How do I set up an AI agent to optimise Meta ads campaigns automatically?
Set it up in this order: fix tracking and account structure, define "optimise" as one goal metric with guard metrics, connect the agent read-only, set a hard spend cap and approval rules, then let it observe and propose for two to four weeks before anything runs on its own. Use Meta's own automation for budget distribution, an agent for diagnosis and proposals, and keep a person on every change that raises spend. Creative decisions need their own loop.
What should be in place before any agent touches the account?
An agent optimises whatever signal you give it. If the signal is wrong, it optimises the wrong thing faster. Check four foundations:
Conversion tracking you trust. The pixel fires on every key page, and server events flow through Meta's Conversions API alongside it. Compare Meta's reported purchases or leads with your store or CRM for the last month. If they disagree badly, fix that first.
One optimisation event per campaign that reflects real value: a purchase or a qualified lead, not a page view.
A simple structure. Fewer, larger ad sets give Meta and any agent more data per decision. Many tiny ad sets starve each one.
A naming convention that encodes campaign goal, audience and creative concept, so the agent can explain results in your terms.
What does "optimise" mean for your account?
Write it down in one line before you configure anything. For example: "Keep cost per purchase under our target while spending the full monthly budget, with ROAS above our floor." That gives the agent:
A goal metric: cost per purchase, cost per qualified lead or ROAS.
Guard metrics it must not break: daily spend pace, frequency on cold audiences, minimum spend on prospecting.
A decision window: how many days of settled data it may use, given that recent conversions keep arriving for days.
A minimum sample: how much spend or how many results an ad set needs before the agent may judge it.
Without this, "optimise" becomes whatever the model infers, and two runs on the same data can recommend opposite things.
Which parts should Meta automate and which should an agent handle?
Meta already automates a lot, and an agent should not fight it:
Budget across ad sets. With Advantage campaign budget, Meta's system looks for the best-performing opportunities across a campaign's ad sets and shifts the budget to them in real time (Meta for Developers). If you use it, the agent's job is the campaign budget and the decisions around it, not micro-moves between ad sets.
Simple if-then rules. Meta's automated rules can pause, unpause, change budgets and bids, rotate ads, notify people, and rebalance budget across ad sets on a schedule (execution types). Use them for stable, obvious logic.
The agent's job: reading across campaigns, explaining what changed and why, spotting problems rules cannot express (tracking gaps, pacing risks, fatigue across a whole concept), and proposing changes with reasons a person can approve in seconds.
Why can't budget automation fix weak creative?
Because budget can only flow to the ads you already have. Meta's retrieval system picks a few thousand candidate ads from tens of millions for each request, and Meta has said it is built for growing creative volume and a more varied set of candidates (Meta Engineering). An agent can move budget toward the best campaign you have, but if every ad in it says the same thing, there is nothing better to move toward. Plan two loops from day one: a campaign loop (budgets, pacing, pausing) and a creative loop (which angles are tiring, what to brief next, which claims you can prove). Nagent gives the second loop to its creative strategy lead, HOOK, so the people approving budget changes also see why an ad is losing. For the wider picture, see can AI plan, launch and optimise Google and Meta ads.
How do I connect the agent safely?
Use a system user, not a person's login, so access survives staff changes and can be revoked in one place.
Grant read access first. In Meta's terms that is
ads_read; addads_managementonly when you move to proposals (permissions).Set an account spending limit in Meta so no bug or bad decision can exceed it.
Limit scope to the campaigns the agent may touch.
Keep the token out of chat. Credentials belong in the tool's integration settings, never pasted into a conversation.
What guardrails should the agent run under?
Set these as settings or code, not as instructions in a prompt:
A hard spend cap, daily and monthly, checked before any change that commits money.
Approval on every raise. Lowering spend can become automatic later; raising it should always need a named person.
A change cap: the largest budget change allowed per campaign per day. Large jumps can reset delivery learning.
Paused by default for anything the agent creates.
A re-check on approval: if the campaign changed between the proposal and the approval, the change should be refused, not applied to a different situation.
A full log of each proposal, its data, its approver and its result.
How should the first 30 days run?
Days 1 to 7, observe. The agent reads the account daily and reports. You check that its numbers match Ads Manager and that its explanations make sense.
Days 8 to 21, propose. It proposes changes with reasons. People approve or reject each one and note why. Track the agreement rate.
Days 22 to 30, narrow automation. If agreement is high for one kind of change, such as pausing a campaign that has clearly missed its target, consider letting that kind run with less friction. Keep raises and creations on approval.
After 30 days, review the log: which proposals were right, which were rejected, and what rule would have caught the bad ones.
Which tools can optimise Meta campaigns?
Per their own sites, checked 24 October 2026:
Omneky runs scheduled automations from a chat instruction, such as pausing ads whose cost per acquisition is above a set figure over seven days, with a dry run first, and its agent can launch ads to Meta (Omneky Automations). Plans from $29 a month (pricing).
Motion offers routines for daily and weekly reports and alerts as part of its creative analytics, from $999 a month for up to $50k of monthly ad spend (Motion pricing).
AdCreative.ai connects ad accounts for creative insights and fatigue alerts (AdCreative.ai).
Pencil launches and tracks ads across Meta and other platforms (Pencil).
Foreplay focuses on creative research and analytics rather than campaign changes (Foreplay pricing).
Where does Nagent fit?
We make Nagent. Two Marketing team leads share this work, and the split matters:
NIA, the performance marketing lead, runs the ad account. Her Meta ads work is in Beta. A workspace admin connects the Meta account and sets a hard spend cap, daily and monthly, on the integration page; the token never appears in chat. In her team room you can ask what the account holds, how campaigns did over the last 1 to 30 days against the days before, or ask her to pause a campaign, lower or raise its daily budget, resume it, or create an empty campaign that arrives paused. Pausing and lowering each wait for an approval. Resuming, raising and creating need a named approver's click every time, and the change must fit what is left of the cap today and for the rest of the month; with no cap set, she refuses. On approval she syncs the campaign again and refuses if it changed.
HOOK, the creative strategy lead (Beta), handles the creative side. It reads the same Meta results nightly, ranks ads against your own account, flags fatigue and writes the next angles. It makes read calls only.
Limits to know: NIA works at campaign level on Meta. She changes campaign budgets, not ad set budgets, and a campaign whose budget sits on its ad sets can be paused but not resumed or re-budgeted by her. Ad sets, audiences, ads and creative stay in Ads Manager. After you approve one of her actions, the same kind of action you fire yourself runs without asking again for an hour, except resuming, raising or creating on Meta, which always wait. She does not optimise on her own on Meta today; she proposes and a person decides. Work is metered in credits; see pricing.
Frequently Asked Questions
How do I set up NIA to optimise Google Ads and Meta ads campaigns?
Connect each account, set the Meta hard spend cap, and ask NIA in her team room. On Google Ads she has 13 live actions behind fixed pre-checks; on Meta (Beta), five campaign-level actions. Every live change waits for approval.
Does AI creative testing automatically pause losing variants and move spend?
Some tools pause on rules. Nagent's HOOK flags losers and writes a refresh plan; a person pauses the ad, and budget changes go through NIA's approvals.
Can an AI agent manage Meta ads without any human approval?
Technically yes, with the right permissions. We advise against it for raises and new campaigns, because spend cannot be taken back.
How long before an AI agent improves Meta results?
Plan for a month of observing and proposing before judging. Gains come from faster, consistent decisions, not from the first day.
See the setup on a test account: book a demo.