guide
AI media-buying agents: L1, L2, L3 autonomy explained (2026)
The full framework - what belongs at each level, how action types graduate, and why full autonomy fails compliance review.
ReadWhere the hours go
Four things that cap how much spend a team can hold.
Creative production is the bottleneck, not the budget
The account can absorb more spend but the team cannot ship enough variants to feed it. Fatigue sets in on the winners, and the next batch is three days out because someone has to write, brief, and assemble it by hand.
Budget shifts happen on Monday, not on Tuesday night
A campaign starts bleeding at 2am. Nobody looks until the morning stand-up. By then you have paid for a full day of a spend pattern anyone would have killed if they had been watching.
Reporting eats the buyer's day
Pulling numbers out of each platform, reconciling them against the affiliate postbacks, and rebuilding the same sheet every morning is hours a week of a senior person doing assembly work.
Full-autonomy tools are unauditable
Tools that promise to run the account for you make decisions you cannot reconstruct afterwards. When spend goes somewhere strange, there is no trace of why - and in a regulated vertical that is not a risk you can carry.
What we build
Six pieces, every one of them auditable.
Creative generation with a brand and compliance gate
Variant generation from your winning angles, passed through a checklist gate before it can reach a human queue - banned claims, required disclaimers, regional restrictions. Output is a draft for review, never a direct publish.
Anomaly detection on live spend
Statistical monitoring on cost per action, conversion rate, and spend velocity per campaign, against that campaign's own baseline rather than a global threshold. It fires an alert with the numbers attached, not a vague warning.
Budget reallocation proposals
A daily proposal that says which campaigns should move budget and why, with the evidence. In L2 configuration a human approves it in one click; in L3 it executes inside pre-agreed bounds and logs every move.
Cross-platform reporting pipeline
One pipeline pulling from each ad platform and your affiliate postbacks, reconciled into a single table. The morning sheet builds itself before anyone logs in.
Landing-page and angle testing support
Generated variants tied back to the performance data, so the next round of creative is briefed from what actually converted rather than from what the team remembers converting.
Audit log on every agent action
Every proposal, every approval, every automated execution written to a log with inputs and reasoning. This is what makes the system defensible in a compliance review and debuggable when a decision looks wrong.
Process
Autonomy is earned, not switched on.
01
Audit the account and the data
We look at how spend is currently managed, where the manual hours actually go, and whether the performance data is clean enough to drive automated decisions. Two to three weeks. If the tracking is broken, fixing that comes first and we will say so.
02
Build at L1, then earn L2
First deployment is advisory only - the agent proposes, humans decide, nothing executes. Once the proposals have been right often enough for the team to trust them, we move approved action types to one-click approval.
03
Bounded L3 where it has earned it
Only specific, low-blast-radius action types graduate to autonomous execution, inside hard bounds you set, with an audit log and a kill switch. Most accounts should keep budget decisions above a threshold at L2 permanently, and we will tell you which.
// FAQ
Questions performance teams ask first.
What do L1, L2 and L3 autonomy actually mean?
L1 is advisory - the agent surfaces a recommendation and a human does everything. L2 is approval-gated - the agent prepares the action and a human approves it in one click. L3 is bounded autonomous - the agent executes inside limits you set, logging every action. Almost every account should start at L1 and graduate action types individually.
Will this replace our media buyers?
No, and we would not pitch it that way. It removes the assembly work - report building, variant production, overnight monitoring - so the buyer spends their day on angles, offers, and channel strategy. The judgement calls stay with the person who has the account context.
Do you touch our ad accounts directly?
Only at L2 and L3, only for action types you have explicitly enabled, and always through your own account credentials with a kill switch. At L1 the system has read access and nothing else.
How is compliance handled for regulated verticals?
Generated creative passes a rules gate before it reaches a human queue - required disclaimers, banned claims, regional restrictions encoded as checks. The gate blocks; it does not warn. The specific rule set is built from your licence conditions during the audit phase.
What does this cost?
Scoped after the audit phase, which is priced separately. Cost depends on how many platforms need integrating, the state of your tracking, and which autonomy levels you want. We quote precisely after we have looked at the account, not before.
Start with the audit, not the agent.
Thirty minutes on the account. We map where the manual hours go and whether your tracking is clean enough to automate against. If it is not, that is the first project and we will say so.
Book a 30-min call