Blog · Growth · 1 August 2026

How to Vet an "AI-Powered" Agency in Fifteen Minutes

Six checks in fifteen minutes, all public info: output trails, live products, named humans, pricing. We ran it on ten UK agency sites, and on ourselves.

You can vet most "AI-powered" agencies in fifteen minutes with public information, because the claims that matter are checkable: real output trails, real pricing signals, named humans, live products. If none of those are visible, the AI story is probably a slogan.

We know this because we recently did it, on around ten UK agency websites, while trying to work out how to position ourselves. What we found was, well, useful. Not damning exactly. Just useful.

Why "AI-powered" claims are usually unfalsifiable

Anyone can say they're AI-native. It costs nothing and there's no regulator checking. The phrase has become a bit like "bespoke" or "world-class", something you bolt onto a homepage because it sounds current, not because it's demonstrable.

Here's the actual problem though: most claims about AI use are structured so they can't be checked. "Our proprietary AI system delivers autonomous development." Okay. Show me. Show me a commit. Show me a log. Show me literally anything the system produced, rather than a paragraph describing what it does.

More often than not, there's nothing behind it. Not because the agency is lying exactly, more that the claim was written by marketing and the AI use, if it exists at all, is a Copilot subscription and someone using ChatGPT to draft emails. Which is fine! Loads of teams use AI sensibly and don't oversell it. But if you're going to put "AI-powered" in your hero banner, you should be able to survive someone asking "powered how". We've written separately about the difference between AI-assisted and AI-operated, because that gap is exactly what the vague claims blur.

The test isn't whether the agency uses AI. It's whether the claim is falsifiable. If there's no way for you, an outsider with no login, to check it, the claim isn't really information. It's vibes.

The fifteen-minute checklist

We built a short checklist for this, using only things you can find without an account, a call, or a favour.

1. Ask to see the output trail. Not a description of the AI system, the actual thing it produced: commit logs, a changelog, a ship log, anything generated by the process rather than written about the process afterwards. If an agency claims an AI agent writes and ships code, there should be a trail of that happening, publicly, continuously. If the only evidence is a years-old blog post announcing the capability, that's not a trail. That's an announcement.
2. Check whether the agency runs any live products of its own. This matters more than it sounds. A portfolio of client work is fine, but skin in the game beats a portfolio of promises, because it means the agency is exposed to the same risks it's asking you to trust it with. If they build software for a living and claim AI-native workflows, do they use those workflows on anything they own?
3. Look for named humans. Real names, with a verifiable footprint, GitHub, LinkedIn, somewhere. Not "our team of experienced engineers". Actual people you could, in theory, message.
4. Look for any pricing signal at all. It doesn't need to be a number. A band, a sprint rate, or even just a clearly stated process for how a quote gets produced, is enough. Total silence on pricing, paired with heavy claims about efficiency and speed, is worth noticing.
5. Stress-test one case study. Take one and ask yourself: could any part of this be checked? Not "is it plausible", checked. If every number is unverifiable and every client is anonymous, you're not reading a case study. You're reading a short story.
6. View the site the way a crawler would. If an agency's entire pitch involves AI visibility, AI search, AI-native anything, but their own site is barely legible to a crawler, their advice on this is secondhand at best.

Fifteen minutes, that's it. No calls, no favours, no NDAs.

What we found when we looked at the market

We ran this over roughly ten UK software, MVP, and AI agency sites, as research for our own positioning rather than as a hit piece, and honestly expected to find a mixed bag. We didn't, really.

• No pricing at all, on almost every site. Not even a band. Not even "starts from". Just a contact form and a promise of a conversation.
• No named team members, anywhere, on most of them, including at least one agency whose entire pitch was "founder-to-founder", which is a strange thing to say when you won't tell anyone the founder's name.
• Case studies with numbers that couldn't be verified against anything. Percentages with no baseline, growth figures with no timeframe, results with no named client.
• "AI-native" and "autonomous" claims with nothing behind them. No logs, no architecture, no sample output, nothing you could point at and say "there, that's the AI bit".
• Literal placeholder statistics live in production, on one site. Zero percent, displayed as a stat, on a page presumably meant to impress people. Nobody had gone back and filled it in.
• A self-ranked "best agencies in the UK" listicle, with one agency placing itself first and no disclosure that the list was self-authored. Which, look, we get the temptation. But it rather undermines the "we're honest and transparent" line that usually sits two paragraphs above it.

None of this means these are bad companies. Some might be excellent, quietly, without needing to prove it to strangers on the internet. But if the whole pitch rests on "trust our AI claims", and there's nothing checkable behind the claims, that's a gap worth noticing before you sign anything.

Apply it to us

Fair's fair, so here's our own answers to the checklist, plainly, no dressing it up.

Output trail: the homepage ship log is a build artefact, not an editorial choice; if nothing shipped that week, it says so.

Live products: yes, we build and run our own. Not client work dressed up as a portfolio, actual products we operate day to day.

Named humans with a footprint: honestly, this is the one we only half pass. The site leads with the agent fleet rather than founder bios. The studio's architecture, the thing that runs it, is public on GitHub, and that's a verifiable footprint you can go and inspect, but it isn't the same thing as named bios, and we won't pretend it is.

Pricing: we don't publish a number on the site, but we're honest about it on a scoping call, and we explain how we work on the services page and answer the obvious questions on the FAQ. No mystery process behind a contact form.

Crawler visibility: our own AI-search advice isn't secondhand, because we build the studio to be legible to crawlers as a baseline requirement, not an afterthought bolted on for SEO reasons.

Now, obviously, we designed this checklist and then applied it to ourselves, so it's a bit rich for us to act surprised we pass it. We built the studio this way partly because we think it's the right way to run a software business, and partly because we knew a version of this article would eventually get written. Both things can be true.

So don't take our word for it. Apply the same six steps to us that you'd apply to anyone else. Go and check the ship log. Go and look at the GitHub. Ask the pricing question on a call and see if you get a straight answer or a dodge. If any of it doesn't hold up, that's useful information too.

What to do with what you find

A word of caution before you go full detective on every software agency site you find: absence of pricing isn't proof of villainy. Enterprise contracts genuinely vary too much for a fixed number on a homepage to mean anything. And some small, good teams just have thin, neglected websites, because they're busy doing the work rather than writing about it.

This checklist finds signal. It doesn't hand you a verdict.

What it does is change the shape of the conversation. Instead of asking "do you use AI", which invites a slogan, ask "can I see something the AI produced". Instead of asking "what results have you delivered", ask for one number you could actually check. Instead of trusting "founder-to-founder", ask for the founder's name.

Most agencies will answer these reasonably. Some will bristle, and that reaction tells you something on its own. Either way, fifteen minutes of looking beats fifteen minutes of being told what to believe.

Drafted by the Adapt Progress Evolve agent fleet; edited and approved by a human.