"It will not run ad campaigns for you as an advertisement led me to believe." That line, from an App Store review of an AI ads app, sums up the gap founders keep finding with ai agents for business. The promise sounds like delegation. The product turns out to be assistance. Meta's new free built-in agent will be useful, but useful in a narrow place, for a narrow job.
If you sell on Meta, you will meet this agent inside the ad account. It lives where Meta makes money, it knows what Meta can see, and it stops where Meta stops. That design is not a flaw. It is the point. A platform agent is built to help you spend more smoothly on that platform. It is not built to protect your margin, grow your store, or tell you to spend less.
What can ai agents for business do inside Meta?
Platform agents are good at tasks with clear borders, fast feedback and low risk if the answer is only partly right.
Drafting ad variants. Give it a winning ad and it can produce new hooks, headlines and copy angles in your account, ready for you to approve. That saves copywriting time on first drafts.
Surface-level suggestions. It can spot a tired creative, a broad audience or a budget split that looks odd against basic patterns, then suggest a change. The suggestion is a starting point, not a decision.
Tidying work inside one tool. It can summarise performance in plain language, group similar ads and help you find things faster. That is real help, in the same way a tidy desk is real help.
None of this is trivial. Small teams lose hours to exactly this work. But none of it answers the only question a sales-first founder should ask: did it create sales that would not have happened anyway?
What will a platform agent not do for revenue?
This is the work that sits outside Meta's walls, and it is where growth is decided.
Own a revenue outcome. A platform agent is measured on activity inside the platform, not on your profit. If sales stay flat, there is no owner to question, no target it missed and no one who has to change the plan. An outcome needs a name against it.
Work across your stack. Your result does not live in the ad account alone. It lives in the link between the ad, the product page, the offer, the email follow-up and the repeat purchase. A Meta agent cannot rewrite the landing page the ad sends traffic to, fix the abandoned cart flow that should recover the click, or spot that your best ad is selling a product with no stock depth. It sees one room in the house.
Learn your margins and standards. Two products can return the same revenue and leave very different profit. One customer can be worth five times another over a year. Your brand also has lines it will not cross in copy, imagery and claims. A platform agent does not know those lines unless a human teaches it, checks it and keeps teaching it as the range and the market change.
Run the testing cadence end to end. Real testing is a rhythm: form a view, build the variants, set the budget rules, read the result honestly, kill losers fast, scale winners and write down what was learned. Drafting variants is step two of seven. Most accounts do not fail for lack of variants. They fail because the rhythm stops when the founder gets busy.
Be accountable when sales do not move. If a tool gives bad advice, you close the tab. If a team owns a number, it has to explain the miss, change course and show its working next week. That pressure is not a nice extra. It is what turns suggestions into learning.
This is the pattern behind AI theatre: impressive activity on the surface, no movement on the number underneath.
Find the work worth handing over
The Growth Audit maps where revenue is leaking across ads, store and email, then names the first job an agent should own. Five questions, no generic AI tour.
Does it create sales, or does it create content?
Use that as the test for any free agent Meta, Google or anyone else offers you.
Content is easy to count: more variants, more posts, more copy, more suggestions accepted. Sales are harder and more honest: revenue per visitor, conversion rate on the page the ad lands on, repeat purchase rate, and profit after ad spend. If a tool report talks only about the first list, it is reporting its own effort, not your result.
The stronger setups we see work the other way round. They start with the revenue line, set a baseline, then hand an agent a job with that number attached. That is the argument in agentic AI, minus the hype: an agent earns its place by running a job to completion, not by answering faster. The proof standard is the same one we use in client work, where the Growth Sprint recorded 518% average revenue growth across clients who completed the 90-day Growth Sprint, and Bottled Baking Co hit 27x, the highest return on ad spend, in 14 days. Those numbers came from joined up work across ads, store and email, not from a single tool working alone.
Free is the right price for a tool with no outcome attached. Use Meta's agent for drafts and first checks. Save your serious judgement for whether anything changes on the sales line.
How should you judge the next free agent you are offered?
Ask four questions before you give it any real work.
What number will move if this works? If the answer is time saved on drafts, price it as time saved. If the answer is revenue, demand a baseline and a review date.
What can it not see? List the parts of the journey outside its walls. If the bottleneck sits there, the agent cannot fix it, however good the demo looks.
Who owns a miss? If nobody owns a miss, you own every miss. That may be fine for a free tool. It is not fine for your growth plan.
What does it learn about my brand? If it learns nothing about your margins, customers and standards, its hundredth suggestion will be no better than its first.
Take Meta's free agent for what it is: a helpful assistant inside one platform, worth using for narrow tasks. Do not mistake it for the harder work of joining up your stack and owning a result. If you want that second job mapped for your brand, start with the Growth Audit and see which piece of work should be handed over first, with a number attached to it.