Before You Fire Your Ads Agency for AI, Read This

Firing your agency and handing Google Ads over to Claude is a big question for a lot of business owners right now. Claude’s the example here because it’s the one showing up most in these conversations, but everything below applies just as much to any AI agent wired into your ad accounts. If a piece of software can genuinely do your ad agency’s job, stop paying for the agency. That’s the obvious move, and we’d make it too in your position.

Big tech is laying off people by the thousands and pointing at AI as the reason. If Google, Microsoft, and Meta are comfortable doing that, why not us? Hand the ad account over to Claude, fire your agency, and get praised for finding efficiencies!

Here’s what that thinking misses, and it’s the same thing a lot of “just use AI” advice misses.

AI is a genuinely great assistant, researcher, accelerator, and amplifier of skills a person already has. It surfaces ideas we wouldn’t have landed on alone. It catches things a tired human eye would miss. Work that used to take five days can get done in one, at the same standard or better. It’s the most useful piece of technology we’ve worked with since the internet itself showed up.

None of what follows is an argument against that. It’s an argument against mistaking the tool for the person driving it.

This is a distinction that matters. “Using AI” actually covers two very different setups.

One is AI as an advisor: it analyzes the account, flags issues, suggests changes, and a person decides and implements what actually goes live. The other is an autonomous agent: it’s connected directly to the ad platform and makes changes on its own, with nobody reviewing before they take effect.

Both get called “using AI instead of an agency.” The risk profiles aren’t close to the same. Stick to the advisor setup and you’ve basically just described what a well-run agency already does with AI. Let it slide into the autonomous version, and an AI agent working off a bad assumption about bidding logic or targeting can burn budget quietly for days before anyone notices, because nobody’s watching each individual change as it happens.

This is the real alternative, and it’s the question the “fire your agency” pitch never forces anyone to answer honestly. For a lot of small businesses, a full-time hire isn’t even the cheaper option once you count salary, benefits, and overhead against an agency retainer, so the savings this pitch promises don’t always exist in the first place. But say they do, in your case. Even then, cost was never actually the problem worth solving.

The real gap isn’t about how good that one person is. It’s structural. An agency keeps seeing new issues across dozens of live accounts as the platform changes, and that knowledge spreads across the team. A solo hire’s experience is frozen at whatever they walked in with, and when something looks off, they’ve got nobody to check it against but the same AI that flagged it as fine, which isn’t a second opinion, it’s the first one talking to itself. Neither gap closes with AI.

Watch a skilled ads specialist run an AI workflow and it looks like magic. Fast, thorough, better than what most humans could produce solo. The temptation is to conclude: the AI did that.

It didn’t. The ads specialist did, using AI as a tool.

Google Ads coach Jyll Saskin Gales wrote about this exact scenario. Her client asked Claude for bidding strategy advice, and Claude recommended switching strategies based on logic that was technically sound. What it didn’t know was that the account’s conversion volume was too thin to support the switch, and that conversion values weren’t syncing properly with the client’s CRM in the first place. Budget spiralled, conversions dropped, and it took an ads expert to catch what the AI couldn’t see.

The dangerous part isn’t that AI gets things wrong. It’s that the output looks just as polished when it’s wrong as when it’s right. A bad suggestion doesn’t arrive looking half-finished or shaky. It arrives looking exactly as confident and clean as a good one, which is precisely what makes it easy to wave through if someone doesn’t have the experience and expertise to spot the difference.

AI’s knowledge has a shelf life. It comes from training data with a cutoff, and Google Ads doesn’t stop moving after that: new campaign types, shifting Smart Bidding behaviour, policy updates, auction dynamics that change what “correct” account structure even looks like. The tool doesn’t always know when its own information is out of date. It’ll confidently recommend something that stopped being true six months ago, because nothing in the model flags that anything changed.

Strip the professional out and keep only the AI, and you don’t get a faster version of the same output. You get an unsupervised system making live account decisions, run by someone who has no way to tell when it’s wrong or how badly. When everyone’s using the same AI tools, the output stops being an advantage—the person directing it is what separates good from average.

Give an AI tool direct access to a Google Ads account, no agency, no PPC background in the room, just a prompt and a login, and the outcome is predictable:

• Campaigns paused that shouldn’t have been
• Budget increases that don’t make sense for the business
• Ad copy pushed live that missed the brand voice entirely
• Bidding strategies changed with zero understanding of context

These aren’t edge cases. This is what happens when the tool gets access and the expertise doesn’t come with it.

“I audited my account in nine minutes with AI” is everywhere on LinkedIn right now, and it’s not true (or it is, and the audit is worth exactly what nine minutes gets you) because it badly undersells what a client-facing audit actually costs in time and years of earned insight and context, even with good AI tooling.

A real audit means hours gathering context on the business (brand history, what each sale is actually worth, why past decisions were made), time spent manually going through the account to understand what’s actually in there, and AI running structured checks across bidding, budgets, and performance in the background. Then someone reads that output, sense-checks it against everything they know about the business, and decides what actually matters. That reading and deciding part is the job. It’s what you’re paying an agency’s hours for, not the software running in the background.

Every example in this post so far shares one root cause: AI producing a bad answer because it didn’t have the full picture. That’s not a flaw that gets patched in the next model update. It’s how these tools work. Feed it thin, generic context, and it gives back a thin, generic answer dressed up to sound specific.

An agency’s context comes from actually working with your business: sitting through the calls where a pricing decision got made, watching a launch go sideways and knowing why, remembering that this client’s ad account always tanks in February for reasons that have nothing to do with the ads. A prompt, even a good one, is a compressed summary of all of that. However well it’s written, it’s still someone’s best attempt to type out what an agency picked up by being in the account for years.

Even Claude’s own skills feature proves the point. Skills exist specifically because a one-off prompt doesn’t carry enough context: you build a folder of instructions and background that Claude loads automatically, so nobody has to re-explain the same details every conversation. That’s a real, useful admission that context is the bottleneck, not model quality. But someone still has to write that skill, and writing it well requires the same account knowledge this whole post is about. If whoever builds it doesn’t have that knowledge, the skill doesn’t fix the gap, it locks it in. Claude will apply that same blind spot automatically and confidently, every single time it loads, instead of just getting one prompt wrong once.

Better inputs get better outputs. That’s true, and it’s also exactly the point. Getting to “better inputs,” both the context and the skill to use it well, is most of the actual work, and it’s the part “just use AI instead” skips over every time.

The usual pushback: “AI in PPC isn’t dangerous if you set it up correctly and manage it well.” True. And setting it up correctly is exactly the expertise the “fire your agency” pitch tells you to skip.

Guardrails aren’t a setting. They’re a person who reviews everything the AI produces, checks it, verifies it, and only implements what actually holds up against what they know about the business and the client.

Ask anyone doing this properly how much of what AI suggests they actually act on, and the honest answer is less often than people’s exaggerated LinkedIn posts would have you believe. That gap between “suggested” and “implemented” is where the judgement call happens, and it’s the part no tool does for you.

That review work doesn’t happen once and then coast. It’s a recurring job:

• Checking what the AI flagged against what’s actually happening in the account
• Adjusting prompts and workflows when its recommendations don’t match reality
• Re-checking assumptions every time Google changes something
• Keeping the client knowledge behind all of it current as the business changes

None of those hours disappear when an agency gets swapped for a subscription. They just move onto whoever’s now responsible for catching what the AI misses (usually the same person who was told AI would save them time).

Starbucks Korea is a cautionary tale. A promotion called “Tank Day” launched on May 18, the anniversary of the 1980 Gwangju Uprising, a date loaded with meaning for exactly the reasons that name suggests. Marketers had used an AI tool for slogan suggestions during development. Nobody in the approval chain caught the connection, partly because some people signing off reportedly never opened the files. The campaign was pulled within hours. The local CEO was fired. The chairman apologized on national television. The company closed thousands of stores for a day of mandatory sensitivity training.

That missing review layer is what an agency, or any team of people who know the terrain, actually is. Remove it, and you’re betting that nobody happens to catch the mistake before it goes live.

What AI can’t be, no matter how good the model gets, is a strategy department, a cultural intelligence layer, institutional memory, or wisdom. What it can do is produce, fast, remix what already exists, and polish a bad idea until it looks clean enough to approve.

The hardest failure mode to catch isn’t one bad suggestion you can point to and fix, like a paused campaign or ad copy that misses the brand voice. It’s the kind that builds quietly over weeks, with no single moment where anything looks obviously wrong.

You don’t know to ask a question, so the AI never flags the issue behind it, and the AI’s confident answer is exactly what stops you from thinking to ask in the first place. Attribution quietly breaks. A structural problem compounds under the surface while the top-line numbers still look fine. Neither side catches it, not because either one is careless, but because catching it depends on someone recognizing what should have been asked and wasn’t.

Who reviews every AI suggestion before it goes live, and do they actually understand the business, strategy, and ad account? “The AI reviews itself” isn’t an answer. It’s the problem, restated as a plan.

Is that person in ad accounts every day, or just familiar with marketing in general? Those aren’t the same skill. Ads specialists build an instinct for what normal looks like on a specific account, the kind of instinct that flags a bad number before anyone can explain why it’s bad. Someone generally comfortable with marketing, checking in occasionally, is reading every number cold, and so is the AI they’re supervising.

If it’s one person and a Claude subscription, what happens the week that person is out sick or gone for good? An agency doesn’t stall because one person took a vacation. A single point of failure does, every time. (Also, if you plan to use AI to make autonomous changes to your account, you’re going to have a bad time.)

What happens when the AI gets it wrong? If the honest answer is “nobody knows,” that’s not a detail to fix later. That’s the entire reason the job exists. AI isn’t perfect. It isn’t infallible. And yes, that’s even for the really well built ones. (There’s a reason why Google has “AI responses may include mistakes.” after its AI Overview responses.)

There’s no honest version of these answers that ends with “so we fired the agency.” If someone’s telling you otherwise, ask them which of these four they actually solved.

AI didn’t build your account’s history. Your agency did. Fire the wrong one and you’ll find out which one was actually doing the work.

Not sure whether your account’s actually being run well right now, AI-assisted or not? We do ads audits for exactly that reason—a second set of eyes from people who live in ad accounts every day.