What should an AI marketing tools selection checklist cover?

What to take away

Define the task first

A checklist that opens with product categories produces a shortlist of things you admire rather than things you need. Begin with the work: which campaign tasks eat hours every week, and which of those could a machine carry?

Write the task as a sentence a colleague could audit. First drafts of product emails from an approved claim bank is testable. Improve marketing with AI is not, and no supplier can be scored against it.

Then set the boundary conditions. Volume, languages, channels, brand rules and who signs off output. Those become pass or fail lines in the checklist, before any weighting starts.

Compliance and data checks

The UK government's AI regulation white paper, published in March 2023, sets out five cross-sector principles: safety, security and robustness; appropriate transparency and explainability; fairness; accountability and governance; and contestability and redress. Its principles-based approach relies on existing regulators, so identify which one oversees your planned use: the ICO for data protection and direct marketing, the CMA for competition and consumer protection, the FCA for financial services, or Ofcom for communications and relevant online-safety matters.

For campaigns, check both UK GDPR and PECR. The ICO's direct marketing guidance explains direct marketing requirements; ask where the supplier stores consent evidence, including when and how consent was obtained and what people were told, and how consent and suppression records travel through its system. PECR requires consent for electronic-mail marketing to individuals unless the soft opt-in applies: check that the contact details were collected during a sale or negotiation, the marketing is for your similar products or services, and an opt-out was offered when details were collected and in each message. If you rely on legitimate interests under UK GDPR, document the balancing assessment; it does not replace any consent PECR requires.

Then ask where data sits, how long it stays, and whether it trains models the vendor owns. Ask for the relevant contract terms and settings, including how to prevent unauthorised reuse and how deletion is handled. Get those answers in writing before the pilot, not after.

Evidence every vendor owes you

Ask every supplier on the longlist for the same evidence. A live sandbox, a named implementation contact and a written statement of how the model treats your inputs are the minimum. Request a data-flow diagram, retention and deletion terms, model-training terms, access controls, support arrangements and written exit commitments. Test the same representative task with each supplier so you can compare the output and the workflow, not just the demo.

Teams also need enough internal skill to judge output. Microsoft's AI learning hub covers prompting and responsible use, a fair benchmark for what your own staff should know before you buy.

Finally, test the exit. Ask how you retrieve your data and what leaving costs. Give the answer its own score.

Weighting and scoring discipline

Scores without weights are decoration. Give each line a weight that reflects the task you defined, then CAP the number of lines so differences between suppliers mean something.

A workable shape is ten criteria with weights summing to 100. Compliance items need a minimum score, and no single line should carry more than a quarter of the total.

Weightings are the part buyers argue about, so write down the reason for each one. A criterion nobody can justify is usually a feature somebody liked in a demo.

Before the first demo, freeze the criteria, weights and evidence required for each line. Do not change a weight to favour a supplier after seeing its results.

Situation and choice table

Buying situations differ, so the emphasis inside the checklist should shift. Use the table to decide what matters most.

Score only the column that matches your situation. Adding rows for every scenario turns a working checklist into a wish list.

Common questions

How long should the checklist be?

Ten to fifteen lines. Enough to separate suppliers, short enough that two colleagues score the same way.

Does a lawyer need to see every shortlist?

Not at the first pass. Bring legal in once the checklist raises a data transfer, consent or contract question you cannot settle from the vendor's paperwork.

What if two suppliers score the same?

Revisit the weightings rather than the scores. A tie usually means the criteria are too broad for the task you defined.

Should we score a free trial?

Yes, but on the same rubric and with your own data. A trial scored on the vendor's sample dataset tells you very little.