Walk me through how you would convince a skeptical enterprise buyer that your AI feature is reliable enough to trust.

Indemnis is Wrensmoor's AI tool for reading enterprise supply and vendor contracts: it flags the clauses that could actually cost real money, before anyone signs. Zell Kase owns its trust and reliability story. Ninety days into a paid pilot, Kestrevale Steel's head of vendor risk, Cateline Feathering, closes her laptop and asks the one question that decides whether the deal happens: why should she let a model anywhere near a contract that could cost her company millions.

How to answer this, stage by stage

Nobody is grading whether you can sound confident. They're grading whether you can hold up under a reviewer whose whole job is to find the hole in your story, and whether you'd hand her the hole yourself before she finds it.

Let's learn

Indemnis reads a contract before anyone signs it and tells a legal team, in plain words, which clauses could actually hurt them.

One contract reads like one document. It's really five different kinds of promise stacked together, and only some of them can bankrupt you if they're wrong.

Before Indemnis, Kestrevale Steel's six in-house contract attorneys and paralegals read every supply and client contract by hand, about 1,100 of them a year, at roughly 3.5 hours each. That's real cost: about $330 a contract once you count the attorney time, and a backlog that made vendor onboarding take close to two weeks. With Indemnis reading first and flagging what needs a real look, the same team clears most contracts in under an hour, and the average cost per contract drops to about $95, blended across the ones that clear fast and the ones that still get a full attorney review.

Five steps. The fourth one, checking value and clause type separately from the score, is the step the whole trust case actually turns on.

Measured against 240 real contracts that were later confirmed as disputed, in court or arbitration, Indemnis misses a genuine liability or indemnification issue about 2.5 percent of the time. For comparison: one attorney doing a single fast first read alone, no second check, misses about 4.1 percent of the same set. Indemnis is already a little better than a rushed human on its own. That's not why Kestrevale should trust it.

Three separate tripwires, any one of them enough on its own. A contract only clears itself when none of the three fire.

Here's the turn. The scary version of this was never Indemnis missing a clause now and then. It's what happens after: if Cateline's team gets spooked by one story like that and forces full manual review across every contract regardless of size, Kestrevale is right back to $330 a contract and a two week backlog, and the entire reason to buy Indemnis disappears.

What I would leave alone: routine purchase orders under $50,000 written on Kestrevale's own standard template don't need any of this. Indemnis clears those fast, and it should, because the boilerplate there hasn't changed in years and nobody has ever disputed a line in it.

The lesson: a reliability story that only reassures isn't evidence. A buyer who's heard a pitch before can tell the difference between "trust us" and "here's the number, here's the gate, and here's what would make me take it back." Only one of those survives a hard question.

Now here is the same thing as a story

The short version above is what you actually say in the room. Read this one when you want to feel why a clause on page fourteen of an appendix nearly mattered more than the whole rest of the deal.

Osaze Marnier has managed Kestrevale Steel's supply contracts for nine years, and she has one habit nobody ever had to teach her: she always flips to the last schedule in any appendix first, because that's where the surprises hide. Payment terms and delivery windows sit up front, easy to write, easy to check. The strange language always ends up on page twelve or fourteen, tucked in after everyone's attention has already left the room.

Indemnis arrived for the pilot in the spring, and for the first two months it was the best thing that had happened to her team's mornings. It read a contract overnight and handed back a scored draft by eight, flagging the handful that needed real attention out of the dozens that came in each week. Osaze kept her old habit at first, reading every appendix in full anyway, the same way she always had, checking Indemnis's high scores against her own eyes.

The scores kept agreeing with her. Week after week, nothing she found contradicted what Indemnis had already flagged clean. So the full read thinned, a little at a time, the way habits do when nothing punishes them. By week eight, she was only reading appendices in full on contracts Indemnis scored below 90 percent. Above that, she skimmed the summary and moved on. Nobody decided this on purpose. It just kept being fine, until the week it wasn't.

Contract 187 was a supply agreement addendum with Kestrevale's largest single alloy vendor, worth just under $1.9 million, a hair under the value gate that didn't exist yet. Indemnis scored the whole document 96 percent sure. Osaze almost let the appendix go with a skim, the way she'd been doing for six weeks by then. She caught herself flipping to section 14.3 anyway, out of the decade-old habit her system had quietly stopped asking of her.

A ten year old habit, not luck, closed a gap the scoring system hadn't been built to see yet.

Section 14.3 read, in the flat language addenda always use, that Kestrevale's remedy for a defective shipment was limited to replacement value only, notwithstanding any other provision in the agreement. On a deal that size, a genuine defect claim downstream, say a batch of alloy that failed in a customer's own product, could easily have run past $4 million in consequential damages Kestrevale had never meant to accept. Scored on its own, that one clause would have come back at 61 percent. Blended into the rest of a clean, well-drafted document, it disappeared into a 96.

Osaze caught it three days before signature, called the vendor's counsel, and got the clause rewritten to match Kestrevale's standard cap. Nobody outside her own team ever knew it happened. She wasn't lucky. She was doing the thing she'd always done, on a system that had quietly stopped requiring it of her.

The decision Zell would take back traces to a kickoff call ten weeks earlier. Someone on the scoping call had asked whether the gate should route by clause type as well as by score. The answer, reasonable at the time, with a launch date two weeks out, was to start with one blended number and add more gates later if they turned out to matter. That was a fine call for a two week setup. It was never built to survive a document that could be genuinely excellent everywhere except one paragraph.

Run Contract 187 again with the gate rebuilt the way it works now, scoring liability, indemnification, and consequential-damages language on its own, separate from the rest of the document, regardless of the overall score. Section 14.3 comes back at 61 percent, well under the 92 percent line, and routes to an attorney automatically, in the same overnight run that produced the draft. Nobody needs to remember to flip to page fourteen. The system does it before the document ever reaches a desk.

That's the story sitting behind the ninety day mark. And the ninety day mark is where Cateline Feathering, Kestrevale's head of vendor risk, closes her laptop and asks Zell the question this whole page is built to answer. Zell doesn't open with the accuracy number. He opens with the position, tells her about Contract 187 without being asked, and ends on the exact number that would make him agree with her if she decided not to trust it. Cateline asks two follow-up questions. She gets straight numbers back for both. Kestrevale signs the three year deal two weeks later, with the gate written into the contract as a requirement, not a feature.

The whole answer to this question, in one picture. A confident average is not the same thing as a confident answer about the one part that could actually hurt you.

What Zell would tell himself, back on that kickoff call: starting with one blended score wasn't careless. It was the sensible call for a two week setup with a deadline. It just wasn't a call anyone had agreed to revisit once the model got good enough that nobody had a reason left to check behind it.

PICK: the four things Cateline needed before she'd believe any of it

Not a way to sound confident under pressure. PICK is what forces a real commitment about which number this claim actually stands on, and which kind of wrong Cateline should actually be afraid of.

One box is small on purpose. The other one is drawn the size it actually costs.

Three things worth stating directly, since this is where the real judgment sits. The alternative Zell's team considered, and rejected, was simply raising the overall blended confidence bar higher, say from 90 to 98 percent, instead of building separate per-clause gates. It lost, because Contract 187 scored 96 percent blended. A higher blended bar would still have let it through; the flaw was never the height of one number, it was scoring the whole document as a single average when one paragraph inside it needed its own answer. The AI-specific failure worth naming by name is exactly that: a document-level confidence score can look excellent while one specific clause inside it is genuinely a coin flip, and nothing about a blended average tells you which clause that is. The guardrail is a risk-class rule, not a smarter threshold: liability, indemnification, and consequential-damages language get scored and gated on their own, at any value, and any contract over $2 million gets a person's eyes regardless of what every clause inside it scored. And the trade-off is real and accepted on purpose: a gated contract takes two to three business days for an attorney's own read instead of clearing overnight, which is slower and costs real attorney hours. Kestrevale accepts that cost on purpose for the slice of contracts where a miss can't be undone once a shipment has already gone out the door.

And if you want to be sure it really works, try it somewhere else

Same four letters, a school district instead of a steel mill, and this time the fragile clause isn't a damages cap. It's who pays when a child gets hurt on a piece of playground equipment.

Recital, built by Tanager Compliance, reads facilities and vendor contracts for public school districts: playground equipment maintenance, bus fleet servicing, food service vendors, the contracts a district signs every year without a legal department the size of a steel company's. Branimir Sorrelfield, Cadenwick Unified School District's director of risk and business services, is deciding whether to renew Recital past its pilot year, and he's already skeptical for a plain reason: a school district can't afford to be wrong about who's liable when a child gets hurt.

Different building, same shape of question. A steel mill's damages cap became a playground vendor's liability cap, and the gate logic transferred without changing.

Recital's own eval set is smaller, 95 confirmed disputed facilities and vendor contracts pulled from insurance claims and litigation across public-sector clients, and its measured false negative rate on liability and indemnification language runs a little higher, about 3.1 percent, honestly reflecting the thinner sample. The gate is built the same way, sized to a school district's own budget: any contract worth more than $250,000, any personal-injury liability or indemnification clause, or any insurance-certificate requirement, at any price, routes to Cadenwick's own risk counsel automatically.

The near miss that made Branimir a believer wasn't caught by a decade-old habit the way Osaze's was. It came from a new contracts coordinator three weeks into the job, reading a playground equipment maintenance vendor's contract for the first time with no old habits to fall back on. She asked a plain question nobody senior had thought to ask in years: why did this vendor's liability cap read so differently from the district's usual template. It turned out the cap, buried in a special-conditions page, limited the vendor's exposure to the cost of the equipment itself, nowhere near what a real injury claim could actually run.

Same rank, different lever, mapped straight onto PICK: the position is the same, measured and gated, not claimed perfect. The impact splits the same way, a risk officer who over-restricts loses the whole point of the tool, a coordinator who trusts a wrong clear loses the district's own protection. The cost asymmetry lands on the same kind of clause, just smaller dollars: a routine supply order missed is cheap, a liability cap missed on a facilities contract is the one that can't be undone once a child is actually hurt. And the kill criteria transfer directly: a false negative rate on liability language crossing roughly double today's baseline, or a model swap without a fresh validation report, both mean stop trusting it until it's proven again.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: measured, not perfect, gated by value and clause type regardless of score, and here's the number that would make me pull it back.
Cost: no budget this quarter to widen the eval set. Ship the gate as it stands, labeled honestly as built on a smaller sample, and grow the eval set with every new confirmed dispute instead of waiting to launch until it's larger.
The model got better, for real: say Indemnis's false negative rate drops to 1 percent. The gate doesn't loosen. A lower rate still isn't zero, and the contracts where a miss can't be undone haven't gotten any less serious just because misses got rarer.

Where people run it wrong.
They lead with the accuracy number because it's the one that sounds most impressive, and let a skeptical reviewer poke a hole in the one claim that was never really the point.
They treat a near miss as proof the whole tool is broken, instead of proof of exactly which gate was missing.
They answer a scare by turning the tool off everywhere, instead of narrowing the gate to where the real risk actually lives.

How to use it live. Before naming any number, ask yourself out loud what the reviewer would ask next: "which mistake here can't be undone, and does that one ever reach someone without a person checking it first?" That question alone buys real thinking time, and it's usually the exact distinction the interviewer is listening for.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0