Renata, and to the notepad at the end of the table. "Full service, one office, a state with no tier-one city. You're exactly the size that can't afford to wait. The big shops have their own people for this. The three-person firms don't feel it. It's you who gets squeezed."

It was a good line. It was also the firm's own dimensions handed back to it as an argument, and Glenda wrote it down in those words.

The demonstration was genuinely impressive. A scanned exhibit went in, crooked, and came back transcribed. A supply agreement went in and came back with the unusual terms flagged by risk. Then several components of the platform passed work between them, one gathering and one drafting and one checking, and produced a memorandum in under two minutes, and Howard made a small sound.

Glenda waited until the end, because that had always been her seat, and then asked her questions in the order she had written them.

Which underlying model was it built on, and what was its knowledge cutoff, and how much could it take in at once? Because every tool of this kind runs on somebody else's foundation model, and those differ in the date their knowledge stops and in how much they can hold at a time, which decides whether a fifty-page agreement is read in one pass or in pieces.¹

What was doing the work in the demonstration: the model, or the layers around it? Because a tool like this is an engineered stack: hidden standing instructions, a retrieval step that goes to a document store before answering, a routing layer that turns text into actions in your systems. If the memorandum had been wrong, she wanted to know which layer she would be complaining about.²

Then the list she had built over two evenings, and which she read out without apology: encryption in transit and at rest; multi-factor and role-based access; audit logging of every call; a current certification with a date on it; where the data physically sits and under whose law; the retention and deletion policy; and whether anything typed into it is used to improve the product.³

And last, the one Warren afterwards said he had not thought to ask in thirty-one years of buying things.

"What happens when we leave? Can we export everything, in what format, and will it load into anything else, or are the connections proprietary, so that leaving means rebuilding?"⁴

Sloane answered six of the nine well. On two she said, plainly, that she did not know and would find out and put it in writing, which was the most credible thing said in the room. On the ninth, the export format, she said the question had never come up in a meeting before.

Afterwards, in the corridor, Howard was almost shy about it.

"How did you know to ask that?"

"I read for three weeks."

"No," he said. "How did you know it was the question?"

She did not have a good answer, and gave the true one, which was that she had spent twenty-six years signing the firm out of things it had signed into carelessly, leases and service contracts and a photocopier, and the only difference here was the vocabulary.

That night she wrote up the assessment and found the hole in it, and the hole was not the vendor's. Every question she had asked was a question about the product. Not one of them was a question about what the firm needed, because the firm had never said.

Chapter Twelve (~2 minutes)

Not One Answer

She took four days and one question, and she wrote the question down so that she would put it identically to everybody.

May this file touch that tool?

Not whether the firm should. Whether this file, in front of this person, may go into that specific thing, today.

Cecily Bourne said no to the affidavit on her desk and yes to the closing letter beside it, and had said as much already. Miles Ferraro said yes to a term sheet with the parties' names taken out and no to the same term sheet with them in, which was a distinction he had not made in September and made now without prompting. Tanya Osei said yes to a chronology built out of a bundle she supplied and no to any authority that had not been through her four checks. Howard Pell said yes to comparing two of the firm's own precedents and no to anything a beneficiary would read. Dot Ackroyd said the sharpest version: it depends on whether the answer is in the file or has to come from somewhere else.

Eleven desks. Eleven answers. Not one of them wrong.

What was underneath them turned out to be a scheme Glenda could have drawn from the material without leaving her desk, except that she would not have believed it until she heard it in eleven voices. Work sorts by what it does to a person and how hard it is to undo: at the top, decisions that touch someone's rights and are difficult to reverse, which need assessment, a human able to override, and a record; in the middle, work that affects the firm's operations and can be caught; at the bottom, drafting and summarizing where somebody reads it before it matters and the cost of error is an afternoon.¹

The checking sorted the same way. A quick read for the low tier, two to five minutes, looking for the obvious tells: confident vagueness, a citation with no case attached to it, a statistic too precise to be real. Detailed verification against sources in the middle, a quarter-hour to three-quarters. Expert review at the top, where somebody who has done the work for years asks what only experience would catch.²

And "good enough" was not one standard either. An internal draft needed to be conceptually sound. Client-facing work needed every fact accurate and every source verified. Anything that carried real liability needed authoritative sources, disclosed limitations, an audit trail, and a named person signing it off.³ Even trust was not a setting you chose once and kept; it was meant to be recalibrated against each piece of work, with the rigor matched to what was actually at stake.⁴

On the Thursday evening she wrote out what she had, and she wrote it slowly because it was the moment the whole thing changed shape.

There was no firm-wide yes. There was no firm-wide no. Anybody who produced either would be wrong at eleven desks simultaneously, and would be ignored within a month, which is worse than being wrong.

What varied between the desks was not the technology. It was identical technology in every case. What varied was the consequence of being wrong, a re-drafted letter here and a custody order there, and that is a fact about clients rather than about tools.

So the document could not be a policy in the sense the hallway expected, which was a page telling everyone the same thing. It had to be a floor: the small set of things that hold everywhere, whatever desk you sit at, with the rest left deliberately open and different by practice group, because the desks were right and the firm was one firm anyway.

She wrote the word floor on the inside cover, next to What are we being asked?, which she had written in February and had not been able to answer until tonight.

Then she wrote the sentence underneath that spoiled her satisfaction, because a floor is only a floor if it holds, and holding is a thing a document cannot do by itself.

Somebody has to own this.

Chapter Thirteen (~3 minutes)

Who Is Answerable

A floor needs a foundation, and the foundation was the question the firm had been not asking since September: who may do what, on whose file, and whose name is on the result.

Glenda started where she always started, with the facts of the present arrangement, and the facts were worse than the absence of a rule. Dot Ackroyd could open every matter in the practice management system, including the two the firm held for the partners' own affairs, because the system had been configured in 2011 by a consultant who had left the state and had given the office manager and the senior paralegal everything so that nobody would be locked out over a long weekend. Nineteen years later Dot had never once looked at anything she had no business in. That was not a control. That was a character reference.

The principle she needed had a name, and it was not written for people at all. Each thing that touches a system should reach exactly the data and actions its work requires and nothing beyond, with credentials held properly rather than shared around, and every flow of information documented so that access can be audited afterwards and a problem traced.¹ Written for connectors and tools. Applying, she thought, rather uncomfortably to the fourth floor of a building on Prentice Street.

The second half was harder, because it was about lawyers.

What could be handed over turned out to sort cleanly. Work that runs on clear rules, that is easy to undo, where the cost of being wrong is small, and where the thing has already proved itself on that kind of task: that can go. Work that turns on judgment, that cannot be undone once it leaves the building, where the consequence is large, or where nobody has done it this way before: that stays with a person.² And between those, three kinds of checkpoint: approve the plan before anything happens; let it run and check the result before it takes effect; or let it run and be told only about the exceptions.

There was a passage on the ways supervision fails, and Glenda read it standing up. Too many alerts and people stop reading them. Reviews that arrive too late to matter. And the one she copied out in full: if you approve work without actually reviewing it, supervision is theater.

She thought of February, and of Tanya saying I read it. It read right.

The rest of it was strict in a way the firm would not enjoy. Delegation does not move responsibility. The supervising lawyer remains accountable for what a non-lawyer or a tool produces under their name, and for whether it complied with the duties they hold personally, and the way you demonstrate you did the job is by recording the instruction given, the review performed, the judgment exercised, and what you accepted, changed, or threw out.³ Research could be done that way; the citations could not be relied on until a lawyer had opened them. A routine notice could be drafted; advice could not be sent. Nothing in it was novel. All of it was unrecorded at Salter Nunn.

The supervision material was blunter still about what makes a checkpoint real: an approval criterion has to be concrete enough to act on. Comprehensive review is not a criterion. Every authority opened and initialed is.⁴ And the early signs of work going wrong, the same thing produced again and again without progress, quiet drift out of the shape you asked for, small errors early and larger ones later, were, she noticed, the signs of a person going wrong too.

She built the matrix over two evenings: who, which tool, whose file, what checkpoint, whose signature. It took four columns and fitted on one page, and when it was done she saw the thing she had been circling since the browser tab.

The firm had never given anybody permission. So nobody had ever exceeded it. Ferraro had not broken a rule; Dot had not overstepped; the reception software that joined a client call in March had not been authorized, and had not been prohibited either. Permission had simply never existed as a thing the firm did.

Which meant that on the day this matrix took effect, several people who had been doing good work in good faith would, for the first time, be capable of being in the wrong. That was the price of the floor, and somebody was going to have to be willing to charge it.

Chapter Fourteen (~3 minutes)

The Client's Schedule

Harrow Valley Dairy had been a client since the second Salter, and its general counsel had never once telephoned the office manager.

"You'll have seen Schedule C," Nadia Okpara said.

Glenda had not seen Schedule C. She said so, because the alternative was worse, and pulled the engagement terms while Nadia waited.

It was two paragraphs and it had been signed in January. Outside counsel were to disclose any use of these systems on Harrow Valley matters, identify the tools by name, warrant that no consumer service processed the company's material, retain records of human review, and notify within five business days of any incident. There was an audit right at the end of it, which Nadia said she had never exercised anywhere and had insisted on keeping.

"Miles signed it," Nadia said, without any edge. "I don't imagine he read past the fee schedule. Nobody does."

"Why is it in there?"

"Because we sell into Europe, and our board made me put a governance program in place, and a governance program that stops at my own front door isn't one." She said it the way people say things they have already said forty times. "Everything I'm asking you for, somebody asked me for first."

It was, Glenda realized afterwards, the first time in five months that the pressure had come from outside the profession altogether. The frameworks she had read about, the binding European regime with its risk tiers and its deadline, the American methodology that everyone builds their governance around, the certifiable standard that turns it into something auditable,¹ had felt like weather in another country. They had arrived on the fourth floor at Prentice Street as two paragraphs in a schedule, signed in January, by a partner who had read the fee page.

The rest of the call was worse and better. Better, because Nadia was not hunting anybody. Worse, because every question she asked had an answer the firm could not give: which tools, under what agreement, with what record of review, and what would happen if something went wrong on a Friday.

That was the whole of the governance material in practice: that you establish testing, approval, change control, confidentiality protection and an audit trail before anything runs on a client's matter, not after somebody notices.² Salter Nunn had run it on eleven of Harrow Valley's matters since September and had established none of it.

And the disclosure question, which Cecily had put in terms of her clients' dignity, turned out to have a colder second form. Where you practice across different courts and clients with different rules, you disclose to the most protective standard that applies anywhere in your portfolio, you put it in the engagement letter at the start rather than mentioning it later, and you keep the client's acknowledgment.³ Nadia had done exactly that to Salter Nunn. Salter Nunn had done it to nobody.

At the end she asked the question that Glenda would be turning over for the next three weeks.

"Who at your firm owns this?"

Glenda had prepared for the tools question, the vendor question, the retention question. She had not prepared for that one, and there was no honest answer that was not a paragraph, so she gave the paragraph: that the firm was in the middle of deciding, that a recommendation was coming, and that she was assembling it.

There was a short pause on the line.

"That's not an answer, Ms Pruitt. That's a person doing a job nobody's given her." Then, more kindly: "For what it's worth, we advise our own clients on this now. It's become a service. Half of them ask us who owns it before they ask us anything else."

The last of that fitted a thing she had read and not believed, that the firms which would get work out of this were the ones that could show a client how they governed it, and that the ability to answer the ownership question was itself the product.⁴

She hung up and wrote the date of the signature at the top of a clean page, because a date is the only argument that never has to be repeated.

Chapter Fifteen (~3 minutes)

What It Would Cost

June Ellery had kept the firm's books for sixteen years and had a way of receiving a proposal that involved no expression at all.

"Give me the hours," she said. "Everybody always starts with the licenses. Licenses aren't the money here and you know it."

Glenda gave her the hours, and they had taken four evenings to build.

The rollout the material described was not a purchase but a six-month shape: a period of telling everyone what was happening and why; then training and a place to practice safely, with the people who were already good at it identified and used; then the part where it gets built into how work is actually done; and only after that, measurement worth having.¹ Training was not one course. It went by role: everybody needed the foundation, the people using the tools needed hands-on work and verification practice, the supervising partners needed something different again. And the measurement came in tiers that arrived months apart, which meant nobody would know for half a year whether any of it was working.

The number that stopped her was in the same passage, and she copied it out for June: the great majority of technology investments that fail, fail on adoption rather than on the technology.

"So it's people," June said.

"It's hours."

The automation arithmetic was the honest kind. Measure how long the manual thing takes now and how often it goes wrong; run the new way alongside the old for a few weeks on the same work; then compare, and set the saving against what it cost to build, and see how many months it takes to get the build back.² Glenda had done that for one candidate, the chronologies Dot already produced, and the answer was that the firm had been getting a return since roughly the previous autumn and had never once counted it, because the hours saved were Dot's and Dot did not bill.

Then the question of build or buy, which Simon Vasco had answered before it was asked by offering to put something together over a weekend.

The comparison was less flattering to Simon than Simon expected. Buying a visual platform gets you running in days, with somebody else carrying the infrastructure and the security updates, at the price of a monthly fee and real difficulty moving later. Building it yourself gets you exactly what you want and a maintenance burden that never ends and belongs to whoever built it.³ Salter Nunn had twenty lawyers, no technical staff, and a history with a bespoke time-and-billing system that a nephew had written in 2009 and that took four years and eleven thousand dollars to get away from.

And there was a specific warning about the weekend build, which Glenda printed and put in the file without comment. Describing what you want and having something functional come back is genuinely possible now, and the thing that comes back is usually the version where everything goes right: no security hardening, no handling of the case where a source is unavailable or the input is nonsense, no thought about what happens at ten times the volume. Before any of it touches live work it needs a security review by somebody who can actually read it, testing under real load, testing of the failure cases, and a written account of who watches it and what they do when it breaks.⁴

June totaled the sheet twice, which she did when she disliked an answer and could not fault it.

"Two hundred and ten hours in the first six months. A hundred and forty of them from people who bill."

"Yes."

"At our rates that's not a cost, it's a small associate." She put her pen down. "Glenda, I've been doing this ledger sixteen years and every line on it has a name against it. Insurance, Warren. Premises, Warren. Library, Howard, God help us. This has no name against it, and I can't carry an unnamed line at that size."

"I know."

"Then that's your recommendation, isn't it."

Glenda did not answer, because it was, and she had known it since the Thursday she had written somebody has to own this and had hoped, without admitting it, that the arithmetic would let her off.

Chapter Sixteen (~3 minutes)

The Floor

She wrote it over a long weekend in October, at the kitchen table, and it came out shorter than anything she had produced in twenty-six years.

The skeleton was not hers. Four things have to stand together or the whole arrangement falls over: written rules about who may do what; training so people know why the rules exist; checking that catches errors before they leave the building; and controls on the tools themselves, so the wrong thing is not available to be misused in the first place. A firm with rules and no training fails when nobody follows them. A firm with training and no checking fails quietly, with well-taught people producing undetected errors.¹

Underneath that sat the tiering she had heard in eleven voices, keyed not to the tool but to what the material was and what it would cost to be wrong, with an approval route for anything above the bottom tier.² And running through it, an idea she had first met in the least likely place, that standards written once and kept where every matter can reach them beat standards argued out afresh on every file, because the arguing is where consistency goes to die.³

So the floor came to six lines, and she fought for every deletion.

No client material into anything the firm has not approved, and the approved set is a list somebody maintains rather than a memory. No authority relied on until a lawyer has opened the authority itself. Whoever signs has checked, and the check is recorded in the matter file, in one line, on the day. Every engagement letter says what the firm does and gives the client the chance to say no. Nobody touches an approved tool until they have done the training, and the training is recorded with a date and a name. And, quarterly, somebody reviews the whole of it against what is actually happening.

Above that, nothing. Deliberately nothing.

That was the part she expected to have to defend. Each practice group would set its own position above the floor and the firm would not require them to agree: the family-law desk could decline everything above the line and never be asked again; the commercial group could go considerably further inside the rules; estates could have its precedent project if estates did its own weeding. The standards material was honest about why this had to be so: a standard too onerous for the work it governs simply will not be followed, and an unfollowed standard is worse than none, because it produces a written policy and a real practice that do not match.⁴

Then she wrote the paragraph she had been thinking about since the browser tab, and wrote it plainly so that nobody could soften it later.

Nothing done before the date this takes effect is a disciplinary matter. Not the term sheets. Not the chronologies. Not the meeting assistant nobody remembers switching on. Anyone who tells the firm now what they have been using, and on what, is thanked and not sanctioned, and the record starts from there.

She had found the idea in the material, where it was put in three cool words about amnesty and enforcement going forward, and she had recognized it immediately as the only version of this document that Ferraro would help her with rather than survive.

By Sunday evening she had two pages. She read them through and found them good, and then found the hole, which had been in the sixth line all along.

Quarterly, somebody reviews the whole of it.

Somebody. Twenty-six years of writing that word and never once minding it, because somebody had always meant a partner, and a partner had always been supplied by the fact that partners owned things. She put the pen down at ten and looked at the two pages for a long time.

A floor is only a floor if it holds weight. What she had written could be read on a Monday, agreed by everybody in the room, and be worth precisely nothing by Thursday, because a document does not stop a matter. A person stops a matter.

Chapter Seventeen (~3 minutes)

The Line She Didn't Want to Write

She tried three versions of line six and threw away two.

The first was the obvious one, and she wrote it because it was true of every other file she had ever closed: the office manager reviews this quarterly and reports. It survived about four minutes. What the floor required was not administration. It was the authority to stop a matter, and the duty behind that authority was a lawyer's duty, held personally, which does not travel downward. Delegation does not move responsibility; the supervising lawyer remains answerable for what is produced under their name, whoever produced it.¹ She could keep the register. She could not be the reason a file stopped.

The second version was the diplomatic one: the partnership reviews this quarterly. She wrote it, looked at it, and heard June Ellery saying that every line in the ledger had a name against it. A thing owned by everybody is exactly the kind of thing that arrives on her desk in February with nobody able to say what is being asked. She had spent eight months proving that.

So the third version was the one she did not want.

A named partner owns this. That partner holds time in the budget for it, reports to the partnership quarterly, and has authority to stop use on any matter pending review. The named partner must first acquire, and then maintain, an actual understanding of these systems.

That last sentence was the trouble, and she knew it while she was writing it. The standard was no longer a matter of taste: a lawyer using one of these things is expected to understand what it does and cannot do and to keep that understanding current, and output that merely looks right is not evidence of anything.² Every word of it was defensible. It was also, on the page, a document written by the office manager telling the partners of Salter Nunn that they did not currently understand something and would have to go and learn it before they could be trusted with it.

She read it back at eleven and felt the thing she had never felt about a recommendation, which was that she did not want it altered.

That was new, and she distrusted it, and she sat with it for a while.

She had been reading, that same week, about how instructions stack inside these systems: a fixed layer set by whoever built the thing, which cannot be argued with and does not negotiate; the conversation's accumulated history; and underneath, the user's own request, which loses whenever it collides with what is above it.³ There was a companion passage about where the leverage actually is: not in the fixed layer, which is not yours, but in the standing instructions you are permitted to set, which quietly shape everything that follows without ever contradicting the thing above them.⁴ The recommended posture was not to fight the layer above you. It was to reframe, and to work through the one you could set.

Glenda read that twice, closed the folder, and went to bed.

She gave it to Aaron on the Tuesday because she wanted one honest reading before Warren's tray.

He read it slowly, all four pages, and then went back to the last sentence of line six and stayed there.

"You know how this lands."

"Tell me."

"Cecily will vote for it. Tanya will vote for it and be relieved somebody said it. Howard will take it personally for a week and then be fine." He put it down. "Warren will read that last sentence as being about Warren, because it is. And three of them will notice that the person who wrote it isn't a partner, and one of them will say so out loud, and after that it will be a thing about you and not a thing about the firm. That's the cost. It's not fatal and it's not nothing."

"Would you take it out?"

"No," Aaron said. "But I'm not the one who has to sit in the room with them for another nine years."

She did not sleep especially well, which surprised her, because she had never lost a night over a recommendation in her life.

In the morning she wrote a covering note, half a page, and at the bottom of it she did something she had not done in twenty-six years of putting documents in that tray. She signed it.

Chapter Eighteen (~3 minutes)

What She Was Offered

It took forty minutes of a partners' meeting that ran to three hours, and it was not the argument she had spent eight months preparing for.

The floor passed almost without discussion. Warren said the honest thing at the start and it took the air out of the room: that a client had already obliged them in January, that the renewal had a date on it, and that the question of whether to have rules had been settled by other people some time ago. Six lines, adopted. The amnesty paragraph went through untouched, which she had not expected, and Ferraro caught her eye across the table when it did.

Then the part she had written and had not been sure would survive: the firm declining to agree with itself.

Cecily Bourne put her position on the record in two sentences and asked that it not be revisited annually. Ferraro's group would carry on inside the rules, with the client's schedule now papered and Harrow Valley notified. Howard got his precedent project on the express condition, moved by June and seconded by Tanya with visible enjoyment, that the forty-six hours of weeding were his own. Nobody pretended these were the same answer. Nobody proposed that they should be. It went into the minute as three positions above one floor, and Glenda watched a firm say out loud, for the first time in her tenure, that it did not have a single mind and did not need one.

Warren mentioned, in passing, the thing Nadia Okpara had said about clients asking who owns it before they ask anything else, and observed that being able to answer that would probably be worth work within two years.¹ Three heads went round the table at that, which told Glenda more about how this firm decided things than the previous eight months had.

Then line six.

"Aaron takes it," Warren said. "Named partner, quarterly to the partnership, standing item on the agenda. Glenda does the work, obviously. She's already done most of it."

It was, on its face, everything she had asked for.

Tanya was the one who found it. "Who stops it on a Tuesday?"

"Aaron does."

"Aaron's in Bellingham on Tuesdays."

There was a small laugh, and it moved on, and that was the moment. She had written four things into that line, in the vocabulary she had gone and got for herself: what the owner is for, what counts as it being done, how far the authority runs, and the condition under which everything stops. They had given her the first three. The fourth had turned into a joke and then into an agenda item, which is how a firm declines something without anyone deciding to.²

Afterwards Warren caught her in the corridor with her folder under her arm.

"You'll run it, then. Aaron's name on it, but you'll run it."

"With what authority to stop a matter?"

"Come to Aaron. Or to me." He said it kindly, and he meant it kindly, and he had been saying versions of it since 1999. "You've never needed a title to be listened to here, Glenda."

Which was true, and was the whole difficulty, and she did not say so.

She had one more thing to try, and she tried it. The last sentence of line six, that whoever owned it had to understand it, had gone into the minute unchanged. If Aaron carried the name, Aaron did the learning, and the hours went in the budget where June could see them and nobody could quietly not spend them.³ Warren agreed to that on the spot, without appearing to notice that it was the only part of the afternoon that would still be operating in three years, when the tools had changed again and the transparency, the proportion and the record-keeping were all that had survived.⁴

So: a floor the firm would keep. Divergence standing in the open, unsmoothed. A named partner who would learn. And, for her, the work without the word for it.

Twenty-six years of things arriving on her desk because they belonged nowhere else, and this one had arrived the same way, with no obvious owner, so it came to Glenda, except that for the first time the arrangement had been said aloud in a room with a vote in it, and had been left exactly as it was.

Warren was waiting.

"Will you take it?"

End · Reasonable Doubts