“The eighteenth of November,” Claire said.
She did not put it on a slide. She wrote it on the whiteboard in Northstar’s small conference room on the thirty-first floor, in green marker, underlined it once, and sat down as though the meeting could now begin.
Daniel had flown to Toronto the night before with Leila, Evan, Rachel, and Hannah. Omar Haddad had taken the earlier flight and spent his evening in the hotel reading the questionnaire for the third time. Across the table, Claire Dubois had brought one person: Grace Adeyemi, Northstar’s director of information security, who had a printed copy of Aureline’s answers to the first forty questions and a pencil she had already used.
“Our regulatory committee meets four times a year,” Claire said. “On the eighteenth it reviews the firm’s position on the new federal privacy rules for three clients. A partner signs that position. If Auralis is in this pilot, the analysis in front of that committee is the first thing prepared with it, and his name is on it.”
“So the eighteenth is the end of the pilot,” Evan said.
“The eighteenth is the only day in the pilot that matters.” Claire looked at Daniel. “Everything before it is preparation, and everything after it depends on how it went. Thirty-eight lawyers in my regulatory group use Auralis for the eight weeks ending that day. One analysis goes to the committee. Nobody has to apologize for it.”
Daniel had spent three weeks arguing that a date meant nothing without a definition. He had not expected the customer to arrive with both.
“What would make someone apologize?” he asked.
Claire turned to Grace, who had been waiting.
“Anaïs Roy is a second-year associate on the Halcyon matter,” Grace said. “Two floors down, we act for Bramwell, which competes with Halcyon. There is an ethical wall between the two teams. Anaïs may not see a Bramwell document, and if she ever did, we would have to tell both clients and possibly withdraw from one of them.” She set the pencil down. “If Anaïs asks Auralis about the market both companies are in, can it cite a Bramwell memo?”
Daniel looked at Hannah, because it was Hannah’s system and because he did not know.
Hannah did not answer at once. He watched her trace the path in her head: the connector reading iManage through a service account, the documents chunked and indexed into Northstar’s partition of Azure AI Search, the workspace boundary that decided what a user could query.
“Today, if both matters are in the same workspace, yes,” she said. “Our isolation is between customers and between workspaces. Inside a workspace, retrieval knows nothing about your walls. The connector reads with one identity, and everyone in the workspace can see what it read.”
Evan began to say something and stopped.
“Thank you,” Grace said, and wrote. “That is the first honest answer I have had to that question from a vendor this year.”
“It is also the wrong answer for a pilot,” Claire said.
“Yes,” Daniel said.
Claire went through the rest of her list the way Daniel imagined she ran a matter meeting: one item, one owner on her side, one question. Her lawyers signed in through Microsoft Entra ID, and when one of them left on a Friday afternoon, access ended on Friday afternoon, not whenever somebody at Aureline noticed. Her general counsel needed to know who had asked what and which documents each answer had shown them, because if a privileged document surfaced in the wrong place, the firm would have to prove where it had gone. Her partners needed a citation they could click in February that still opened the passage it had pointed to in November. Grace needed a name and a telephone number for seven o’clock on the morning of the eighteenth, Toronto time, if anything misbehaved while the analysis was being finished.
“That is four in the morning in Vancouver,” Evan said, and then visibly wished he had not.
“I know where Vancouver is,” Claire said mildly. “The last item is the one you will think is small. If you change the model underneath my lawyers in week five, I want to know before they do. Not afterward. Before.”
Evan had held back for an hour, which Daniel knew had cost him. He leaned forward now.
“We can commit to all of that for the eighteenth,” he said. “Identity is in place. We have audit. Hannah’s team can extend the permissions. We will staff support.”
Daniel felt the room settle on the sentence the way a room settles on a signature. Leila was looking at the table. Rachel had stopped writing.
He thought of a box on Elena’s wall, ingest-halvorsen, four hundred and twelve commits behind the product, and of the contract signed on a Friday that had made it.
“May I ask the question the other way?” he said to Claire. “If on the eighteenth we had done everything on your list except one item, which one would end the pilot?”
Claire did not need long. “The wall. I can manage around the others. I can tell a partner that support was slow or that a citation needed a second click. I cannot tell a client that a lawyer on the other side of a wall read their memo through a tool I chose.”
“Then that is what we promise first,” Daniel said, “and we prove it before the eighteenth, not on it. I would like to send you a written promise by Monday. Each item on your list, what we commit to and by which date, how your team can check it, and what we are not offering in this pilot. Nothing in it that exists only for Northstar.”
Grace looked up. “Why does that last part matter to us?”
“Because if we give you a separate implementation without a team responsible for keeping it current, you will be depending on something the product can leave behind,” Daniel said. “We have done that once. I would rather tell you now what we will not do.”
Claire studied him for a moment and then capped the green marker.
“Monday,” she said. “And Daniel—I would rather have six honest lines than twenty hopeful ones. My partners can tell the difference.”
Evan waited until the taxi was on the Gardiner before he said it.
“I had the deal in the room.”
“You had a yes in the room,” Daniel said. “I am not sure it was the deal.”
“She asked for six things and I said yes to six things. That is what customers want to hear. Now she gets a document with a column called not offered.” Evan watched the lake slide past, flat and grey. “You know what Victor says about our pipeline? That we are very good at first meetings.”
“I know what happened to Halvorsen when they got everything,” Daniel said. “They have their own ingestion worker, and nobody has deployed it since June. Somebody said yes to them on a Friday.”
Evan was quiet for a block. “That was before me.”
“It was. That is my point. The person who makes a promise is rarely the person who keeps it. If you and I write this one together, we are both.”
Evan almost laughed. “You are recruiting me.”
“I am asking you to sign it with me. If it says only what we can prove, you can sell it without flinching.”
Leila spoke from the front seat without turning around. “And if it says only what we can prove, the second customer gets it too. That is the only way a design partner does not become the product.”
On the flight home, Leila gave an assistant the questionnaire, Grace’s marked-up answers, and the transcript of the meeting, and asked it for a requirement map: every distinct thing Northstar had asked Aureline to do or prove, with duplicates removed, grouped by kind, and the question numbers behind each item. It returned sixty-one requirements in six groups—identity, isolation, evidence, audit, support, and governance. She spent the rest of the flight checking it against the questionnaire and found three requirements it had merged that should have stayed separate, and one that no question had asked for.
By Friday morning the map was on the wall of the small room, beside the photograph of Elena’s drawing.
Hannah took the wall first.
“It is two problems,” she said. “The first is the known one. The connector stops reading as a service account alone. It also reads each document’s access list from iManage and stores the group identifiers with every chunk. At query time we resolve the user’s groups from Entra and pass them to Azure AI Search as a security filter, so a Bramwell chunk never comes back for Anaïs at all. Not hidden by the interface. Never retrieved. Three weeks, with two of my people.”
“That leaves us a few days before the eight weeks start,” Daniel said. “The wall passes our tests before any lawyer gets access. If it does not, we tell Claire the pilot cannot start on time.”
“And the second?”
Ravi answered from the screen. “The second is mine. To keep model cost down, we cache generated answers. The key is the normalized question and the workspace. The hit rate is fourteen percent.” He paused, and Daniel watched him see it as he said it. “If Anaïs asks a question somebody on the Bramwell side asked yesterday, she gets yesterday’s answer. With yesterday’s citations.”
Nobody said anything for a moment.
“Can the key carry the permissions?” Daniel asked.
“It can carry a fingerprint of the user’s effective access. That is the right fix and it is not a three-week fix. For the pilot I would rather switch the cache off in any workspace with walls and pay for the misses.”
“Then switch it off,” Daniel said. “And put the fingerprint in the plan, with a date, so that it does not become another sentence that means never.”
“There is a third,” Hannah said. “similar-docs. The panel queries the vector store directly. It has never heard of a filter.”
“Its retirement was set for next month.”
“Its retirement is now this week.”
Priya had been listening with her laptop closed, which Daniel had learned meant she was assembling something.
“Then the wall needs a test, not a review,” she said. “Build two synthetic matters with a wall between them, using real documents from the public corpus. Generate a few hundred questions designed to cross it—paraphrases, the other client’s name, the cached path, the similar-documents panel, the export, a conversation shared across the wall. Run all of them on every deploy. Any document from the wrong side, in any answer, fails the build.”
“Who writes the questions?”
“An agent writes most of them, because it will think of phrasings we will not. My team writes the twenty it will not think of, because those are the ones a curious associate will try.”
Omar took identity. Entra sign-in worked. Provisioning did not: accounts were created when people first logged in and removed when somebody at Northstar remembered to file a ticket. SCIM provisioning from Entra would close it, and Omar thought it could ship in four weeks. Until then, a departing lawyer kept access for as long as their last token lived.
“Then the promise says so,” Rachel said. “Until a date, removal is manual and takes up to an hour. After the date, it follows Entra. Grace will respect the first sentence a great deal more than a pretense.”
Noah took support, and he had clearly been thinking about the early start since hearing about the meeting.
“Our regular support covers seven in the morning to five in the afternoon, Pacific. Northstar starts three hours ahead of us, and committee prep can run past our closing time. I can add a pilot rotation from four to seven in the morning, Pacific, for eight weeks, and an on-call extension until ten in the evening, Toronto time. Three engineers, paid, with an escalation path that reaches a manager in fifteen minutes. On the eighteenth itself, the early shift starts an hour sooner, with one of us on a call with Grace from six, Toronto time.”
“Who?”
Noah hesitated just long enough for Daniel to see what he was about to say.
“Not you,” Daniel said. “Whoever is on the rotation that week, and the rotation is in the capacity plan, not on anyone’s Saturday.”
Noah did not argue, and Daniel noticed that he looked relieved.
Sofia took evidence. The citation migration from the previous week gave her a starting point. Before the pilot opened, Grace would receive a set of citations tied to document versions and passages; Sofia’s team would rerun them after every index rebuild and keep them resolvable through February. A source the user could no longer access would return an access-denied message, not its text. Omar took the audit check as well as identity: before the first lawyer signed in, Grace would run a scripted session and reconstruct its questions, users, and sources from the audit export. Rachel asked him to include a denied request. Grace needed to see what the system had refused as well as what it had shown.
The cost arrived at the end, as Daniel had known it would. Hannah could build the wall or consolidate the Halvorsen fork in the next six weeks. She could not do both with the people she had. Daniel moved Halvorsen to the first quarter and told Evan he would need to explain it to Halvorsen himself, with Hannah beside him. Controlled actions came out of the pilot entirely: Auralis would draft the recommendation and prepare the action for review, as the reviewed claims had said in September, and it would not execute anything in Northstar’s systems. The other exclusions were connectors beyond iManage and SharePoint, a separate Northstar ingestion worker, and round-the-clock support. Ravi pinned the model deployments in Foundry and wrote the change rule himself: ten business days’ notice before any model or prompt change, with the Northstar evaluation results attached.
Daniel and Evan wrote the promise on Saturday. Rachel edited it on Sunday with a single rule, which she wrote across the top of the draft in red: Every “will” has a date and a check. Anything without a check is “intend.”
It came to eleven commitments and four exclusions. The wall came first. Beside it, under How Northstar can check, Daniel had written a line he expected Grace to read twice: Northstar may attempt to cross the wall at any time during the pilot. Aureline will treat any success as a severity-one incident.
Claire replied on Monday at nine-forty, Toronto time.
She had changed one line. She had moved Northstar’s own wall test from the third week of the pilot to the first. Daniel checked the schedule with Hannah before accepting it: their internal tests still had to pass before the pilot opened, and Grace would challenge the wall herself before the wider group began using it.
If it fails, she had written, I would rather know before my lawyers have learned to like it.
Below that, her signature, and Grace’s.
Evan forwarded it to the sales channel with no comment except the words eleven lines. Daniel was still reading it when Olivia’s message arrived.
Good. Victor wants the launch date at Thursday’s board prep. Not the pilot. The launch.
He looked at the promise on his screen: one date, eleven commitments, each with a check beside it. It had taken a week to write a promise for thirty-eight lawyers. On Thursday, the board would want one for the company.
The deadline a board sets is usually a guess about the market. A customer’s deadline is anchored in an event inside their own work—a committee, a filing, a renewal, an audit—and that event tells you what success means far more precisely than any internal launch definition. The useful move is not to accept or resist the date but to find the event behind it and ask what would make it go wrong. Claire’s date was not negotiable. The scope of the promise could be negotiated once they understood what that date required.
The first customer for a newly defined release can make its requirements feel like a custom project. Those requirements help turn the product definition into a specification. Some become capabilities other customers can use; others need configuration or work specific to the customer. The danger is a separate implementation with no owner, no maintenance plan, and no way to follow the product as it changes. That can be faster this month and become an abandoned fork the month after.
| Area | What the customer asks | What it actually tests | Promise as product | Promise as fork |
|---|---|---|---|---|
| Identity | “Does access end when someone leaves?” | Account lifecycle | SSO and SCIM provisioning for every tenant | A manual script for one customer |
| Isolation | “Can it cross our ethical wall?” | Permissions inside a tenant | Document access enforced at retrieval, cache, and every derived store | A hand-configured workspace per matter |
| Evidence | “Will this citation open in February?” | Durability of the evidence path | Resolvable citations with a regression check | A frozen export of the answers |
| Audit | “Who saw what?” | Reconstruction after the fact | A record of each question, user, and source shown | A log extract on request |
| Support | “Who do I call at seven in the morning?” | Operating commitment | A defined window, a paid rotation, an escalation path | A founder’s mobile number |
| Governance | “Tell me before you change the model” | Change control | Pinned deployments, notice, evaluation results | A promise not to touch anything |
Not every requirement carries the same consequence. Asking “which one would end the pilot?” separates the items a customer can manage around from the one they cannot explain to their own clients. That core deserves different treatment: a test that runs on every deploy rather than a review before launch, and an open invitation for the customer to try to break it early. A failure in week one is information. A failure on the day that matters is a broken promise.
The core also exposes where the risk really lives. Permission filters on the search index were the obvious work. The answer cache and a forgotten convenience panel were the actual paths across the wall, because each was built after the access model and never learned about it.
AI can turn a two-hundred-question security questionnaire, a customer’s annotations, and a meeting transcript into a deduplicated requirement map with traceability to every source question in hours rather than days. It can draft answers that link to evidence, propose the adversarial queries that test a boundary, and check a draft commitment for the missing date or the missing test. In a promise like this one, it is the fastest way to make sure nothing the customer asked was lost.
It is also confident about structure it inferred. The map merged requirements a lawyer would keep separate and invented one nobody had asked for. Treat its output as a draft to be reconciled line by line against the customer’s own words, because the customer will read it that way.
Humans must decide:
The person who signs a customer promise is accountable for it long after the meeting, and is rarely the person who will keep it. That is the reason to write it with the people who will.
One page per pilot or first enterprise commitment, written jointly by the accountable R&D leader and the commercial owner, reviewed by legal, and signed by the customer sponsor.
Wrightery
PILOT PROMISE
Customer: Sponsor:
Aureline owners (R&D / commercial / legal):
Pilot dates: Event that defines success:
Users in scope: Work in scope:
1. THE CORE
The one commitment that ends the pilot if it fails:
How it is proved before the event:
How the customer may test it, and from when:
2. COMMITMENTS
For each:
Commitment:
Date: How the customer can check it:
Owner: Status today (live / building / manual until ____):
3. NOT OFFERED IN THIS PILOT
Capability: Reason: Earliest reconsideration:
4. OPERATING TERMS
Support window and time zone:
Escalation path and response time:
Change notice for models, prompts, and retrieval:
Incident definition and notification:
5. PRODUCT TEST
Commitments that exist only for this customer (target: none):
Product work this promise moves, and what it displaces:
The promise is worth more for its exclusions than for its commitments. A customer who sees what you will not do believes the rest. A team that wrote the exclusions knows exactly what it has agreed to keep.
Wrightery is in private build. Join the waitlist and we'll bring you in as soon as a seat opens — then describe one deliverable you already produce, and watch your AI analysts research it, cross-check it and draft it for your review.
No credit card. We'll email you the moment your seat is ready.
Rather talk it through? Book a 30-minute meeting ↗