At Andwise, a physician uploaded an employment contract and got back the clauses worth arguing about: the tail-coverage sentence where nobody had assigned the run-off premium, the non-compete radius that reached their next job, the call-duty line that allocated coverage without saying how much. A named reviewer signed the analysis before it went out, on a deadline. About 1,200 physicians used the product.
Then they signed the contract and stopped opening it.
That was the part I understood least well while building it. We had made something physicians trusted. We had not made something they returned to. Those are different products, and the difference turned out to be the whole argument.
What the training installs
The standard story about physician software is that physicians are busy, so the interface should ask less of them. That is true and it is thin. It treats the physician as a user with a short attention span, which is not the interesting fact about them.
The interesting fact is what seven to eleven years of training does to how someone handles a decision. Medical education is an apprenticeship in one move: you receive incomplete information, you commit to a reading of it, someone with standing corrects you, and the correction stays. Morning report. Boards. Morbidity and mortality conference. The attending who asks why you ordered the test.
The loop repeats thousands of times, and someone grades it every time. What comes out the other end looks like a personality — checks the thing everyone else skipped, says the uncertainty out loud, wants the reasoning and not the answer — but it is a trained response to an environment that corrected them.
The loop needs a grade
The transfer usually fails here, and it fails for a mechanical reason.
The loop requires feedback that arrives while the decision still matters, from someone with the standing to say you were wrong. In the hospital that arrives in hours: a lab result, a complication, a senior physician’s correction. In a contract negotiation it arrives years later, attached to a job already accepted, and nobody tells you which clause caused it.
So a physician can be careful and calibrated at the bedside and careless at the negotiating table. That is not a character flaw or a gap in intelligence. The environment stops grading them, and the habit has nothing to attach to.
Supplying what the clinic supplied
None of this says what to build. The move from “the loop needs a grade” to “a product should supply one” is an argument, not a finding. It holds if you accept that the product’s job is to build judgment rather than deliver an answer. I accept it. A product that delivers the answer is faster, and it leaves the physician as dependent as it found them.
Any product that wants the loop has to design around the constraint, and most do not. A content library grades nothing. A dashboard grades nothing. A recommendation engine hands over the answer, which is the opposite of the loop, because it removes the commitment the correction is supposed to land on.
What Andwise got right and what it skipped
We built the honest version. The software surfaced the relevant language and stopped there; a flag was not a legal conclusion. The physician decided whether to accept a clause, raise it, or escalate to a lawyer. We refused the advisor-referral model that would have made the physician the product, and we took no payment for listings in the advisory directory.
The values were right. The mechanism was half-built.
The transaction problem
A contract review is a transaction. One document, one analysis, one signature, and the need is gone. There is no second attempt, no correction, no record of improving judgment. We built the payoff step of a learning loop and skipped the loop.
That explains the payment problem better than “physicians won’t pay.” Physician subscription could not fund growth at any price we tested, because we were selling a one-time review to someone who needed it once. A loop is a different business: it can charge for the accumulated record rather than the single transaction. We never tested that. I do not know that it works.
What a grade looks like outside the clinic
You cannot grade “the right answer” on a contract. Which clause creates the most exposure depends on the employer, the state, the specialty, and how the physician plans to leave. Any product that pretends otherwise is selling an answer key it cannot defend, to an audience trained to notice.
You can grade process, and process is what the training optimized. Did you find the clause. Did you ask the question that changes the outcome. Did you know what you did not know, and did you route it to someone who did. Those are checkable against the document in front of the physician, and they hold up when the answer is genuinely uncertain.
A process grade is also the position I would defend. It is hard to copy, because it requires the reviewer network, the escalation clock, and the record. An answer key is a prompt away.
The relationship the interface implies
An interface does more than carry information. Its wording sets up a relationship: who knows, who it’s helping, who it’s watching.
A consumer product usually plays the guide. Three options, a recommendation, a cheerful confirmation. For someone deciding which subscription to cancel, that is the right shape.
For some physicians, the same shape can land as supervision. Physicians spend years being evaluated by people with more authority, and compulsory quizzes, praise, and gated progression may make decision support feel like supervision and discourage use. The capability is identical. The role is not. This is a hypothesis to test: hold the advice constant and compare uptake when those steps are compulsory with uptake when the physician can proceed or challenge the reasoning.
The alternative is a colleague: assembles the evidence, states the uncertainty, expects an argument. That framing carries a cost the guide framing avoids. A colleague has a name and a license. Software can invite challenge; it cannot absorb responsibility.
So the question is not whether the system seems collegial. It is what happens when the physician overrides it and is wrong, and what happens when the physician follows it and it is wrong. If the answer is “the physician owns both,” say so on the screen, because they will find out.
There is a version of this that fails. Making assumptions inspectable lets a physician check the reasoning. Inspecting every assumption may add work a physician at 2 a.m. with twelve patients cannot spare. Inspection should scale with what the decision costs, not with how sophisticated the user is. Available always, mandatory never.
Two things are true at once
There is a version of this argument that says physicians are sophisticated users, so build for their sophistication. There is another, in How to Design for Cognitive Scarcity, that says assume fifteen percent battery.
Both hold. A physician reading a contract at 11 p.m. after a shift is depleted and also capable of more than a form. The product I would build absorbs the ambiguity and demands the judgment on the same screen: low floor on effort, high ceiling on thinking. Most products pick one and call it empathy.
The motivations that help and the ones that extract
Medical training does not only build calibration. It builds motivations that institutions find useful: fear of being unprepared, comparison against peers, responsibility accepted without control, the sense that self-sacrifice is the professional baseline.
Those are commercially attractive. A product can sell continuous assessment, more credentials, more preparation, and the anxiety that funds them. It would sell. It would also reproduce the arrangement the physician is trying to escape, with a subscription attached.
The line between the two is not a matter of tone. It is a question about what the product does at the moment the user has learned enough to need it less. Does it make leaving easier, or does it make leaving feel like falling behind? An engagement metric counts both answers as a user who came back.
The uncomfortable part
I built the version that made leaving easier. It did not fund growth, and I wound it down in 2024. The extractive version of the same insight looks like the better business on paper: keep the physician anxious, keep them assessing, keep them paying. I would not build it.
Recurring revenue is not the objection. A physician who comes back because the product keeps sharpening their judgment is the business I wanted, and that is the loop I never tested. So whether the physician’s interest and the margin line up here is a claim I can neither support nor rule out.
What I can support is narrower. The training creates a competence that generalizes. The products physicians get mostly ignore it. A product that grades process instead of selling answers can use it without exploiting it.
Whether that is a company is a separate question from whether it is true.
Revision
12 November 2026: retitled from “1,200 Physicians Trusted a Contract Tool, Then Stopped Opening It”, which read as physicians rejecting the tool; they stopped because the contract was signed. The closing section now separates a subscription that keeps the physician anxious from one that keeps improving their judgment. The earlier version set margin against virtue as if those were the only two options.
6 November 2026: retitled from “1,200 Physicians Trusted It, Then Stopped Opening It”; on a card or in search, “it” had nothing to refer to.
29 October 2026: the interface discussion now states the “another attending” explanation as a hypothesis about some physicians and specifies a comparison that could test it. The workload discussion distinguishes access to assumptions from compulsory inspection.