Scalewell / Blog

AI vendor due diligence for marketing: the questionnaire to send before you sign

Before signing an AI vendor or partner, settle four things: whether it works on your data, what data it touches, what it can change in your accounts, and how you leave.

A marketer inspecting a polished demo dashboard with a magnifying glass, with a messier real dashboard behind it.

AI vendor due diligence for a marketing team means settling four things before you sign: whether the tool does what the demo showed on your own data, what customer data it touches, what it can change in your ad accounts and CRM, and how you leave. The questionnaire below covers all four for any AI vendor or partner.

The demo is usually where it starts. The vendor's AI found wasted spend in a sample account, wrote a tidy weekly summary and proposed budget moves in seconds, and the room was impressed. Then someone asks whose account that was, and the answer is a demo account built to show the product well. Nobody has seen it run on your accounts, your CRM or your conversion data.

This guide is for the CMO, VP of Marketing or head of demand gen who signs the contract. Instead of ranking AI tools, it gives you the pre-signing checklist, the full questionnaire with what good answers look like, a pilot design and the red flags. The risk register that sits behind it, covering vendor risk and lock-in once a tool is in use, is in our article on the generative AI risks a CFO or legal team will ask about.

Vendor AI claims need checking on your own accounts

Vendor AI often underdelivers against what was sold. In a 2025 Gartner survey of martech leaders, reported by MarTech on November 26, 2025, 45% said the AI agents their vendors provided don't deliver the promised business performance (MarTech). That's close to half.

Regulators have noticed the same gap from the other side. When the FTC announced Operation AI Comply in September 2024, a set of enforcement actions over deceptive AI claims, its chair said there is "no AI exemption from the laws on the books" (FTC). Those cases were mostly consumer schemes, but the message for buyers is the same: an AI claim is a marketing claim, and it deserves the same scrutiny you'd give any other.

So the job of due diligence is to move every important claim from the demo into your own data before you commit. Most of what follows is a structured way to do that without turning procurement into a six-month project. A team that runs these checks signs on evidence from its own accounts rather than a demo, and Scalewell, the partner that helps paid and demand gen teams use AI across strategy, planning, implementation and operations, can help your team run them.

AI vendor due diligence happens in four stages

Four stages cover the decision, and each one has a different owner. Running them in order saves time, because a vendor that fails the first stage doesn't need a legal review.

StageQuestion it answersOwnerOutput
1. FitDoes the tool do a job we actually have, better than what we already own?Head of paid media or demand genA one-line job description and the current way it's done
2. ProofDoes it work on our accounts and our data?The person who'll own the jobA pilot result against a written pass mark
3. Data and accessWhat does it see, store and change?Marketing ops with privacy or ITCompleted data and access sections of the questionnaire
4. Contract and exitWhat are we committing to, and how do we leave?Legal and financeSigned terms with data, notice and export clauses
A four-stage funnel of narrowing gates, with a different person at each gate and a document coming out of each, ending in a signed contract.
Four stages, four owners: each stage ends in a concrete output before the next begins.

Stage one is the most skipped and the most useful. Write down the job in one line, such as "draft the weekly paid media report from Google Ads, LinkedIn and HubSpot data," and how long it takes today. If the vendor's tool doesn't map onto a job on that list, the evaluation can stop there.

The AI vendor questionnaire, with what good answers look like

The AI vendor questionnaire below is written for a B2B marketing team buying a tool or service that uses AI on ad accounts, CRM data or marketing content. Send it before the second meeting, ask for written answers, and treat a missing answer as an answer.

Section A: What the tool does

#QuestionA good answerA red flag
A1Which jobs does the AI do, and which steps does a person still do?A specific list, with the human steps named"The AI handles everything"
A2Which model or models does the product use, and can that change without notice?Named models or providers, and a stated notice policyWon't say, or "proprietary AI" with no detail
A3What does the AI produce: suggestions, drafts or live changes?A clear split, with live changes off by defaultLive changes on by default
A4Can you show it working on a customer's account, with their permission, rather than a demo account?Yes, or a reference customer we can speak toOnly the demo account
A5How do you measure whether the AI's output was right?A named measure, such as acceptance rate or accuracy against a checkNo measure

Section B: Data

#QuestionA good answerA red flag
B1Which of our data do you need, and which do you store?A field-level list, with storage separated from processing"All of it"
B2Do you or your model providers train on our inputs or outputs?No, in the contract, for the plan we're buying"Not by default" with no clause
B3Where is our data processed and stored, and for how long?Named regions and a retention periodUnknown or indefinite
B4Do you sign a data processing agreement?Yes, standardOnly for enterprise plans above our size
B5Which subprocessors, including AI model providers, see our data?A published list with change notificationsNo list

Section C: Access to ad accounts and the CRM

#QuestionA good answerA red flag
C1Which permission level do you need on each ad account and on the CRM?The lowest level that does the job, starting read-onlyAdmin access by default
C2Which actions can the tool take in our accounts?A list of actions, each one switchableUnlimited actions
C3Can every change require approval by a named person?Yes, with a log of who approved whatApproval is optional or unavailable
C4Do you keep a change log we can export?Yes, with timestamps and the before and afterNo log, or log only in the vendor's interface
C5Which connection route do you use for each platform, and who holds the token?Named route per platform; tokens held in our workspace where possibleShared credentials or a personal login

Section D: Cost, support and exit

#QuestionA good answerA red flag
D1How does pricing change as usage, seats or ad spend grow?A written price table, including any usage-based lines"We'll discuss at renewal"
D2What happens to our data, prompts and workflows if we leave?Export in plain formats, then deletion with confirmationNo export, or export only on request at a fee
D3How much notice do you give before retiring a feature or model we depend on?A stated notice period in daysNo commitment
D4Do you indemnify us against IP claims on AI outputs?A clause, with exclusions our counsel can readNo indemnity at any plan level
D5Who do we contact when an AI change goes wrong in a live account, and how fast do they respond?A named support route and response timeEmail only, no stated time

The contract questions in section D overlap deliberately with the five contract answers in the AI risks article linked above, so the same answers serve both the signing decision and the risk register afterwards.

Ask exactly what the tool can change in your ad accounts

Access is where marketing due diligence differs most from a standard IT review. A tool that can change bids, budgets or targeting can spend money, so section C deserves a careful read before anything else is signed.

Connection routes vary a lot in what they allow. As of November 2026, Google's own Google Ads MCP server reads accounts and makes only partial changes, such as creating assets and labels with propose-and-confirm validation, and its developer page says it "cannot create, modify, pause, or enable campaigns, ad groups, ads, budgets, or bids". Microsoft's own Advertising MCP server is free, hosted and read-only, while LinkedIn's own MCP server, which reads and writes ad accounts, is in beta and in use but not yet publicly documented. Our guides to the Google Ads MCP server, the Microsoft Ads MCP server and LinkedIn Ads MCP describe what each route can read and change.

The AI client matters too. OpenAI's help center says full MCP support in ChatGPT, including write actions, is available on Business and Enterprise/Edu plans on ChatGPT web, while Pro users can connect custom MCP servers with read and fetch permissions only (OpenAI). So a vendor's answer to C2 depends on both the route and the client, and the honest answer names both.

Our rule is simple: start every new tool read-only, give write access one action at a time once a person has reviewed a month of its suggestions, and keep budget increases and launches with a named approver whatever the tool can technically do.

Run a paid pilot on your own data with a written pass mark

A pilot is the only way to answer the proof stage, and it works best when the pass mark is written before it starts. A vendor that won't agree to a written pass mark is telling you something useful.

A pilot spec we'd use for a tool that drafts weekly reports and proposes budget moves:

  1. Scope. One ad account, the CRM read-only, four weeks, suggestions only.
  2. Baseline. Log how long the current weekly report takes and how many budget moves the team made last month.
  3. Pass mark. The draft report needs fewer than 30 minutes of edits by week three, its numbers match the platform and CRM to the dollar, and at least half of its budget suggestions are ones the team would have made or wants to test.
  4. Fail conditions. Any number that doesn't match the source, any suggestion that would breach a budget ceiling, or any attempt to change the account without approval ends the pilot.
  5. Price. A paid pilot at a fixed fee, credited against the first year if you sign. A fee gives both sides a reason to take the four weeks seriously.

Keep the pilot's suggestions in a shared sheet with the team's verdict on each one. By week four that sheet is a better guide to the tool than any demo, and it doubles as evidence for finance.

Choosing an AI partner takes different questions from choosing a tool

An AI partner, such as an agency or consultancy that builds AI workflows for your team, needs a different set of questions, because you're buying people and their output as well as software. The key difference is what's left when the engagement ends.

Five questions we'd add for a partner:

  1. Who builds the work, and what are their backgrounds? Look for people who have run paid media as well as people who build software.
  2. Where does the work live? In your accounts and tools, or in the partner's own platform that you rent.
  3. Who owns the prompts, workflows and documentation afterwards? Your company, in writing.
  4. How is your team trained to run and change what was built? A plan with named sessions and an end date.
  5. Can we speak to a client whose engagement ended a year ago? That conversation tells you whether the work survived the handover.

Our comparison of in-house marketing vs agency for AI in paid media covers when a partner makes sense at all, and our survey of AI tools for B2B marketing sorts the tool options by job if you decide to buy software instead.

Vendor due diligence fails in four predictable ways

The most common failure is judging the tool on the demo. Demo accounts are built to show the product well, with clean data and obvious wasted spend. Insist on a pilot on your own account before price negotiations start.

Skipping the access review. Teams often grant admin access on an ad account to get a trial running and never revisit it. Grant the lowest level the pilot needs, such as Read-only under Admin, Access and security in Google Ads, or the Viewer role in LinkedIn Campaign Manager (managed in Business Manager if the account sits there), and write the access level into the contract.

Letting procurement run without the job owner. Legal and IT can review terms and security, but only the person who'll own the job can judge whether the output is good. Put that person on the pilot and give them the final say on stage two.

No exit plan. A tool that holds your prompts, workflows and history in its own format becomes expensive to leave. Ask for an export of everything during the pilot, and check you can actually open it.

Buying the AI tier of every existing tool. Each vendor's AI can only see its own data. Run the questionnaire on the add-ons too, and ask whether one connected assistant would do the same job.

How Scalewell helps your team run the evaluation

Your team gets a written job to judge every vendor against, and a working setup once it has chosen. It all opens on a free 30-minute call. Then comes an audit of your team's week, recording each recurring job and how it runs today, which is the stage one line this questionnaire asks for. Those jobs are scored into a ranked roadmap that shows which one is worth buying for first. The choice of tool stays with your team. After that, a build sprint, usually lasting 8 to 12 weeks, sets the first jobs up inside the tools your team already uses, including the one you picked, and your people are trained to run and change them without Scalewell.

Where to start

Write down the one job you want the vendor's tool to do and how long it takes today, then send sections A to D before the next call. If you'd like a second pair of eyes on the answers or the pilot design, Scalewell is happy to look at it with you on a free call.

Frequently asked questions

Do you need to re-run due diligence when an AI vendor changes its model?

Re-check the parts the change touches rather than the whole questionnaire. A new model or model provider can change where data is processed, which subprocessors see it and whether the training terms still hold, so ask the vendor to confirm those answers in writing again, and re-run a week of the pilot checks on your own account to see whether output quality moved.

Should agencies you already work with answer an AI vendor questionnaire?

Yes, if they use AI tools on your ad accounts, CRM data or content. The data and access sections apply to the tools the agency runs on your behalf, so ask which tools and plans they use, what access those tools hold, and put the same data and approval rules in the agency contract.

Can you use AI to review a vendor's questionnaire answers?

An assistant on a business plan with written data terms can compare the written answers against the questionnaire, list the questions left blank or answered vaguely, and draft follow-up questions. The contract itself still goes to legal, and the pilot result still comes from the person who owns the job.

Should you sign an annual AI contract after a successful pilot?

Only once the exit terms are in place. Ask for the export, deletion and notice clauses from the questionnaire to be written into the first contract, check that the pilot fee is credited, and ask whether usage-based lines in the price table have a cap for the first year.