· Updated 2026-10-16 · buyer due diligence
AI agency due diligence: a buyer checklist with an exit plan
Before hiring an AI agency, check whether the people who will deliver the work can explain its limits, show evidence for relevant claims, and hand it over without keeping the buyer dependent. A polished pitch is not enough to approve a pilot. Use the same questions for every shortlisted provider and write down any answer that remains unverified.
This checklist is for a buyer comparing providers for a defined project. It is a procurement aid, not legal advice or a certification of a vendor.
1. Give each provider the same brief
Describe the user, workflow, current workaround, intended outcome, systems involved, data sensitivity, deadline, budget boundary, and who will accept the work. Say what is out of scope. Ask each provider to respond to the same brief so their delivery plan, exclusions, and assumptions can be compared.
Keep discovery separate from delivery. A paid pilot should answer a named uncertainty, produce inspectable work, and end in a decision to continue, revise, or stop. Ask what evidence will be available at the end—not only what demo will be shown.
2. Verify the team and the evidence
- Who will do the discovery, engineering, evaluation, and ongoing support? Ask to meet the people named in the proposal.
- Which similar project can the provider discuss, and what part did this proposed team actually deliver?
- Can a reference explain the original scope, what changed, the handover, and what remained with the customer?
- Which claims are based on measured results, and which are estimates or demonstrations? Ask for the method and limits behind each number.
- What happens when the model is wrong, data is missing, or the integration fails?
With permission, call a relevant customer yourself. Ask what they would repeat, what surprised them, how the supplier handled a problem, and whether another team could operate the result. Compare this with the provider’s reference-call questions and case-study claim checks.
3. Map data and operating responsibility
Ask for a simple data-flow map: what enters the system, where it is processed, which model or subprocessors receive it, what is retained, and who can access it. Confirm who approves production data use, who monitors failures, and who can pause the workflow. Request the vendor’s actual policies and contract terms; do not infer compliance from a slide or a model-provider logo.
Agree what the buyer can inspect during the pilot: test cases, evaluation results, error examples, version changes, run costs, and the fallback path. Make clear who responds to incidents and how quickly an urgent issue is escalated under the proposed agreement.
4. Put ownership and exit terms in the proposal
Before work starts, identify who owns or can use the code, prompts, evaluation set, documentation, and project-specific data. Clarify third-party licences and any usage fees. The exact terms depend on the contract and jurisdiction; have the buyer’s legal adviser review language that matters to the transaction. See the guide to IP and code ownership clauses.
Write an exit plan while the relationship is healthy. It should answer:
- What source code, configuration, runbooks, test cases, and architecture notes will be handed over, and in what format?
- How can the buyer export its data and verify that vendor-held copies are deleted or retained under an agreed rule?
- Which accounts, credentials, keys, and deployment permissions will transfer or be revoked?
- What support period covers handover, and what does it cost?
- Can another team run the system without a private tool or undocumented step controlled by the vendor?
Use the vendor transition plan guide to turn these answers into named owners and dates.
5. Compare the full cost and decide
Compare the same scope across proposals. Separate discovery, build, integration, model usage, hosting, maintenance, support, and buyer-side review. Ask what could make each line grow and what is excluded. A low pilot price can still lead to a costly handover or ongoing dependency.
Score fit, evidence, delivery ownership, data controls, commercial terms, and exit readiness using one provider scorecard. Put unresolved risks beside the decision instead of averaging them away. Pause selection if there is no accountable buyer, the provider cannot name its delivery team, material claims have no evidence, or the buyer cannot retrieve and operate the result.
Hire AI Agencies describes its directory, intake, and comparison approach on the services page. If you are comparing live proposals, the contact page explains how to start an intake with your brief and constraints.