You opened three CRM demos this quarter. Each one called itself AI-first. Each one ran a demo where the AI worked perfectly on the vendor's seeded data, and you walked away with no clear idea what it would do on yours. The pitch deck closed and you were back to the spreadsheet, more confused than before.

The right question is not "is this AI good." The right question is "will this AI feature still be useful on the day my data is messy, my team is busy, and the pitch deck is closed?" Twelve questions follow.

This is a vendor-neutral AI CRM checklist. It does not pick a winner. It filters hype from product on any CRM you are evaluating in 2026, whether the AI feature ships built in, bolted on, or sold as a separate tier. The twelve questions are the same questions to ask a CRM vendor about AI whether your team is two people or fifty. Print it, share it, or take it into your next demo call. The right answers will look the same regardless of who you are buying from.

1. Can the vendor name a measurable outcome?

Before anything else, ask the vendor what the AI feature is supposed to do that you can measure. Not "improve productivity." Not "save time." A specific outcome you could write on a Tuesday-morning standup whiteboard.

A good answer sounds like: "It writes a one-paragraph summary at the top of every deal record, and the team stops scrolling through 40 emails to figure out where the deal is." A red-flag answer sounds like: "It augments your workflow with intelligent insights."

If the vendor cannot name the outcome in one sentence, the feature is a pitch.

2. Can you turn the AI off without breaking the CRM?

The fastest test for whether AI is product or pitch is whether you can disable it without losing the rest of the tool. If your CRM stops working when you turn the AI feature off, the AI was the feature. Make sure the table, the pipeline, and the import you actually use every day still work the same way.

A good answer: "Yes, it is a per-workspace toggle. The rest of the CRM behaves identically." A red-flag answer: "Why would you want to?"

A 1 to 50 person team that already lost a quarter to setup-heavy tools needs an off-switch that does not punish the rest of the workflow. This is the load-bearing test for any AI CRM for SMBs: the rest of the product has to stand on its own.

3. Will it work on your data, not theirs?

Most AI-CRM demos work because the vendor seeded the data. Yours arrives messy, partial, and inherited. Ask the vendor to run the AI feature on a CSV you exported from your current tool. Not their sample data. Yours.

If they refuse, or schedule a "data engineering call" before the test, you have your answer. The feature was built for the demo, not for your Tuesday morning. The answer to that one question is the entire answer to whether you should buy the AI tier.

4. Is the AI bolt-on, or is it the UX you have to live in?

Two shapes of AI feature show up in CRMs today. One sits as a utility next to the table you already use, runs when you click it, and leaves your workflow alone the rest of the time. The other replaces the table with a chat surface and asks you to type to your CRM instead of editing it.

If the AI is the UX, you are buying a chat surface, not a CRM. That is a different product. Some teams want one. Most small teams that already lost a quarter to setup-heavy tools want their table, their pipeline, and a real client book they can scan in five seconds.

Ask the vendor how a brand-new hire on Tuesday morning gets work done. If the answer starts with "they ask the assistant," the AI is the UX.

5. Where does your data go during inference?

When the AI processes your contact list, your deal notes, or your client conversations, ask where the data travels and which third-party model providers see it. The honest answer names specific providers (OpenAI, Anthropic, Google, the vendor's own model) and specific data-residency regions (EU, US, your own region).

The dodgy answer says "we use industry-leading security." That is not an answer to the question.

This matters more than usual for European SMBs and for any agency holding client data under contract. Ask for the data-processing addendum in writing during evaluation, not after the contract.

6. Is the answer prompt-shaped or rule-shaped?

Two different things share the AI label inside CRMs in 2026.

A rule-shaped feature is automation: when a deal moves to stage X, send email Y. It runs deterministically. It does the same thing every time. It is testable. Automation has been working in CRMs for ten years.

A prompt-shaped feature is judgment: read the email thread, write a summary, suggest a next step. It is non-deterministic. It will say one thing today and a slightly different thing tomorrow on the same input. That behaviour is fine for some jobs, dangerous for others.

Ask the vendor which features in your CRM are rule-shaped and which are prompt-shaped. If they cannot tell you, they are selling the label, not the feature.

7. What changes about the price at general availability?

Most AI features in CRMs are in beta or preview in 2026. The price you see today is not the price you will pay next year. Ask what happens to the AI tier at general availability: does the cost double, does it move from per-workspace to per-user, does it become a separate purchase.

A good answer: "Beta pricing is locked for 12 months from your sign-up date, and we will give 60 days' notice before any change." A red-flag answer: "Pricing will be evaluated based on usage patterns."

This question alone has saved buyers from a Q3-2026 sticker shock they could see coming if they asked at evaluation.

8. Who owns the AI's output?

The data the AI generates (summaries, scores, suggested fields, drafted emails) is also your data. If the AI writes summaries into your records, ask where those summaries live, and whether they export with the rest of your data. If the export drops them, you are renting the output, not owning it.

The canonical answer here is the one your CRM should give about all of your data: "Your data is yours. Export to CSV or JSON anytime, including after you cancel." If the AI output does not pass that test, the feature is locked inside the vendor's tier UI, and you will lose it the day you switch tools.

Ask the vendor to walk you through a real export of a record with AI-generated fields, on the demo call, and confirm what arrives in the CSV.

9. What happens when the AI is wrong?

Every prompt-shaped feature is wrong some percentage of the time. The right question is what the team does on the day it is wrong.

Does the bad output write straight into a customer-facing field with no review? Does it queue for human approval first? Is there an undo? Does the team know which records were touched by AI and which were touched by a human?

If the vendor cannot answer this in specifics, the feature has not been operated long enough to know. Ask for the error rate on the metric the vendor tracks internally. A vendor that has shipped the feature into production knows the number.

10. Where is the human in the loop?

Related to the question above, broader in scope. Ask where the AI hands off to a human and how much work that handoff asks of the team.

A good answer: "The AI drafts the email, the rep edits and sends. The CRM logs both the draft and the final version, side by side." A red-flag answer: "The AI handles end-to-end customer communication."

Small teams cannot afford an AI that needs more babysitting than the work it replaced. Specifically ask how many minutes per record the human-in-the-loop step takes on the vendor's own benchmark. If the answer is "it depends," you are looking at a feature whose real cost is paid in your team's attention, not in the per-user line item.

11. How does the vendor evaluate the feature?

Ask the vendor how they measure whether the AI is working. Not how they measure customer happiness in general. Specifically the AI feature you are buying.

A good answer names a metric they track in production, on real customer data, and shares the recent direction of that metric. A red-flag answer says "we constantly improve based on feedback" or "our model has industry-leading accuracy."

If they cannot show you the evaluation method, the feature is not being evaluated. A feature that is not being evaluated will not be improved. A feature that will not be improved will be charged for at the rate of a feature that is.

12. What is the exit cost if you cancel?

If you cancel the AI tier in eighteen months, what do you lose? The summaries the AI wrote into your records. The scores it added to your deals. The custom prompts your team built. The integrations that depended on it.

A good answer: "All AI-generated content exports with your data. Custom prompts export as JSON. Your records read the same after cancellation, the AI simply stops adding new content." A red-flag answer: "Most customers do not cancel."

The exit cost is the truest measure of how much you actually own. A CRM that costs you a quarter of work to leave is not a CRM you bought. It is one that bought you.

How to use the twelve questions

You will not get a perfect answer to all twelve from any vendor. Three or four red flags should make you walk. One or two are normal at this stage of the category.

The pattern that matters: vendors who take the questions seriously will answer in specifics, in writing, with the second-level question already anticipated. Vendors who brush the questions off are telling you what your evaluation will feel like for the next twelve months.

Use the checklist twice during evaluation. Once on the first demo call, to filter hype from product. Once on the second call, with the same questions in the same order, to see whether the answers stay consistent. Vendors whose answers drift between calls are vendors whose pricing will drift too. This is the cheapest way of evaluating AI in a CRM without committing weeks to a paid pilot.

This is the checklist we wish we had had in 2024 when we were the buyer instead of the builder. Three CRMs in, the team was still in the spreadsheet. The AI tiers had grown louder, the demos had grown slicker, and not one of them ran the AI feature on the actual messy data we had on hand. So we wrote the questions down. We use the same twelve when we evaluate the tools we build with. The questions do not stop being useful after the contract is signed.

Get the printable version

We've prepared a printable reference card to keep on hand while you build or audit your CRM. One page, ready to print.

Download the printable 12-point checklist

AI in a CRM: a buyer's checklist for SMBs