Red flags in an AI agency proposal, read from the page

The red flags in an AI agency proposal are mostly things the document leaves out. Look for work described as activity instead of deliverables, no test for “done”, pricing that can grow, benchmarks with no source, no number for running costs, and a handover that is promised but not written. Any one is fixable. Three together mean you are buying a conversation.

This is a guide to reading the page itself. For what to ask across the table, see our questions to ask an AI agency before you sign. Here, we assume the PDF is already in your inbox.

Why read the proposal and not the pitch?

A proposal is the only artifact that exists before money moves, and it is the only one you can hold a builder to later. The pitch deck describes what is possible. The proposal should describe what will be delivered, tested, priced and handed over. If a point is not written there, it is a verbal promise, and verbal promises do not survive a change of staff.

Buyers with more procurement muscle than most of us read it the same way. The US Office of Management and Budget’s memo M-25-22, dated 3 April 2025, tells federal agencies to use performance-based requirements so they can understand and assess vendor claims before contract award. Read before you sign, in other words, not after.

Flag 1: the scope is activities, not deliverables

If the scope section is built from words like “discovery”, “AI strategy”, “ideation” and “ongoing optimization”, nothing in it can be finished, so nothing can be late. A sound scope names objects: an agent that does one named job, a report that lands on a stated day, a repository you receive.

The UK government’s AI Playbook, published in February 2025, gives the same starting point for its own buyers: start with your problem statement. A proposal that never restates your problem in your words was probably written from a template.

Flag 2: nothing says how “done” will be tested

A proposal with no acceptance criteria hands the definition of done to the builder, who will define it on the last day. Sound criteria name the test set, the pass rate and who runs the test. They are written before the build and measured on your real inputs, not on a demo.

Anthropic’s evaluation documentation, current as of October 2026, shows the difference. “The model should classify sentiments well” is its example of a bad criterion. Its good one names a score of at least 0.85 on a held-out test set of 10,000 posts. Specific, measurable, tested on data the builder did not tune against. Our guide to a statement of work that holds up covers how to write that section into the scope.

Flag 3: the price can grow

Open-ended hourly rates, “estimates” with no ceiling, and phases priced only after the previous one finishes all put the cost risk on you. The risk is real because the builder is not the only one who gets AI costs wrong. A Mavvrik and Benchmarkit survey of 372 enterprise organizations, published on 10 September 2025, found 85% missed their AI cost forecasts by more than 10%, and 24% by 50% or more.

Mavvrik sells AI cost tooling, so weigh it accordingly. The takeaway still holds: nobody forecasts these costs well, so a price that can grow will.

Sound pricing is a fixed number for a fixed scope, with change handled as a new written order. If the work cannot be priced yet, the honest proposal says so and sells you the scoping step first.

Flag 4: the benchmark has no source

“Up to 95% accuracy” with no dataset, no date and no definition is decoration. So is “saves 20 hours a week” with no description of who measured it. A sound proposal either cites a named, dated source, or says plainly that the number will be measured on your data during the test described under flag 2.

Ask of every figure: measured by whom, on what, when. If the answer is “a previous client”, ask for the part you can check yourself. Our own numbers are dated for the same reason: the agent fleet proof page reports 41 agents, 5,450 runs and zero failures over 13 days, as of July 2026, with the measurement window stated on the page.

Flag 5: the running cost is missing

Delivery is not the whole bill for an AI system. Model usage is billed per call, models get retired, and someone has to watch the output. A proposal that ends at “go-live” has priced the part that makes it look affordable.

The UK Playbook tells buyers to budget for hidden costs and to consider ongoing support and maintenance when drafting requirements. In a proposal, that becomes a monthly estimate at your expected volume, a named account for the model key, and a stated maintenance path: who owns it after delivery, and what that costs.

Flag 6: handover is promised, not written

“Full documentation and knowledge transfer” is a sentence, not a clause. A sound proposal lists what you receive: repository, prompts, evaluation set, configuration, documentation and the accounts everything runs on. It says the accounts are yours from day one. A sentence about “IP to be discussed” means the agency has not decided yet, and the default favours the builder.

OMB’s memo makes the same point for its buyers. It asks for knowledge transfer, clear data and model portability, clear licensing terms and pricing transparency as the lock-in protections that proposals should offer. Ask for each of those as a named line item.

The six flags side by side

On the page	What it usually means	What a sound proposal says
”AI strategy”, “discovery”, “optimization”	Nothing to finish, nothing to be late	Named deliverables with a delivery date
No acceptance section	Done is decided on the last day	Test set, pass rate, who runs the test
Hourly or “estimate only”	The cost risk is yours	Fixed price, change as a written order
”Up to 95% accuracy”, no source	Marketing figure, not a measurement	Cited source, or measured on your data
Ends at go-live	Run costs left for later	Monthly estimate, your key, a maintenance path
”Knowledge transfer included”	Handover undecided	Itemized handover on your accounts

How we write ours

For our own work, the proposal states the scope as deliverables, how each deliverable is accepted, and the price as a fixed number. You receive the delivered code, data and documentation, and the accounts sit with you. Change is a new written order, never an invoice surprise.

If the problem is that you cannot yet write the scope, that is what the AI operations audit is for. It costs $1,900, fixed, runs 10 working days, and returns a ranked build list in writing. If we find nothing worth building, you get a full refund. The list is written so any competent team can build from it, including the one that sent you the proposal you are now reading.

Start async

Send us the proposal you are weighing at /start. You get a written reply within one business day, no meetings, with the flags we found in it and which ones we would ask the author to fix.