Skip to content
AI stack, guide

AI agents in paid user acquisition: delegate, gate or keep

Last verified

Hand AI agents the work that reads data and drafts output: daily reports, anomaly flags, creative briefs, naming checks. Gate anything that spends or changes delivery, such as budgets, bids and launches, behind human approval and a spend cap. Keep strategy, targets and kill decisions with the media buyer.

What to delegate to an agent in paid UA
What to delegate to an agent in paid UAThree lanes: the agent drafts, checks and reports; the person approves budget and publishing at a gate; the ad platform executes. Every step that spends money passes an approval gate first.AgentproposesPersonapprovesAd platformexecutesDraftads, briefsCheckpolicy, capsApprovepublishingLaunchspendsReportread-onlyApprovebudgetBudgetchangespendsNEXT ROUNDApproval gate before any step that spends moneyWhat to delegate to an agent in paid UAThree lanes: the agent drafts, checks and reports; the person approves budget and publishing at a gate; the ad platform executes. Every step that spends money passes an approval gate first.AgentproposesPersonapprovesAd platformexecutesDraftads, briefsCheckpolicy, capsApprovepublishingLaunchspendsReportread-onlyApprovebudgetBudgetchangespendsTHEN THE NEXT ROUNDGate before any spend
Illustrative flow, simplified.
The agent drafts, checks and reports. The person holds the gate in front of every step that spends money, and only then does the ad platform execute.

Delegate, gate or keep: the task table

An AI agent in paid UA is a program that uses a language model to read ad data, pick a next step and, if you give it access, act in an ad account through tools. Sort each task by what the agent can touch. Delegate means the agent reads data and drafts output. Gate means the agent may propose a change, but a person approves it before anything spends or changes delivery. Keep means the decision stays with the media buyer.

Delegate, gate or keep by task
TaskTierReasonControl
Daily spend, CAC and ROAS report from exportsDelegateReads data and drafts a summaryTotals reconcile to the platform and backend before anyone reads it
Anomaly flags: CPM spike, conversion rate drop by source, placement or creativeDelegateFlags only, changes nothingFlag goes to a channel, a person decides the action
Creative briefs from past test resultsDelegateDrafts text from your own dataBuyer edits before the brief reaches designers
Naming and UTM convention checksDelegateDeterministic comparison against a patternViolations listed, fixes made by a person or a gated script
Competitor ad research from public ad librariesDelegatePublic data only, no account accessSource links kept with each finding
Kill, scale and pacing recommendations from written rulesGateDrafting is safe, applying the rules changes deliveryAgent outputs a list, buyer approves each line
Budget changesGateSpends money directlyPer-change cap, daily account ceiling, named approver
Bid, target CPA or ROAS changesGateShifts delivery and can reset learningPreview of old and new value, approval, weekly edit limit per ad set
Launching campaigns, ad sets or adsGateCreates spend and brand exposurePreview before write, launch in paused state, human turns it on
Automated stop and restart rulesGateActs without a person in the loopRule set signed off once, spend cap, log of each trigger
CAC target, payback ceiling and margin assumptionsKeepDepends on finance and business contextBuyer or head of UA owns the number, agent reads it
Strategy, channel mix and kill calls on ambiguous testsKeepNeeds judgement the data cannot supplyAgent may summarise evidence, the decision is recorded under a person's name

No measured results of agents in paid acquisition were found in the research behind this page. Treat vendor claims about autonomous media buying as unverified until you see a test on your own account.

Approval gates and spend caps

Enforce the gate in the access layer, not in the prompt. A prompt that says "ask before spending" can be ignored. A read-only token cannot write. Give the reporting agent read-only credentials and issue a separate, scoped credential for any agent that writes.

  • Split credentials by tier: read-only for delegate tasks, write access only for gated tasks.
  • Require a preview of the old value and the new value before any write, with a named approver.
  • Set a per-change cap and a daily account ceiling in the tool layer, outside the model's control.
  • Limit edits per ad set per week so frequent tweaks do not reset learning.
  • Create new campaigns in a paused state so a person turns on the spend.

Illustrative example: if the per-change cap is a percentage of the current daily budget, the largest proposal an agent may make is the current budget multiplied by one plus that percentage. A proposal above that figure is rejected without review. The account ceiling works the same way: any set of approved changes that pushes total daily budget above the ceiling is blocked.

Logging, naming and rollback

Each agent action needs a record that lets a person answer three questions in a minute: what changed, who approved it and how to undo it. Without that record, a bad week turns into an investigation.

  • Log the timestamp, agent name, object id, old value, new value, approver and the reason the agent gave.
  • Prefix every object an agent creates, for example agt_ plus date plus a ticket id, so agent work is easy to filter.
  • Save a snapshot export of the affected campaigns before each write.
  • Keep a revert script or a documented manual step for each write type, and test it on a small campaign.
  • Store logs outside the ad platform, in a sheet or database the agent cannot edit.

Failure modes and fixes

  • Bad input data: an agent that reads unreconciled platform, MMP and payment numbers will recommend the wrong action with confidence. Reconcile attribution first and have the agent state which source each figure came from.
  • Small samples: rules without a minimum volume kill ads too early. Put event thresholds into the rule text and make the agent report below-threshold items as undecided.
  • Permission creep: MCP servers run with the permissions of the user who installs them. Check the scopes each server requests and drop any you do not need.
  • Unreviewed third-party skills: the safety scans in the research read file lists and pattern matches only and executed nothing. Review the code yourself before granting write access, and treat a flagged script layer as a reason to read it.
  • Data leaving your machine: some audit skills fetch landing pages or send image briefs to outside model providers. Check each tool's data flow before you feed it account data.
  • Persona prompts: a prompt such as "act as an expert media buyer" has no inputs, thresholds or procedure. Replace it with written rules the agent applies to an export.

Audit your setup

The ai-marketing-system-audit skill maps a marketing system as a loop of acquire, convert, measure, automate, intelligence and decide, then ranks the gaps. Run it before you widen any agent's access. If the measure step is weak, agents will automate on numbers you cannot trust.

  1. 1.List what agents touch

    Write down each agent, the tools it connects to and whether its access is read-only or write.

  2. 2.Run the audit skill

    Give the skill your current setup description and exports, and ask for the gaps ranked by what to build first.

    Run ai-marketing-system-audit on this setup. Inputs: agent list, access level per agent, latest campaign export, current reporting process. Output: ranked gaps and the first build.
  3. 3.Fix measurement before write access

    If attribution or payment data does not reconcile, repair that first. Agents inherit every error in the data.

  4. 4.Assign tiers

    Place each task from your own list into delegate, gate or keep, and set the cap, approver and log location for each gated task.

  5. 5.Pilot on one account

    Start with one small campaign, review the log after a week, test a rollback, then widen scope.

Questions on agents in paid UA

Questions

What can an AI agent safely do in an ad account?
Read-only work is the safest: daily reports, anomaly flags, creative briefs, naming checks and research from public ad libraries. These tasks draft output and change nothing in delivery. A person still reads the output before acting on it.
Should an AI agent change budgets on its own?
Put budget changes behind approval and a spend cap. An agent can propose a change with the old and new value, but a named person approves it. A per-change limit and a daily account ceiling in the tool layer stop a bad proposal from spending.
Are read-only ad platform MCP servers safe to install?
Read-only reduces risk but does not remove it. MCP servers run with the permissions of the user who installs them, so review each server's code and scopes. Check where tokens are stored and whether any data leaves your machine.
Who decides what to kill in a campaign?
The media buyer keeps kill decisions, especially on ambiguous tests. An agent can apply written kill and scale rules to an export and list recommendations, such as with the ua-kill-scale-rules skill. A person approves the list and owns the outcome.

Sources

Each source shows the date it was last checked.

  1. GitHub github.com/AdsMCP/tiktok-ads-mcp-server
  2. GitHub github.com/googleads/google-ads-mcp
  3. GitHub github.com/kLOsk/adloop
  4. GitHub github.com/RoyAzran/mcp-ads
  5. GitHub AgriciDaniel/claude-ads: LICENSE
  6. GitHub ua5er/user-acquisition-skills: LICENSE
  7. Model Context Protocol modelcontextprotocol.io/docs/getting-started/intro
  8. GitHub AgriciDaniel/claude-ads: README.md
  9. GitHub AgriciDaniel/claude-ads
  10. GitHub ua5er/user-acquisition-skills

Want your own AI and acquisition setup checked for gaps and risks? Start with a growth audit.

See the growth audit