“Read the last line,” Evan said.
Daniel scrolled to the bottom of the opportunity record. Two hundred and forty thousand dollars. Eleven months of conversations. Closed as lost the previous Thursday.
Reason lost: Customer already has Microsoft 365 Copilot. Could not articulate what Auralis does that their existing tools do not.
“I wrote that myself,” Evan said. “I could have written something kinder.”
It was the seventh day of Daniel’s four-week assessment, and Olivia had asked for decisions rather than findings. The executive team had gathered in the boardroom again, though nobody had bothered with the harbor view. The rain had stopped. On the screen, Evan’s pipeline review had reached the part everyone skipped in good quarters.
Eight losses in the last two quarters. Three to a bundled assistant already inside the customer’s enterprise agreement. Two to a general-purpose assistant the employees had adopted without procurement. Three to no decision at all, which Daniel had learned to read as the most honest verdict a market could give.
“Before anyone says it,” Evan continued, “I am not blaming the product. I am telling you that I cannot win a comparison I cannot explain.”
Olivia glanced at the competitor names. “We pay one of those companies to run our product. Then we meet them in the deal.”
“Foundry is the fastest way to reach good models,” Marcus said. “It is also somebody else’s distribution channel.”
Daniel wrote it on the board without comment. It belonged in the strategy rather than in the grievance.
Leila brought up the response marketing had prepared. A comparison grid filled the screen: Auralis in the first column, three competitors in the others, fourteen rows of capabilities. Auralis carried a checkmark in every row. Two competitors carried four each.
The room looked at it for a moment.
“We win,” Marcus said.
“We wrote it,” Daniel said.
Olivia laughed once, without much humor. “That is not a disqualification. Everyone writes their own grid.”
“Everyone does. That is the problem with using one to decide strategy.” Daniel walked to the screen. “Every row here is a capability we can name. Not one is a reason a professional would change how they work on a Tuesday. If our answer to a bundled competitor is fourteen checkmarks, we are asking a buyer to prefer our list to a product their company has already paid for.”
Evan spread his hands. “Then give me the sentence. One sentence. I have four calls this week.”
“I don’t have it yet. That is what today is for.”
Ravi unmuted from the screen, and the frame showed a whiteboard behind him covered in evaluation notes. “I have a version of the sentence. We are better at the work. I can defend that with numbers.”
“Go ahead.”
“Last week we ran the evaluation harness against Northstar’s public policy corpus. Two hundred and forty questions, written with Leila’s team. An answer passed only if it was correct and its citations supported it. With our current model and the relevant documents supplied directly in chat, sixty-one percent passed. With the same model inside Auralis, seventy-nine percent passed.”
Evan sat forward for the first time. “That is the sentence. We are thirty percent better.”
“It is eighteen points on a curated subset,” Ravi said. “I would not put it on a slide yet.”
Daniel wrote the two numbers on the board. “Ravi, how much of the seventy-nine is the model?”
“Meaning?”
“If you swap our current model for the strongest frontier model available through Foundry tomorrow, what happens to seventy-nine?”
Ravi paused long enough that everyone understood he had already run it.
“Eighty-two.”
“And if you keep our model and remove our retrieval, our chunking, our permission filter, our citation verification, and our prompt assembly?”
“The chat setup gives us sixty-one. But that comparison tests the whole system. It does not tell us which parts earned the gain.”
Daniel wrote the third number and stepped back so the room could see all three.
“Three points from swapping the model in this test,” he said. “Eighteen from the system around it. We have spent four months describing ourselves by the part we buy.”
Marcus had been quiet. He closed his laptop halfway, which Daniel was beginning to recognize as the gesture that preceded disagreement.
“That reads better than it is,” Marcus said. “The three points cost us nothing. We buy them. The eighteen points cost four months of engineering and they are the part a competitor can copy.”
“Can they?”
“Retrieval is not a secret. Permission-aware indexing is not a secret. Anyone who wants to build it can.”
“Agreed. The question is what gets better as customers use it.” Daniel underlined the eighteen. “The next model upgrade is available to our competitors too. Their gains will differ, but we cannot count access to it as an advantage. What do we learn from the work that they cannot buy with the same API call?”
“Neither does anyone build one on a retrieval pipeline,” Marcus said.
“No. But we are not talking about a pipeline. We are talking about whether a professional can defend the output to someone who outranks them.”
Olivia stood and walked to the window, which was how she thought when she disliked the direction of a conversation without yet knowing why.
“Say more,” she said. “And be careful, because I think you are about to make us small.”
Daniel had expected the objection and respected it. Olivia had raised eighteen million dollars on a platform story, and the next financing would not be easier. Victor’s question from the board meeting had not changed: was Aureline building a feature or a durable company? A narrow answer could settle it the wrong way.
“Then let me take the risk out loud,” Daniel said. “I am not proposing we shrink the product. I am proposing we choose the comparison we intend to win, and accept the comparisons we intend to lose.”
“Those are the same thing to a buyer.”
“Only if we choose work the buyer does not value. Right now we lose comparisons we never intended to enter. We are compared with a bundled assistant on price, because we sound like a bundled assistant. We are compared with a general assistant on conversational quality, because we demonstrate conversation. We do not get compared on the work the customer cannot currently do at all, because we have never made that the subject of the conversation.”
“Which work?”
He did not answer immediately. He had a hypothesis and one week of evidence, and the room deserved to know which was which.
“That is the decision I want today,” he said. “Not the wording. The subject.”
Rachel Stein arrived eleven minutes late with a laptop under one arm and the expression of a general counsel who had spent the morning reading something instructive.
“I read Northstar’s questionnaire,” she said. “All of it.”
Two hundred and fourteen questions had arrived the previous afternoon, attached to a message that had also moved the meeting up. Daniel had assumed it would become a security project. Rachel had treated it as market research, which had not occurred to him.
“Here is what they ask about,” she said. “Six questions ask which models we use. Five ask about model quality or benchmarks. Eleven ask whether a person can be prevented from seeing something. Nineteen ask about retention, deletion, and where data physically lives. Twenty-three ask how an answer can be traced to its source. Thirty-one ask what happens when it goes wrong—incident response, notification, rollback, liability. The rest are identity, isolation, subprocessors, audit history, and export.”
Evan said what everyone was thinking. “Eleven out of two hundred and fourteen are about the thing we argue about every day.”
“Claire already has access to intelligence,” Rachel said. “Her lawyers use a general assistant every day. These questions suggest she is buying the ability to put her name on what comes out of it.”
Leila had been writing while Rachel spoke. She turned her notebook around.
The work a professional has to sign.
“That is the job,” she said. “Not search. Not chat. Not summarization. A senior professional produces something another person relies on—a policy analysis, a regulatory position, a recommendation to a committee—and their name is on it. Today they can get a draft from any assistant in four minutes. They still spend two days checking it, because they cannot defend a citation they did not verify themselves. We are not competing with the four minutes. We are competing for the two days.”
The room was quiet in the way that meant something had landed.
Daniel wrote it on the board.
We compete for the verification, not the draft.
“Then say what that costs us,” he said. “A position is not real until it removes something.”
He worked through it with them, and it took longer than the framing had.
If Auralis competed for defensible professional work, then general conversational breadth was not a battleground. Aureline would lose to a consumer-grade assistant on writing a birthday message and would stop treating that as a defeat.
If the differentiator was evidence, then training a foundation model was out of scope. Aureline would buy capability through Foundry and approved providers and direct its engineering budget toward evidence and verification. Marcus had never proposed training a foundation model, but the roadmap had carried an ambiguous line about model specialization for two quarters, and it came off.
If the differentiator was the traceable path from source to statement, then agent autonomy could not run ahead of it. Controlled action stayed proposal-and-approval until the audit record was as strong as the answer.
And if the pilot was Northstar, connector breadth was not the launch story. Two connectors done with permission fidelity beat nine done with a service account.
“Which brings me to the conference,” Daniel said.
Sofia, who had not spoken, looked up.
Her team had spent three weeks of evenings on an assistant builder for the October demo—a gallery of templates, an animation Daniel had to admit was beautiful. It was the most impressive thing Auralis could show a stranger in ninety seconds. It also demonstrated breadth, which they had just decided was not the subject.
“I am not cutting it,” Daniel said. “I am asking you to change what it proves. Same stage, same ninety seconds. Instead of nine templates, show one analysis and let someone in the audience try to break the citation.”
“That is a smaller demo,” Sofia said.
“It is a harder one.”
She considered it. “My team will hear that three weeks were wasted.”
“Then tell them what I am telling you. The builder is not wasted. The claim it was built to make is the one we are retiring. If I am wrong about that, it is my mistake and not theirs, and I will say so on the same stage.”
Sofia wrote something down and did not look entirely convinced, which Daniel took as an accurate reflection of the evidence available.
Olivia returned to the table. “I will agree to this with two conditions.”
“Go ahead.”
“First, narrow is a starting position, not an identity. I am not going to describe Aureline to investors as a citation company.”
“Describe it as the company that makes AI usable for work that carries consequences. That is not smaller. It is only more specific.”
“Second, I want the proof inside the pilot. Not a promise that evidence will eventually differentiate us. A number Claire’s own team produces, on their own corpus, that they would repeat to a peer.”
“That is the right condition,” Daniel said. “It is also the one that can fail. If Northstar measures us and we are not materially better on the work they actually sign, the position is wrong and we will have spent the pilot learning it.”
“Then we will have learned it in eight weeks instead of eight months,” Olivia said. “Write it down. One page. I want the version Evan can read in a hallway.”
Evan was already typing. “Give me the question to ask, not the claim to make. That worked last week.”
Leila answered without looking up. “Ask them: what is the last thing you personally check before your name goes on the analysis? Then ask how long it takes. Then ask what would have to be true to trust a system with it.”
“And if they say they would never trust a system with it?”
“Then we have found the boundary of the product, which is worth more than the deal.”
The meeting broke up at half past one. Daniel photographed the board and wrote the page himself before the conviction faded, because strategy written the next morning tends to come back softer.
Marcus was waiting in the corridor.
“You just told the company that our advantage is the layer between the retrieval index and the model,” he said.
“I told the company that is where we have to prove it.”
“It is the weakest part of the system. Context assembly was written in the second month, before we had tenants.” Marcus was not being defensive; he was being precise, which was worse. “If the strategy depends on it, somebody should rebuild it properly.”
“Somebody will. With Hannah and Ravi, and with a design the managers can staff.”
Marcus nodded and went back toward his office.
Two mornings later, Daniel’s repository feed carried a merged pull request against the context-assembly service. Forty-one files. A new retrieval stage, a rewritten prompt builder, and a schema change to the citation payload.
Opened by Marcus Hale. Merged by Marcus Hale. No reviewer.
The advantage they had chosen two days earlier was already being rewritten by the man who had built it.
Most competitive analysis begins with a grid of vendors and capabilities. It feels rigorous, and a company that builds its own grid wins its own grid.
The useful question is not which product the customer compares you with, but what they are trying to accomplish, what they do instead today, and what moment forces them to choose anything at all. The alternatives are rarely only other products: the tool already bundled into an existing agreement, the assistant employees adopted without asking, an internal build, and doing nothing—usually the market leader. A position that cannot explain why the customer abandons their workaround is a feature list.
Every AI product is assembled from layers with different supply characteristics, and confusing them is how a company spends its scarcest year on ground it does not own.
| Layer | Who supplies it | What erodes the advantage | What can compound |
|---|---|---|---|
| Frontier model capability | External model providers | Competitors can buy access to the same releases | Access alone offers little lasting advantage |
| Infrastructure and hosting | Cloud vendors, often also competitors | Shared services make basic capability widely available | Your operating expertise and efficiency |
| Context, retrieval, and evidence | You, shaped by customer data | Competitors can reproduce the techniques | Evaluations, failure analysis, and improvements from real use |
| Governance, audit, and trust posture | You, validated by customers and auditors | Controls become standard; failures can undo trust | Operational history and customer references |
| Customer knowledge and workflow fit | Built with the customer | Workflows change and competitors learn | Knowledge earned through repeated delivery |
A model upgrade may improve every product differently, but access to the upgrade is rarely exclusive. The strategic question is what remains distinctive once competitors have it too. Buy shared capability deliberately. Invest in the customer knowledge, evidence, and operating experience that make your use of it harder to reproduce.
Strategies are easy to adopt because most cost nothing to adopt. A statement about where a company will win becomes real only when it names what the company will lose. Test any proposed position against four questions:
If the answers are empty, the organization has adopted a slogan, every team keeps optimizing what it already believed, and the product expresses the average of those beliefs rather than a decision.
AI lowers the cost of understanding a market. It can cluster win and loss reasons across a CRM, extract the requirements buried in a security questionnaire, compare competitor documentation against your own claims, and turn customer interviews into candidate positions. It can also help design controlled comparisons and analyze evaluation results to test how much quality changes when you swap the model or remove parts of the surrounding system. Those comparisons need consistent inputs and grading; a single aggregate score cannot establish which component caused a gain.
It cannot supply the comparison you are willing to lose. That depends on ambition, runway, the board’s appetite, and an honest reading of the organization’s capability. A model can propose a plausible strategy from the reasoning of companies in other positions and other years. The leader still has to test whether that reasoning fits this company, these customers, and the time available.
Humans must decide:
These are commitments about identity and consequence rather than analyses. They determine what the organization stops doing, and stopping requires authority rather than intelligence.
One page, written by the accountable leader rather than by committee, with a review date chosen in advance.
Wrightery
COMPETITIVE POSITION RECORD
Position owner:
Date written:
Mandatory review date:
1. THE JOB WE COMPETE FOR
Professional who does the work:
What they must defend, and to whom:
What it costs them today:
2. THE REAL ALTERNATIVES
Bundled tool already paid for:
Assistant already adopted without procurement:
Internal build, manual process, or doing nothing:
Why each is chosen today:
3. WHAT WE MUST BE TRUSTED WITH
Outcome the customer relies on:
Proof demanded before relying on it:
4. WHAT WE REFUSE TO COMPETE ON
Comparisons we concede:
Capabilities we will not build:
Roadmap items removed this quarter:
5. WHERE THE ADVANTAGE COMES FROM
Rented layers, and what the supplier's next release does to us:
Owned layers, and how they compound:
6. EVIDENCE AND FALSIFICATION
Measurement supporting the position, and who produced it:
Customer workflow and evaluation scope:
Measurement the customer would repeat to a peer:
Required improvement over the current alternative:
Evidence owner and due date:
Result that would prove the position wrong:
Who reports it, and who may change the position:
The record is not a marketing document and does not need to be persuasive. Its purpose is to make one expensive decision visible enough that the organization can act on it, argue with it, or discover it was wrong while the company can still afford the correction.
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 ↗