How to Sanity-Check AI Vendor ROI Claims

Every AI vendor will tell you their product saves time, reduces cost, and delivers measurable returns. Most of them are quoting from generic industry benchmarks, not from organisations like yours. The gap between the headline ROI figure and what your organisation will actually see is almost always wider than the sales deck suggests.

These five questions separate real ROI from sales fiction before you sign anything.

Question 1: Show me ROI for organisations similar to mine

Generic claims — “customers save an average of 40% of processing time” — are meaningless without context. What counts as a similar organisation? Same industry, comparable headcount and revenue, same use case, and case studies from the last two years.

Question 2: What are the full implementation costs?

Vendors quote software costs. They frequently omit everything else. When building a realistic year-one picture, the hidden costs include professional services to configure the system (typically £15,000–£50,000), staff training (£5,000–£15,000), integration work with your existing tools (£10,000–£30,000), any custom development required (£20,000–£100,000), and ongoing support contracts at 10–20% of software cost annually.

Request a full year-one P&L before any commercial decision. If the vendor resists producing one, that tells you something.

Question 3: Who measures the ROI — and how?

If the vendor measures ROI on your behalf, they have an obvious incentive to show positive numbers. The right answer is that you own the measurement, the vendor suggests the metrics, and your finance team validates the results. If ROI measurement will happen through the vendor’s platform, ensure you can export the underlying data for independent verification.

The metrics that actually matter: hours saved per month from time tracking data, error rate reduction before and after deployment, cost per transaction over time, staff satisfaction scores, and customer satisfaction indicators.

Question 4: What happens if the projected ROI does not materialise?

The honest answer is that some organisations are poor fits for some AI tools. A good vendor knows this and will have a structured approach to handling it. A 90-day success guarantee with defined exit or remediation terms is a sign of a vendor confident in their product. “Everyone sees ROI with this tool” is a sign of a vendor that has not thought it through.

Question 5: How long before break-even?

Realistic break-even timelines by use case:

• Simple automation (document processing, scheduling): 6–12 months
• Complex AI (predictive analytics, customer insight): 12–24 months
• Strategic AI (market entry, product innovation): 18 months or more

Any vendor claiming month-one ROI is either misrepresenting the implementation timeline or defining ROI in a way that does not map to your actual business outcomes.

The Sanity-Check Formula

When a vendor gives you a projected ROI figure, apply three adjustments before treating it as credible:

• Reduce the benefit estimate by 30% — vendors are optimistic
• Add 20% to the cost estimate — hidden fees and scope creep are real
• Extend the timeline by six months — implementation always runs long

If the adjusted numbers still support the business case, you have something worth pursuing. If they do not, the vendor’s original claim was not credible.