How to read an AI vendor proposal

What the numbers mean, what questions to ask, and what is almost always missing from the pitch deck — a practical guide for organizations evaluating AI vendors without a technical co-pilot.

AI vendor proposals have developed a distinctive genre with recognizable conventions. There is a slide showing the exponential growth of AI capability. There is a case study from a recognizable company in a tangentially related industry. There is a slide on "seamless integration." There is an ROI projection. There is a pricing table that becomes comprehensible only after a second conversation.

What is rarely in the proposal: a clear account of what the system will not do, where it will fail, what happens when it does, and what the full cost of ownership actually looks like over a three-year horizon.

This guide is about reading the proposals you receive — and the space between what they say and what you need to know.

What to watch for immediately

Before getting into the substance, there are signals in the first few pages of a proposal that tell you a lot about what you are dealing with.

Red flags — reconsider or probe hard

!

ROI projections with specific percentages and no methodology explained

!

Case studies from industries structurally different from yours

!

"Seamless integration" with your existing stack, stated without specifics

!

No mention of limitations, failure modes, or edge cases

!

Pricing that requires a follow-up call to understand

!

Reference to your data being used to train future models (often buried)

Amber flags — ask for clarification before proceeding

~

Accuracy claims without specifying the benchmark dataset used

~

Implementation timeline that seems faster than your internal bandwidth allows

~

Human review described as "optional" for consequential outputs

~

Change management or training described in a single bullet point

~

SLA defined around uptime but not around output quality

Green signals — a vendor worth taking seriously

Proactively describes what the system will not handle well

Proposes a pilot scoped to a specific, bounded use case

Asks about your data before proposing a solution

Clear data ownership terms — your data stays yours

References to specific compliance frameworks relevant to your industry

The questions to ask in the follow-up

The proposal is a marketing document. The follow-up conversation is where you get to conduct due diligence. These are the questions that consistently yield the most useful information.

1
What is the failure mode? When your system is wrong, how is it wrong? Does it fail silently or does it signal uncertainty? How would one of our employees know that an output needs human review?
2
What data was the model trained on? Can you describe the training data sources? Is our industry represented? What are the known gaps in coverage?
3
Who owns the data we put into your system? Is our data used to train future versions of your model? Can we request deletion? What are the data residency options?
4
Can you show us a failed implementation? Not the polished case study — a real engagement that underdelivered. What happened, and what did you learn from it?
5
What does the full three-year cost look like? Include implementation, training, integration maintenance, model updates, and the internal headcount required to manage the system.
6
What happens if we want to leave? How do we extract our data? What is the migration path? What proprietary formats or dependencies would we be locked into?
7
Who else at our company will this touch? Beyond the team using the tool directly — which other teams, processes, or systems will be affected? Has the vendor mapped this?

The number that is almost always wrong

The ROI figure in most AI vendor proposals is constructed from a productivity gain estimate multiplied by a headcount and an average salary. It looks precise. It is almost always wrong in the same direction — it counts the upside and ignores the cost of the denominator.

The denominator includes: implementation time (typically 2–4× the vendor's estimate for organizations without a dedicated AI team), integration maintenance (ongoing, rarely one-time), training and change management (underestimated in every proposal we have seen), and the productivity dip during transition, which is real and which the ROI model treats as zero.

A simple correction: take the vendor's ROI figure, halve the numerator, double the denominator, and push the break-even point 12 months further out than they project. If the investment still makes sense after that adjustment, it probably does make sense. If it doesn't survive that stress test, you need better numbers before proceeding.

The question the proposal never answers

Vendor proposals are built to answer the question: "Is this system capable?" The question they do not answer — and the one you most need answered — is: "Is this system right for us, at this moment, in this context?"

Those are different questions, and only you can answer the second one. What the proposal gives you is input. What you need to bring to it is a clear picture of your own organization — your actual processes, your actual data, your actual people, and the specific problem you are genuinely trying to solve.

If you don't have that picture before you start evaluating proposals, the proposals will fill the gap with their own answers. And their answers will serve their interests, not yours.

get independent guidance → download the readiness guide