Every AI agency pitch deck has a number in it. A percentage saved, a count of clients served, a multiple returned. Almost none of those numbers come with a way to check them. That is not necessarily dishonesty, a lot of it is just habit: marketing writes a number once, nobody re-runs it, and by the time you are reading it on a sales call it has been repeated so many times it sounds like fact.
Your job before you sign anything is to turn each claim back into a question. Not "is this true" in the abstract, but "what would I need to see to believe this specific line." Below is the question for each type of claim you will actually run into, and what a real answer looks like versus a dodge.
ROI and percentage claims
"We deliver a strong ROI" or "clients see meaningful time savings" are the most common lines in this industry, and the hardest to check from a deck alone, because ROI is a ratio of two numbers you cannot see: what the client spent, and what improved because of it.
Ask three things. First, what is the baseline this percentage is measured against, and who measured it, the agency or the client? Second, is this figure an average across every client, or the best result cherry-picked to lead the pitch? Third, can they show you the calculation, not just the output? A real answer names the metric ("support tickets resolved without escalation"), the before and after values, and the time window. A dodge restates the percentage in a different sentence and moves on.
If a number cannot be recalculated by someone outside the agency, treat it as marketing copy, not evidence. That is a general rule, not just ours: it is also why we keep a public claims registry for our own numbers, so anyone can rerun the command that produced them rather than take our word for it.
Client counts and "trusted by" claims
"Trusted by dozens of businesses" is a headcount claim, and headcount claims are cheap to inflate because almost nobody asks for the list. Ask for it anyway: not full case studies, just names or a category breakdown you can sanity-check against the size of the agency. A small agency claiming a client roster far larger than its team could plausibly service should prompt a follow-up question about what "client" means in that count, a one-off audit call, a monthly retainer, and a free trial can all get bundled into the same number.
Ask whether the count is active clients right now or every client the agency has ever had, including ones who churned soon after signing. Both are legitimate things to say, but they are different claims, and an agency confident in its numbers will tell you which one you are looking at without being asked twice.
Testimonials
A testimonial with a first name, a stock photo, and no company is not evidence, it is a placeholder that never got swapped out. Ask for the person's LinkedIn, or ask to speak to one reference client directly. A legitimate agency has done this before and has an answer ready. An agency that gets cagey when you ask to verify a named testimonial has told you something, even if it never says it out loud.
We keep named testimonials off our own site until we have real, permissioned ones from finished engagements, for the same reason: an unverifiable name attached to a quote is worse than no quote at all.
Case studies and outcome claims
Case studies sit in a genuinely awkward spot: most agencies cannot name the client, because most client contracts have confidentiality clauses, and that is normal, not a red flag by itself. What matters is whether the claim inside the case study is still checkable even with the name removed. "This client saved several hours a week" should come with what task, over what period, measured how. If the agency cannot answer those follow-ups without the client's name, the anonymisation is hiding the mechanism, not just the identity, and that is the part worth being wary of.
Guarantees
A guarantee is only as real as what happens when it is not met. "We guarantee results" with no defined metric and no defined consequence is not a guarantee, it is a sentence. Ask three things: guaranteed against what specific, pre-agreed number; measured by whom; and what happens contractually if it is missed, refund, extra work at no charge, or nothing at all. If the answer to the third question is vague, the guarantee was never a real commitment, it was a persuasion device.
A pattern we ran into ourselves
While building our own claims-verification process, we came across an agency whose website carried two different ROI counters, one on the homepage and one on a case-study page, each citing a different multiple for what read as the same client engagement. Neither page linked to the other, and neither cited a source. It was not necessarily dishonest, most likely one page got updated and the other did not, but it is exactly the kind of thing that a five-minute cross-check catches and a first read-through does not. If an agency cannot keep its own numbers consistent across its own pages, that is worth noticing before you rely on any of them.
A free way to run this check yourself
We built a short, free tool for exactly this: the AI claims checker. Paste in the claims from a pitch deck or a website and it walks you through the same verification questions above, one claim type at a time, so you leave with a specific list of what to ask for rather than a vague feeling of doubt. No sign-up, and it is not gated behind a call.
What we do differently, and how to check that too
We would rather you hold us to the same standard than take our word for it. Every number-bearing claim we publish sits in a public registry with its source and the exact command or method to re-derive it, and our receipts page is where that evidence lives. If a figure of ours cannot be checked that way, we do not publish it. That is a low bar, and it should be the bar for anyone you are about to pay.
If you want a second pair of eyes on an agency's pitch before you sign, or you want to run this same checklist against your own shortlist with someone who does this for a living, book the free audit. Bring the deck, bring the numbers that made you suspicious, and we will work through them together, no obligation either way.