How to Sanity-Check an AI Vendor's ROI Claims Before You Sign

A vendor ROI figure is written by the vendor's marketing team. They choose the best customer, the best month and the best metric. That is their job. It is not dishonest. It is selective.

Regulators now look at this. DLA Piper reported in May 2026 on an FTC action over AI-washing claims in a business-to-business setting. The point for you is simple. Overstated AI claims are a recognised enforcement issue in the US, so you should assume they exist in the UK market too.

Your job is to move the claim from marketing to evidence. Five checks do that.

A claim of 70% faster means nothing until you know faster than what. Ask the vendor three questions. How long did the task take before the tool? Who measured it? Over what period?

A good answer names a process, a before figure, an after figure and a date range. A bad answer says staff felt more productive. Feelings are not a baseline. If the vendor cannot give you the before figure, the after figure has no value.

A case study from a firm with 5,000 staff tells you little about your firm of 40. Ask for a reference customer of your size, in your sector, using the same product version you are buying.

Then ask to speak to them with no vendor on the call. Ask what they would change. Ask what the tool costs them in staff time to run each week. Ask whether they measured the saving themselves or accepted the vendor's figure.

If the vendor refuses a reference call, treat that as your answer.

Take the vendor's claim and rebuild it. Use your headcount, your hourly cost and your adoption rate. This example is illustrative, not real data.

A vendor claims the tool saves each user 5 hours a week. You have 20 users at a loaded cost of £25 an hour. The claim gives you 20 x 5 x £25 = £2,500 a week, or £130,000 a year. Now test adoption. If 8 of the 20 use the tool each week, the saving falls to 8 x 5 x £25 = £1,000 a week, or £52,000 a year. If the vendor measured its keenest users and your staff achieve half of that, the saving falls to £26,000.

If the licence costs £30,000 a year before training and support time, you lose money on the realistic case. The vendor's figure said you would make £100,000. Run this arithmetic before every signature.

Vendor ROI figures usually count the licence and ignore everything else. Add the cost of integration, data cleaning, staff training, and the person who owns the tool each week. Add the cost of checking the output. An AI tool that needs a human to review every result saves less than the headline.

Ask the vendor for the total cost of ownership over three years. Write the figure down. Compare it to your rebuilt saving from Check 3.

A claim in a brochure binds nobody. A claim in the contract does. Ask for the promised outcome to be written in as a measurable service level. Name the metric, the baseline, the measurement method and the review date.

Then ask for a remedy. A fee reduction or a right to exit without penalty if the outcome is missed at the 90-day review. A vendor that believes its own number will agree. A vendor that refuses has told you how much it trusts the figure.

Take the ROI claim from the vendor you are considering now. Do these four things before you sign anything.