Verify the AI decision before you sign.
This is the whole method to check an industrial AI decision before you sign the vendor or scale the pilot — the four moves that separate a bet you can defend from one you can't, and five tools that turn each move into a page your own team fills in. It's yours. No email, no gate, no sales call to unlock it.
why this exists
Almost every AI assessment a mid-market operator sees is a sales document. A vendor or a consultancy scores your "readiness," finds you almost ready, and — conveniently — sells the thing that closes the gap. The party that diagnoses your AI problem usually sells you the cure. That structure, not the technology, is why the failure numbers look the way they do.
Tektari sells no implementation and takes no vendor money. That's the one thing that lets us publish the material the seller can't afford to: the buyer's-side playbook for not getting burned. Giving it away costs us nothing we were willing to charge for, and it's the honest version of the same discipline our paid work runs on. The numbers this guide is built on are worth sitting with — not to scare you, but because each one traces back to a step a checklist would have caught:
- RAND interviewed 65 practitioners and put the failure rate of AI projects above 80% — roughly twice the rate of IT projects that don't involve AI. Their #1 cause: "misunderstandings and miscommunications about the intent and purpose of the project." Not the models. The problem statement. (rand.org, RRA2680-1)
- MIT's NANDA project found that roughly 95% of organizations show no measured P&L impact from their generative-AI spend — on an estimated $30–40 billion of investment. That study is preliminary and contested, so we use it in exactly that bounded form: not "95% fail," but "almost nobody can show the money on a P&L." (primary PDF)
- Gartner predicted at least 30% of generative-AI projects would be abandoned after proof-of-concept by end-2025 (republished release), and now predicts more than 40% of agentic-AI projects will be canceled by end-2027 (via MarTech).
- S&P Global found the share of companies abandoning most of their AI initiatives jumped from 17% to 42% in a single year, with the average organization scrapping 46% of its proofs-of-concept (via CIO Dive).
- The Manufacturing Institute and PwC surveyed manufacturing leaders in April 2026: 58% describe their own executive leadership's AI use as "limited," and 54% have low or very low confidence in their frontline leaders' ability to lead AI change. (nam.org)
Every one of those failures was, at some point, a proposal on an executive's desk that nobody was equipped to interrogate. This guide makes the interrogation routine — before the money moves, in language an operator already speaks.
the method in four moves
The failures above cluster into four disciplines. Verifying a decision means running all four before you commit — in order, because each one depends on the last. Skip one and the failure finds the gap you left.
Make the problem exist on paper first.
Before a tool, a vendor, or a budget line: one written problem statement, a sponsor-signed baseline number, a pass/fail line, and a kill rule. This is RAND's #1 failure cause — the intent nobody wrote down — turned into a page. If the problem can't survive being written plainly, you've found the finding, and it cost you an afternoon instead of a quarter.
Measure the "before" your own team can trust.
You cannot prove an AI worked if you never measured what it replaced. Locate the records, trace them, and pull a current-state number from the system of record — tagged for whether it was counted or estimated, and signed by the person who owns the P&L it touches. AI fitted to miscoded records repeats the miscoding, faster and with more confidence.
Find out who holds the pen — before they do.
Who wrote the scope, who defined "delivered," who measures it, who carries the cost of failure, and what you keep if you walk. Most mid-market AI arrives on vendor paper; MIT NANDA tied roughly two-thirds of deployments to external partners. This is where most of the risk gets signed — usually by the buyer, usually without noticing.
Make the call on the record — and check who carries it.
Put the go, no-go, or not-yet in writing, against the baseline and the acceptance test, with kill criteria named. Then ask the MI/PwC question about your own shop: can leadership defend this under questioning, and can frontline supervisors actually carry the change? A decision no one can defend and no one can run is a failure with a later date.
the toolkit
Five tools, one per move (the decision move gets two: the vendor scoping done before you sign, and the record made after). Each page is a short read plus a template your team fills in — no theory you can't act on. Start at the top; a skeptical COO can put the first one to work today.
The tools are ordered, but not locked. If you already have a written problem statement, start at the baseline. If a vendor proposal is on your desk this week, jump to the scoping checklist and the acceptance test, then come back and do the problem statement before you sign — because that's the one that decides whether the other four are even worth filling in.
download the whole toolkit
Want all five templates in one file you can print, mark up in a meeting, or hand to a direct report? The complete toolkit is a single PDF — the same content as the five pages, formatted to fill in by hand. No email required; it's a direct download.
Download the Verification Toolkit (PDF)
Every template in the guide is also copy-buttoned on its own page — click copy and paste it straight into a doc. Nothing here is watermarked, gated, or licensed back to us. Use it internally, adapt it, put your own logo on it. It's a checklist, not a product.
how to use it — and why it's free
How to use it. On the next AI proposal that reaches your desk, don't debate the model or the vendor's demo. Open move 1 and ask for the problem statement — one page, signed by the person who owns the number it's supposed to move. Most proposals stall right there, and that's the point: the cheapest failed pilot is the one you never started. If it clears move 1, walk it through 2 through 5 before a dollar moves. The whole method is designed to be run by your own people, on your own data, with nobody from Tektari in the room.
Why it's free. Because we can't sell you the cure, publishing the diagnosis costs us nothing but the time to write it well — and it proves the charter better than any claim could. Under Tektari's Verification Charter, we sell no implementation, take no vendor commissions or referral fees, publish our prices, and put part of our fee at risk. A firm that made its money building the pilots could not afford to teach you how to kill one.
Where the paid work fits. This guide is the method done by you, on your own. The Problem-Definition Audit is the same discipline done with you, in three weeks, on one real decision — $12,500 fixed, and $2,500 of that is invoiced only when the scoped pilot passes the acceptance test your team runs on your own data. Same moves, same pages; the difference is that we do them alongside you and put money on the result. If the guide is all you ever need from us, that's a win we designed for. And if you'd rather see where you stand before you open a single template, the free Industrial AI Scoreboard scores you against the four disciplines in twenty questions — its published methodology carries the same failure evidence this guide is built on.
If a firm makes its money building the pilot, it can't afford to teach you how to kill one.
This field guide runs under Tektari's published Verification Charter: no implementation sold, no vendor commissions or referral fees, prices published, and part of the fee at risk. Every template is meant to be run by your own team, on your own data, without us — the acceptance test included. The worked examples in the guide are clean-room synthetic; a live engagement uses only your own records.
See where you rank first.
The guide is free and the toolkit is yours. If you'd rather start with a number, take the free Industrial AI Scoreboard — twenty scored questions across the same four disciplines, no sales call — and see which band you're in before you open a template.
Take the ScoreboardThat doesn't look like an email — try name@company.com