The honest answer is yes for most of the work and no for the parts that decide whether the work was worth doing — and the line between them is sharper and more useful than either the enthusiasts or the sceptics make it sound. Here is exactly where it falls.
What AI genuinely does well now
Audience construction
Reading a platform's live targeting catalogue and assembling layered segments. Mechanical, rule-governed, and better done by something that never gets bored on the four hundredth interest.
Keyword research
Pulling real search volumes and structuring ad groups. The caveat is real data versus generated lists — see below.
Creative production
Generating a scene and drawing your offer over it as readable text. Twenty variants instead of two, and you pick.
Campaign structure
Objectives, budgets, placements, the ad set mechanics. This is where most self-run campaigns quietly fail, because the errors produce no warning.
Watching and alerting
Noticing that spend spiked, delivery stopped or a lead arrived. Automation is strictly better than a person here because it never sleeps.
What it cannot do, and will not learn to
Three things, and they are not temporary limitations waiting on a better model.
It does not know why customers choose you
Not that you open on Sundays, or that your quote includes the parts, or that people come because you actually answer the phone. Those are what make an ad work, and they exist only in your head until you write them down.
💡 The single highest-leverage thing anyone does with one of these tools is write the business description properly. Two honest sentences about who your customer is and why they pick you will produce a better campaign than any amount of switching between products. Most people type something generic, get a generic campaign, and conclude the category does not work.
It does not know what a customer is worth to you
This is the number that makes every other number interpretable, and no tool can supply it. A cost per lead of 200 is excellent if a client is worth 15,000 and ruinous if an average sale is 80 — same figure, opposite decisions. Until you work this out, neither a tool nor an agency can tell you whether your campaign is succeeding.
It does not carry your legal responsibility
An AI tool will write "the best in the city", "guaranteed results" and "trusted by 5,000 customers" without hesitating, because that is what persuasive ad copy looks like in its training data. Every advertising regulator holds the advertiser responsible, and "the AI wrote it" has no standing anywhere. The model has no idea those sentences carry weight.
Three failures that are the tool's fault, not yours
If you have tried an AI ad tool and it did not work, one of these is the likely reason — and knowing them lets you test a product before paying for it.
Targeting shipped inside the product instead of read from the platform
Meta exposes a catalogue of roughly 284 behaviours through its API, and returns the same catalogue regardless of which country you ask about — the country-specific entries sit inside one shared list. A tool with a hardcoded shortlist silently discards everything outside it.
We found exactly this in our own product: a hand-maintained list of fifteen behaviours was filtering the model's selections and throwing away 161 of 177 chosen entries for having no matching ID. The list had been written for one market, so that market's campaigns looked fine and every other market ran without the signals that existed for it. Nothing errored. Nothing in any report showed it.
Keywords invented rather than researched
Ask whether keyword selection uses live Google Keyword Planner data through the official API, or whether a language model is asked which keywords are popular. The second is common — Planner access requires an approved developer token and guessing does not — and produces lists that read beautifully and contain terms nobody searches.
💡 Test it in two minutes: take five suggested keywords into Keyword Planner with your country selected. If several show no volume, the list was generated. The campaign built on it will get almost no impressions and nothing will explain why.
Constant editing of live campaigns, sold as optimisation
Every edit to a live ad set restarts the platform's evaluation of it. An ad set adjusted several times in a morning can therefore serve nothing for the rest of the day on an account that is delivering perfectly well elsewhere — and nothing in the interface tells you that is what happened.
So treat "continuous AI optimisation" as a claim requiring evidence. Adjustment on real signal after a learning period is valuable; rewriting live ad sets daily is interference wearing a better word.
What a realistic division of labour looks like
You decide
What you sell, to whom, what a customer is worth, and whether a claim is true. Ten minutes of thinking that nothing can do for you.
The tool builds
Audience, keywords, creative, campaign structure, and the payload the platform will accept. Hours of work that produces no differentiated advantage when done by hand.
You approve
Every image and headline before it publishes. This is a compliance step, not a control preference.
The tool watches
Delivery, cost per result, and telling you the moment a lead arrives. Response time decides conversion, and nothing human is awake at 2am.
So: can it?
It can build and run a technically correct campaign better and faster than most people manage by hand, and it can watch that campaign more reliably than any human will. It cannot tell you what to sell or whether the numbers are good. If you can answer those two questions, the answer is yes. If you cannot, no tool and no agency will rescue the campaign — and that is worth knowing before you spend money on either.
Build it in ten minutes, approve it before it runs
Adyft builds Google and Meta campaigns from a short description of your business, launches into accounts you own, and shows you everything first. 14-day free trial, no card.
Start free — no card needed