The tool is never the first problem: what a seven-figure ad audit taught me about AI

This is a story about a Google Ads account, not an AI project. I am telling it here anyway, because it is the clearest example I have of the thing that quietly wastes the most money in AI too. The expensive problem is almost never the tool. It is that the business cannot see itself clearly enough to know which tool to point where.

A while ago I reviewed the paid advertising of an established UK marketplace. They had been on Google for the better part of a decade, had spent well over £1M across the account's life, and were running at six figures a year. Signups had grown every year. Cost per signup sat stable in the low tens of pounds. On the headline numbers it looked healthy, which is exactly the kind of account most people leave well alone. I found three things, and none of them showed up in the headline numbers.

One: it was funding about a tenth of its own demand

Every live campaign was showing for only about a tenth of the relevant searches. It was losing most of its available impressions to budget, not to relevance. That distinction matters enormously. The ads already matched those searches and were eligible to appear. The money simply ran out each day before the demand did. On a deliberately conservative model, the budget caps were turning away several times the volume of demand that was already matching. Not new demand to go and find. Demand that was raising its hand and being shown the door.

Two: it had not been modernised in years

All of the live ads, roughly a hundred of them, were a legacy format Google froze back in 2023. Across more than two dozen ad groups, there was not a single modern responsive ad. A seven-figure account was running entirely on manual bidding. None of this was negligence. It was the ordinary result of an account that worked, left running by people with a business to operate. But the tools most likely to lift results were sitting in the box, unused, while everyone watched the stable cost per signup and concluded nothing needed doing.

Three: it could not see the most important thing about itself

This is the one that matters, and the one I think about most. It is a two-sided marketplace. It needs both sides to grow, and the two sides are not interchangeable. Yet every signup was counted as one, with no distinction between which side it came from, and placeholder conversion values sat throughout the account. One side of the marketplace was barely targeted at all. It was missing an entire high-value category of search, and spending its small budget on the wrong intent.

So the account could not answer the only question that actually governs a marketplace's growth: which side are we short of, and are we feeding it? It was optimising hard towards a number that blurred the two things it most needed to tell apart.

You cannot optimise your way out of a measurement problem. A system that cannot see the distinction that matters will get very efficient at the wrong thing.

The insight was a better question, not a better tactic

For a marketplace, growth is not "buy more demand". It is "fund the constrained side". This account could not tell which side it was feeding, so it could not make that call, so more budget would simply have poured faster into a blur. The most valuable output of the whole review was not a media tactic. It was moving the question from "how do we spend more on ads" to "which side of the business should we be spending on at all". Everything useful followed from that reframe.

Where AI actually comes into it

Here is the part that connects to the work I normally do. Google's smart bidding is, in the plainest sense, AI: machine learning that sets bids far faster and more granularly than any human running manual bids across dozens of ad groups ever could. It was available to this account the whole time. Switching it on would have felt like the modern, obvious move.

It would also have been a mistake, at least as the first move. Smart bidding learns from the conversion data you give it. Point it at an account that counts both sides of a marketplace as one undifferentiated signup, with placeholder values, and it will optimise confidently towards the blur. You do not get out of the hole faster. You get driven into it faster, with a tool that is genuinely good at its job. The AI was never the missing piece. Clean measurement and the right question were, and the AI only becomes an advantage once those exist.

I see the same pattern in AI implementation almost every time. A business reaches for the clever tool, the model, the automation, while the real constraint is that it cannot yet see itself clearly enough for the tool to help. Layer AI on top of a process nobody has mapped and data nobody trusts, and you get confident, expensive output that is subtly aimed at the wrong thing. That is how a promising pilot turns into Pilot Purgatory. The technology works. It is just answering a question nobody checked first.

The recommendation, and the honest framing

What I handed over was deliberately unexciting: a staged plan, one lever at a time. Fix the foundations first, which meant modern responsive ads, clean targeting, real conversion values, and the under-served side of the marketplace rebuilt from the ground up. Then uncap the proven campaigns in steps, so growth could be watched rather than gambled. Only then introduce smart bidding, once there was a trustworthy baseline for it to protect and learn from. No big-bang changes to an account that was, after all, still working.

I want to be honest about what this is. It is the diagnosis stage, not a victory lap. I am not showing you a graph that tripled. The point of the story is what a proper review surfaces in an account everyone had assumed was fine, and that the biggest wins on the table were structural, not clever. Nobody needed a growth hack. They needed the account to be able to see itself, and the right question asked of it.

That is the whole of my method, whether the subject is paid media or an AI system. Before anyone talks about the tool, I work out what the business actually cannot see, and which question it is quietly getting wrong. The tool is the easy part, and it is never the first problem.

This is also, more or less, what a First AI System Recommendation is: that diagnosis, pointed at where AI would matter most in your business, before a line of anything is built. If you would rather work it through in a room of other owners, I run The Implementation Room for exactly that.