On a Tuesday afternoon last year, a client of ours got an email that said one of their category pages had dropped nine positions over the weekend, here is what we think caused it, here is what we are doing today.

She replied asking how we spotted it so fast. The honest answer was that nobody spotted it. A monitor did, at some point on Sunday, and an alert was waiting for the account lead on Monday morning, and he looked into it and wrote to her on Tuesday.

The reason I remember it is what she said next, which was that in eight years of working with agencies, she had never once heard about a problem before her CEO did.

That is the change. Working with an AI powered agency should mean earlier problem detection, less administrative chasing, and strategy calls spent on decisions and insight rather than reciting numbers.

If you already work with an agency, or you are weighing up a new one, that difference matters quickly. It changes how fast issues surface, how much time your team spends on follow-up, what still needs human judgment, what should never be automated, and how to tell whether an agency is actually using AI well or just talking about it.

Three things a client actually notices

The chasing stops

The report arrives on the second, every month, without anyone asking. The follow up from the call arrives the same afternoon rather than the following week. The thing you mentioned in passing does not need mentioning again.

This sounds small written down. It is not small to live with. A large part of what clients quietly hate about agencies is the administrative load of managing one, and most of that load is chasing. Remove it and the relationship feels different before a single strategic thing has changed.

Problems arrive early

Rankings move, a page breaks, a campaign underdelivers in week two rather than week four. In the old model you find out when a human happens to look, and humans happen to look on a schedule that suits the agency.

The change is that things are being watched continuously, so the gap between something going wrong and someone telling you shrinks from weeks to about a day. You will notice this most in the conversations you no longer have to have, which is a strange thing to notice and takes a couple of months to register.

There is a failure mode here worth naming, since we hit it. Early on we tuned the monitoring too sensitively and a client got three alerts in a fortnight, two of which were nothing. That is worse than silence. Being told about a problem that is not a problem trains people to ignore you, and the whole value of arriving early evaporates the second the client starts skimming your emails. We loosened the thresholds and added a rule that a human reads anything before it goes out.

The strategy call has strategic planning in it

The monthly call used to be forty minutes of walking through numbers the client could have read, and ten minutes of actual thinking at the end when everyone was tired.

Now the numbers arrive beforehand and have already been read, so the call opens at the interesting part. This is the change that clients like most in hindsight and notice least at the time.

One more thing clients notice, though usually only when they compare notes with someone at another agency: the answers stop depending on who is available. Ask about something from March and you get the same answer whether you get the account lead or their colleague, because the history is in one place rather than in one person. Agencies lose an enormous amount of client goodwill to staff turnover, and most of that loss is not skill. It is context walking out of the building.

What does not change

Everything that was ever worth paying for.

The judgement is still a person’s. Someone still decides what the data means, which is a genuinely different activity from producing the data, and one that nothing has got close to automating.

The argument is still a person’s. Any agency worth its retainer will at some point tell a client that the thing they want is a bad idea, and hold the position through two rounds of pushback. No system does that.

And there is still someone whose name you know, who is accountable, and who you can ring. If you rewire an agency and the outcome is that you now talk to fewer humans, something has gone badly wrong with the design.

What no client should let an ai automation agency automate

I want to be specific here, because a vague version of this section is worthless.

No client should accept an automated recommendation to change budget. A machine can tell you a channel is underperforming. Whether to move money is a business decision with context the agency does not fully hold, and it needs a person who is accountable for being wrong.

No client should accept anything published to their site or their social accounts without a named human approving it. Not reviewed by a system. Approved by a person, who can be asked about it afterwards.

No client should accept an automated response to a complaint or a crisis. The first reply after something goes wrong sets the tone for the entire recovery, and it is exactly the situation where a plausible-sounding generic response does the most damage.

And no client should accept an agency that will not tell them what is automated. This is the one I would push hardest on. If you ask which parts of your work involve AI and the answer is a shrug or a sales line, that is a governance problem, not a technology problem. Regulators have been fairly direct about this: the FTC’s guidance on AI claims is essentially a long way of saying that if you say a thing about AI in your business, you should be able to substantiate it.

We hold those four lines, and we have moved one of them twice as we learned more. Rewire covers where we draw that line, and why it moved, in more detail than belongs in this post.

How to tell whether your agency has actually done this

Three questions. They take about ten minutes to ask and the answers are hard to fake.

Which of your processes has improved operational efficiency and has fewer steps than it did a year ago?

Not faster. Fewer steps, fewer handoffs, fewer people touching it. If the answer is a tool name rather than a process, nothing structural has changed.

Who reviews the work before it reaches me, and what are they checking for?

A good answer is specific and names a person. A bad answer describes a system.

What happened the last time your AI generated content got something wrong on my account?

This is the one. Every agency doing this seriously has a story, because everyone doing this seriously has been wrong. An agency that has never caught an error is an agency that is not checking.

If you want a structured version of this conversation rather than three questions in a meeting, our Revenue-First Diagnostic is built for exactly that, and it looks at the commercial side too.

The point

Clients notice the operations change months before they notice the strategy change, and by the time they notice the strategy change they will have stopped being able to tell you why the relationship feels different.

It was Tuesday instead of Thursday. That is most of it.