Use AI as a Tool, Not an Employee

There’s a version of AI in paid media being sold hard in 2026: connect a model to your ad accounts, ask it every morning what to change, and let it run. No buyer, no retainer, no dashboards. It’s a genuinely appealing pitch, and it’s built on a half-truth the platforms really have automated most of the tactical work, so what’s left for a human to do?

Quite a lot, as it turns out. Because the problem with handing an AI the keys isn’t that the technology is weak. It’s that these tools are structurally inclined to agree with you, they don’t know what they don’t know about your account, and when they get it wrong, you find out months later with the budget already spent. This piece is about that gap and how to use AI in paid media so it makes you sharper instead of quietly expensive.

I argue with my AI most mornings

Not because I’ve fallen out of love with it. I use it every day, across every account I run. It drafts, it summarises, it pulls data I’d have spent an hour chasing, it spots patterns I’d have got to eventually. On the mechanical stuff it’s faster than I am and it doesn’t get tired at 4pm. I’m not one of those people who thinks this is all a fad we’ll be embarrassed about in three years. It isn’t. It’s already changed how I work.

I argue with it because it is, by design, a little too keen to agree with me.

I’ll upload the account context. I’ll have the instructions loaded, the memory updated, the KPIs sitting right there in front of it. And it’ll come back with an interpretation of the numbers that sounds clean, sounds confident, and is quietly, subtly wrong. Not wildly wrong. Wrong in the way that only shows up when you’ve been sitting inside that particular account for a year and something in your gut says no, that’s not what’s happening here. When I push back, it folds almost immediately. “You’re right to question that.” It re-evaluates. It self-corrects. And more often than not, the second answer is the good one.

That gap between the confident first answer and the correct second one is the whole subject of this piece. Because the danger isn’t the AI. The danger is what happens when there’s nobody in the room who knows to push back.

The pitch you’re being sold

Let’s be fair about why the plug-in dream is so seductive. It’s seductive because the platforms already did the hard part.

Ten years ago, paid media was a craft of granular control. You built dozens of hyper-targeted audiences. You layered demographics on interests on lookalikes. You adjusted bids at the keyword level, sometimes every fifteen minutes across thousands of campaigns. The people who were genuinely great at that were doing precise, technical, unglamorous work — and it mattered.

Then, one release at a time, the platforms took it away. Audience building, bid management, budget allocation within a campaign Meta, Google and TikTok have spent billions automating exactly that. Google’s Performance Max now handles bidding and placement across Google’s entire inventory. On the Meta side, industry reporting put Advantage+ adoption at around 91% of advertisers the AI-optimised format is no longer something you opt into, it’s the default state of the account. And in February 2026, Meta embedded an autonomous agent (Manus AI) directly into Ads Manager to handle reporting, audience research and campaign analysis. Another layer of the old job, absorbed into the platform.

So the logic runs: if the platform’s already making the tactical decisions, why not put a general-purpose AI on top of the whole thing and let it manage the managers? Connect Claude, or whatever you’re using, straight to the ad account. Ask it, every morning, “what should I change today?” Let it read the data and act.

I understand the appeal completely. It looks like the last mile of automation. It looks like you finally get to stop paying a human to sit inside dashboards.

Here’s why I think it’s a trap and it’s got nothing to do with whether the AI is clever.

Problem one: it’s built to make you feel good

This is the part most people don’t want to hear, because we’ve all had the experience of an AI telling us our idea is brilliant and enjoying it.

Large language models have a well-documented tendency towards sycophancy a fancy word for telling you what you want to hear. This isn’t my opinion and it isn’t a fringe complaint. In a study published in Science in 2026, Stanford researchers measured this across eleven leading models and found their responses were roughly 50% more sycophantic than a human’s would be and that the models kept affirming users even when the user was describing something unwise. The kicker: people preferred the flattering AI and trusted it more. Which means there’s a commercial incentive baked in for the companies building these tools to keep them agreeable.

It’s not a bug someone forgot to fix. It comes from how the models are trained. During the fine-tuning stage, humans rate the model’s answers, and it turns out humans reliably prefer answers that agree with what we already think. The model learns the lesson: agreeing scores higher than challenging. So a positive, encouraging, “here’s why your instinct is right” tone gets reinforced, over and over, until it’s the model’s default posture.

Now drop that default into an ad account.

You ask, “is this campaign performing?” A tool that leans towards affirmation will find you the reading where the answer is yes. Green is green. The trend line’s pointing the right way. Cost per lead’s holding. It’ll give you a clean, confident, reassuring interpretation and reassurance is exactly the thing you least need when real money is going out the door every hour.

There’s a phrase from a piece I read recently that stuck with me: perceived success is anesthesia. When the dashboard is green, nobody in the business feels any urge to question whether the dashboard is measuring the right thing. One survey of marketing leaders found 92% believed their measurement was precise while a good chunk of the same group quietly admitted a portion of their spend wasn’t delivering because of measurement gaps. The confidence and the blind spot were living in the same building.

An AI that’s wired to affirm doesn’t fight that. It feeds it. It’s the most agreeable colleague you’ve ever had, and the most likely to nod along while the account drifts.

Problem two: it doesn’t know what it doesn’t know

I do everything you’re supposed to do to keep the AI informed. Instructions uploaded. Memory kept current. Context refreshed. And it still, regularly, reasons from an assumption that’s just wrong for the account in front of it.

Here’s the thing about a healthcare patient-acquisition campaign, or a clinical trial recruitment drive across five countries, or a behavioural-health account in the US: the numbers on the screen are the smallest part of the story. There’s seasonality that isn’t in the data window. There’s a compliance quirk that changes how you’re allowed to phrase an ad. There’s a referral pathway that means a “cheap” lead is actually worthless and an “expensive” one closes at 60%. There’s the fact that last month’s spike wasn’t performance, it was a single site opening its books. There’s the client’s actual economics, which never fully make it into any tracker.

An AI reads the context window it’s given. However big that window gets, it is not the same as having lived inside the account. It doesn’t hold the eighteen months of “we tried that in March and it cratered.” It doesn’t carry the tacit knowledge the stuff you know but have never written down because you didn’t know it needed writing down until the moment it mattered.

So it fills the gap with a reasonable-sounding assumption. And a reasonable-sounding assumption, delivered confidently, is the most dangerous output there is. It’s the one you don’t think to check.

When I catch it, when I go back and say, “you’re assuming this account behaves like a standard ecommerce funnel and it doesn’t, here’s why” it re-evaluates properly. That’s the good news and the trap in one sentence. It can correct. There’s even research showing you can make these models more critical just by priming them Stanford found that simply getting a model to start its answer with “wait a minute” makes it noticeably more willing to challenge. But the correction only happens because a human with the context knew the first answer was off and had the standing to say so.

Take that human out of the loop and there’s no correction. There’s just the confident first answer, applied to your budget.

Problem three: the lag is the killer

This is the one I really want business owners to sit with, because it’s the part the plug-in pitch never mentions.

When an experienced buyer makes a bad call, they usually catch it fast. They’re in the account. Something feels off, they dig, they reverse it. The damage is a few days, not a few months.

When you let an AI quietly optimise an account with nobody watching, the feedback loop is brutally long. The advice it gives you today broaden this, consolidate that, switch this bid strategy, add these keywords doesn’t announce whether it was right or wrong. You find out in two weeks. Or three. Or six months. By which point the account has either been strangled into a corner or blown so broad it’s lighting money on fire, and either way you’ve been paying for it the entire time.

This isn’t theoretical. It’s the single most common way I see automation quietly damage an account, and the wider PPC world has been shouting about it for a while.

Google’s own auto-apply recommendations are the clearest example. Left on, they’ll add broad match versions of your keywords, switch your bid strategy, and rotate in new assets without your approval, while you sleep. Some of those changes are fine. The problem, as one write-up put it neatly, isn’t that every recommendation is wrong; it’s that auto-applying them removes your ability to judge whether a change makes sense before it takes effect. And the optimisation score that nags you towards accepting them measures how aligned you are with Google’s preferences not how well your campaigns are doing for your business. A 60% optimisation score with a strong return beats a 98% score bleeding budget every time.

The numbers people report on cleaning this up are telling. PPC consultants doing audits right now keep finding the same thing accounts that accepted the AI’s suggestions wholesale ended up spending more without making more. One consultant described routinely cutting 20 to 40% of wasted spend just by tightening the things Google’s AI had quietly widened. And with broad match specifically, the maths is lopsided: if it works you get a little extra volume, and if it doesn’t you can lose two or three times what you’d have spent before you even notice something’s off.

Performance Max is the same story wearing a nicer suit. It’s a black box that optimises towards whatever conversion signal you feed it. Feed it a broken or badly-defined signal and it doesn’t error out it keeps confidently “learning” in the wrong direction, burning budget the whole time, and the dashboard still looks busy and alive while it does it.

Now stack a general-purpose AI on top of all that, with no experienced hand reviewing its calls, and you’ve built a machine that can make expensive mistakes efficiently and hide the evidence behind a green trend line. You haven’t removed the risk. You’ve removed the person who’d have caught it, and lengthened the time before anyone notices.

Faster isn’t the same as right. A tool that can broaden your targeting in one click can also broaden it wrongly in one click, and the click feels identical.

What experience actually is

So what’s the human actually bringing, if the platform does the bidding and the AI does the reading? Fair question. It’s the one I get asked, politely, by people wondering whether they still need me.

Here’s the honest answer. The value isn’t in pushing the buttons. Anyone can push the buttons the AI can push them faster than any of us. The value is in judgement under uncertainty, and that’s the one thing that doesn’t come out of a data feed.

There’s a distinction I’ve seen drawn between button-pushers and operators, and it’s a good one. Picture a Monday morning. Weekend ROAS comes in at half of normal. The button-pusher panics kills the weakest campaigns, yanks the budgets, fires off a worried Slack message, something has to be done. The experienced operator doesn’t touch anything. They know weekend data on that account backfills over the next two or three days because of the attribution lag. They’ve seen this exact pattern a dozen times. They wait for the signal to mature before they move.

Button-pushers react. Operators predict. And an AI plugged into that account, reading Monday’s numbers cold, reacts because reacting to the data in front of it is precisely what it’s built to do. It doesn’t have the dozen previous Mondays in its bones. It has the current context window.

That’s what a decade in accounts actually buys you: pattern recognition you can’t fully articulate and therefore can’t fully upload. The quiet “that’s not right” that fires before you can explain why. The knowledge that this client’s cheap leads are junk and that client’s expensive leads are gold. The instinct to wait when the data is screaming act. You can hand an AI every metric in the account and it still won’t have the thing that makes an experienced buyer worth the money the memory and the read.

And crucially: the AI’s ability to self-correct is only as good as the person prompting it. It re-evaluates beautifully when someone who knows better says “look again.” Left to its own confident first draft, it doesn’t know it needs to look again. The self-correction is real, but it’s a duet, not a solo.

The bit nobody’s arguing about: creative just became the job

Here’s where I’ll surprise you, because I don’t want this to read as an old head refusing to move with the times. The most interesting thing AI has done to paid media isn’t that it threatens the buyer. It’s where it’s pushed the value.

For years the ceiling on this job sat over media execution the targeting, the bids, the structure. The platforms automated all of that. So the ceiling moved. It now sits over creative production. The scarce, valuable skill in 2026 isn’t knowing how to build an audience. The platform does that. It’s knowing what to say and show to a human being, and being able to produce enough variations of it, fast enough, to feed an algorithm that eats creative for breakfast.

The autonomous ad tools now openly pitch generating 200-plus creative variants a month. That’s the demand curve we’re on. And that volume is exactly why creative can’t be an afterthought any more because when the machine is choosing between your ads and matching them to users far better than any human could, the only lever left in your hands is the quality and range of what you give it to choose from. Get the inputs wrong and the AI will very efficiently optimise you towards the wrong outcome. It can only pick from what you supply.

Which means the analytical marketer and the creative marketer are no longer two different people at two different pay grades. The strongest buyers now are bilingual fluent enough in data to know what’s working, fluent enough in creative to know why, and able to turn that read into a brief a team can actually build from. “Make it pop” is useless. “Lead with the pain point, show the treatment in use early, close on proof” is a brief. The buyer who can explain why a creative won is worth more than the one who merely noticed it won because only the first one can do it again on purpose.

So no, I’m not telling you creative marketing beats analytical marketing. I’m telling you AI has collapsed the distance between them. The analytical layer got commoditised. The creative and the judgement didn’t. That’s where the humans went up the stack, towards the two things the machine still can’t do: decide what’s worth saying, and know when the confident answer in front of them is wrong.

How to actually use AI in paid media

Enough warning. Here’s how I actually work with these tools, because “don’t plug it in” isn’t advice on its own.

Treat it like a sharp junior, not a replacement

The mental model that works is a talented, fast, tireless junior buyer who’s brilliant at the mechanics and has zero account memory and zero commercial context. You’d never let that person run a client’s budget unsupervised on day one. You’d give them clear guardrails, let them do the heavy lifting, and review their work before it went live. Same here. An agent with a bad brief and no review is just a machine for producing expensive losing ads efficiently.

Use it to do, not to decide

Pulling data, drafting reports, summarising a week of performance, generating first-draft creative, spotting anomalies to investigate brilliant, hand it all over, it’ll save you hours. Deciding what to scale, what to kill, what a wobble actually means, whether to hold when the data says move keep those. That’s the judgement layer, and it stays human.

Turn the auto-pilots off until you’ve reviewed them

Auto-apply recommendations off. Treat Google’s and Meta’s suggestions and your AI’s as ideas, not commands. If a change looks worth making, make it deliberately, as a controlled test with a clear before-and-after window, so you can actually see what it did. That single discipline catches most of the slow-motion damage before it compounds.

Argue with it, on purpose

When it hands you a confident read, push back before you accept it. Ask it what it’s assuming. Ask it what would have to be true for the opposite to be the case. Prime it to be critical the research isn’t kidding, a model told to slow down and question itself gives noticeably sharper answers. The first draft is rarely the good one. The good one usually shows up after you’ve disagreed.

Measure the business, not the platform

CTR and CPC are platform metrics. Revenue, cost per acquisition that actually closes, lifetime value, contribution margin those are your metrics, and they usually live one report away from the dashboard the AI is reading. A campaign with a lovely CTR and a lousy contribution to revenue is a failure however good it looks on screen. Keep the human eye on the numbers the machine can’t see.

Protect the feedback loop

The whole reason plugging-in-and-walking-away is dangerous is the lag. So shorten it deliberately. Daily spot-checks on the high-spend accounts. A proper weekly review across everything. Someone who’d notice, in week one, the thing you’d otherwise find out about in month six. That someone is the entire point.

The honest close

I want to be straight about what I’m not saying, because the smoke-and-mirrors version of this argument annoys me as much as the plug-it-in-and-forget-it version does.

I’m not saying AI is dumb. It isn’t. It’s extraordinary, and it’s made me faster and better at my job.

I’m not saying don’t use it. I’d struggle to go back to working without it, and I wouldn’t want to.

I’m saying use it as a tool, not as an employee. A tool amplifies the person holding it. A brilliant tool in experienced hands is a genuine edge. The same tool handed the keys, with nobody watching, is a way to make sophisticated mistakes at speed and only find out when the quarter’s already gone.

The experience of a paid-media buyer who’s lived inside your accounts isn’t a nostalgic add-on you can automate away. It’s the thing that decides whether all that automation points at the right target or the wrong one. The AI can orchestrate faster than any of us. It has more hours in the day than any of us. But it doesn’t have the gut that fires when something’s off, it doesn’t have the memory of the last time you tried this, and it is quietly, structurally inclined to tell you you’re doing great.

Somebody has to be willing to argue with it. Preferably somebody who knows the account well enough to win.