You have a PDF on your desk. Arrows, a logo, three case studies, a number at the bottom. Somebody wants you to sign it. This post is the checklist to run before you do — ours, theirs, anyone's.

We wrote it the way we wish every prospect would read our own proposals. If a proposal — including one of ours — does not survive this checklist, do not sign it.

The frame: testable promises, not a brochure

A good AI services proposal makes promises that can be checked while the work is happening, not only when it is done. Everything below pulls those checks forward into the document, so the day you sign is also the day the falsifiers are agreed. If a section cannot be turned into something you can inspect, it is decoration.

What to look for

A named deliverable for week one. Not "kickoff" and "discovery" — an artifact. A file, a configuration, a working endpoint, a written read of your current systems. If week one produces only meetings, week twelve will too.

Evidence classes on the claims. A proposal that says "we will integrate with your CRM" should say how you will know — a configuration receipt (Class C) you can open, a test row you can submit, a dashboard panel you can refresh. If the vendor cannot tell you what form the evidence will take, they have not thought about it yet.

A falsifier per milestone. For every claim — "the form delivers to your CRM", "the model classifies tickets at the agreed accuracy", "the dashboard reflects the live database" — one sentence describing the observation that would prove the claim wrong, and the day it will be made.

A ledger, or its equivalent. Some written, append-only place where claims, evidence classes, and falsifiers are recorded as the engagement runs. Format does not matter. Existence does.

Named humans, not just a logo. Who does the work. Who owns the receipts. Who picks up the phone in week six when something breaks.

What to red-line

"AI-powered" with no specific model, data, or decision. Ask which model, on which inputs, making which decision, evaluated how. If those four answers are not in the document, write them in before signing.

Accuracy numbers without a test set. "94% accurate" is meaningless without the data it was measured on and the data the system will see in production. Replace the number with "to be measured on a sample of the actual production data in week N, against falsifier X."

Undisclosed vendor lock. If the proposal quietly commits you to a single platform, API, or prompt format only this vendor can edit, that is a clause, not a footnote. Pull it into the open.

Training on your data, by default. Anything touching your customer data or proprietary corpora needs an explicit clause about whether the vendor or their model provider retains, retrains, or evaluates on it. Default-open is the wrong default.

What to refuse outright

Guarantees of outcome on language-model behavior. No honest vendor can promise a specific accuracy, conversion lift, or "delight" number unconditionally. They can promise a measurement plan and a remediation path. Demand the second.

"Endorsed by" claims that are actually placements. Being listed in a directory, included in a roundup, or named among several pathways is a factual placement, not an endorsement. Hold every vendor to that distinction — including us (Class E).

Pricing with no exit ramp. A twelve-month commitment with no clean stop point at week four bets on the vendor's first month being honest. Build in a falsifier-gated checkpoint.

Deck-only IP. If the proposal references methodology, framework, or "proprietary system" without a single file path, repository, or document you will be able to read during the engagement, you are buying the deck, not the work.

A note on the market you are actually in

The AI services market right now is two markets at once: a small one where people who can do the work are doing it, and a large one where decks are being sold to buyers who cannot yet tell the difference. Alicia Maren's interview with Jay Kumar Chimata of JobFirst.ai — see Meet Jay Kumar Chimata: JobFirst.ai And The Real AI Job Market — is a candid look from a 22k-user talent platform at which AI skills employers actually pay for in 2026. We read it as confirmation, from a different angle, that "show me what you have done that I can inspect" is now the buy-side question (Class E). If your vendors cannot answer that on the proposal, they will not be answering it in week twelve either.

The minimum viable proposal

A proposal worth signing has a list of claims, an evidence class per claim, a falsifier per claim, a named human per claim, and a checkpoint date per claim. Everything else is paint.

Receipts Before Pitch Decks: How To Read An AI Proposal