Paid Media
Agentic Marketing Platform Permissions to Check First
The account access, spend limits and approval controls to confirm before an agentic platform touches your Meta or Google ad accounts.

An agentic marketing platform needs read access to performance and conversion data first, and write access only for the specific campaign actions you approve. It should never hold billing control, user administration or account ownership. Confirm the platform role, the API scope, a spend cap enforced outside the ad platform, an approval gate for launches and budget rises, and a tested way to revoke access.
Agentic marketing platform permissions should start at read-only and widen one action at a time. An AI agent needs to see campaign and conversion data to be useful, but it does not need billing control, user administration or account ownership. Before any platform touches your Meta or Google ad accounts, confirm four things: its role, its scope, its spend limits and its approval gates.
Key takeaways
- Start every agent on read-only access to one ad account, and move to write access only after its recommendations have proved sound.
- Keep billing, user management and account ownership with named humans. No agentic platform needs them to optimise campaigns.
- Spend limits must be enforced somewhere the agent cannot edit: the ad platform, the billing profile or the platform's policy layer. Ask which one.
- Require human approval for first launches, budget increases above a set size, tracking changes and anything destructive.
- Test the revoke path before the first live change. An access removal nobody has practised is documentation, not a control.
- Ask for a change log that records what the agent saw, what it proposed and which rule allowed it.
What is the quick answer on what permissions an agentic marketing platform needs?
The short answer is a ladder with four rungs. Each rung adds one capability, and you only climb when the previous rung has earned it.
- Observe: read-only access to campaign, spend and conversion data.
- Draft: the ability to create paused campaigns and propose edits, with a human pressing go.
- Act within caps: routine edits such as pauses, negatives and small budget moves, inside limits set outside the agent.
- Administer: billing, users and ownership. This rung stays human.
Vendors describe this the same way. BattleBridge, an ad-agent vendor, BattleBridge, September 2026, sets out an observer stage, a supervised operator stage and a bounded autonomous stage, and says advancement should depend on evidence rather than elapsed time. This guide turns that ladder into a checklist you can run against any platform, mapped to the real roles in Meta and Google Ads.
Why are agentic platform permissions harder than ordinary agency access?
Giving an agency access to an ad account is a familiar, human-paced act. An agent works at machine speed, can make many changes per cycle, and can be steered by text it reads, so a bad permission is costly faster. The risk is not only a rogue agent; it is a correct agent doing the wrong thing at scale because its limits were too wide.
The category is also moving into the ad platforms themselves. DesignRush, DesignRush, September 2026, reports that Meta is adding Partnership Ads to its Meta Ads MCP connector, so AI agents can create ads and manage creator permissions. That means the question is no longer only what a third-party tool can do. It is also what any connected agent can do with the access you have granted.
Round-ups of the best agentic platforms usually compare features, not permissions. Our guides to agentic marketing platforms for paid media teams and to what to test before buying cover the feature side. This page covers the part a feature table skips: what the agent is allowed to do on day one.
What are the layers of control in an ad account?
Permissions live in four layers, and a safe setup uses more than one. Teams often confirm the first and assume the rest.
| Layer | What it controls | Who sets it | Question to ask the vendor |
|---|---|---|---|
| Platform role | What the connected identity can do in Meta or Google Ads | You, in the ad account | Which exact role do you request, and can it be lower? |
| API scope | Read versus write calls the app may make | The app, approved by the platform | Does the integration request read-only scope until I enable writes? |
| Spend limits | How much money can move, and how fast | You, billing and the platform's policy layer | Where is the cap enforced, and can the agent edit it? |
| Approval gates | Which actions wait for a human | The platform's configuration | Which actions are gated by default, and can I add more? |
| Audit and revoke | Records and the off switch | You and the vendor | Can I export the log, and how do I cut access in one step? |
BattleBridge makes the point that advertising platforms provide roles and account permissions, but an agent needs another control layer that validates every proposed action before it runs. Its example rules include allowlisted account IDs, fixed budget-change thresholds, a cap on changes per cycle and automatic stop after repeated errors. Treat these as questions to put to any vendor, not as features every vendor has.
Step 1: How do I start with read-only access, and why?
Action: connect the platform to one ad account with read-only access, not to your whole business portfolio. In Google Ads that is the Read-only access level. In Meta it is a role or app permission limited to viewing and reporting.
Result: the agent can see performance, spend and conversions and can write recommendations, but cannot change anything.
Why: a read-only period shows how good the agent's judgment is before it can act on it. Supermetrics, Supermetrics, September 2026, recommends starting with one read-only account, then adding metric definitions and decision rules, testing against past problems, and only then introducing limited write access with human approval.
Read-only does not mean harmless. The data is sensitive, so the same questions about retention, sub-processors and who at the vendor can view it still apply. Include them in the first call.
Step 2: Which Meta and Google roles should I grant?
Action: pick the lowest role that supports the current rung of the ladder, and write it down next to the account ID.
Result: a single table you can show a colleague, an auditor or the vendor that says who holds what.
Why: role names are the most common place where access quietly widens, because Admin is the easy default.
On Google Ads, Meticulosity, Meticulosity, checked November 2026, lists five access levels: Admin, Standard, Read-only, Email-only and Billing. It describes Read-only as view campaigns and reports with no edits, and gives the practical rule of granting the lowest level that still lets someone do their job. Exchangefour, Exchangefour, checked November 2026, adds that for an agency Standard is usually enough, Admin stays with the account owner and billing stays on your own payments profile.
On Meta, Leadsie, Leadsie, 2026, notes that without a Business Portfolio you can be added under ad account roles as Admin, Advertiser or Analyst. Kitchn, Kitchn, checked November 2026, explains that Meta uses two broad business access levels, Partial access and Full control. Check the current labels in your own account, because platforms rename them.
| Ladder rung | Meta | Google Ads | Billing, users, ownership |
|---|---|---|---|
| Observe | Analyst-type, view-only access | Read-only | Human only |
| Draft | Partial access with campaign management on the ad account | Standard, with a policy layer blocking activation | Human only |
| Act within caps | Advertiser-type role, plus caps set outside the agent | Standard, plus a spend limit and gates | Human only |
| Administer | Full control or Admin | Admin or Billing | Not for an agent |
This mapping is our own framework, built from the role descriptions in the sources above. It is a starting point; confirm each role's exact abilities in the platform's help pages before you rely on it.
Step 3: What API scope should the platform request?
Action: ask the vendor which API permissions it requests from Meta and Google, and whether read and write are separate grants you control.
Result: you know whether turning on writes is a deliberate step or already baked into the connection.
Why: the role in the ad account and the scope of the app are two different gates, and an agent needs both before it can edit anything.
On Meta's Marketing API, adsap.ai, adsap.ai, 2026, explains that an app that reads reports asks for ads_read, while an app that creates or changes anything asks for ads_management. Climent Media, Climent Media, 2026, advises granting ads_read if you only read. If a vendor says it needs ads_management just to show you reports, that is a question worth asking.
Supermetrics shows the two-gate model in practice: it states that write access is configured separately for each client ad account, and that you also need edit permissions in the ad platform itself. The platform scope alone is therefore not enough, and neither is your role alone.
Step 4: How do I set spend limits an agent cannot raise?
Action: define three limits: a per-campaign daily budget ceiling, a per-change size limit, and a whole-account daily or monthly ceiling. Then confirm where each is enforced.
Result: a worst-case loss you can state as a number before the agent goes live.
Why: a limit the agent can edit is a suggestion. A limit in your billing profile or the platform's own policy layer is a control.
BattleBridge lists explicit controls: a maximum daily budget by campaign, a maximum percentage increase per change, a maximum total account spend per day, a cooldown between adjustments and human approval above a dollar amount. Its worked example lets an agent reduce a budget immediately, increase it by no more than 15% in a 24-hour period and request approval for anything larger. It also warns that the ad platform may not enforce every such rule natively, so the application controlling the agent must.
Account-level spend limits exist on the ad platforms, though the details change. Wevion, Wevion, 2026, covers how account limits, daily limits and budgets differ across Meta, Google and TikTok, and notes that Google can apply an account-level daily spending limit to new accounts and accounts with recent changes. A step-by-step guide from Arc Intermedia, Arc Intermedia, checked November 2026, describes an account-level monthly spend limit under a Google Ads manager account that pauses ads once the cap is reached. Check your own billing settings, because availability varies by account type.
Worked example: why a per-change cap is not enough
Take the BattleBridge example of a campaign budget at $200 and a 15% cap per 24 hours. One increase takes the budget to $230, which matches the figure BattleBridge uses in its change-log example. The per-change cap looks safe.
Now let the agent apply the maximum increase on three consecutive days. The arithmetic, which is ours: $200 becomes $230, then $264.50, then about $304.18. The budget has risen by more than half in three days, with every single change inside the rule.
The fix is a second limit: a cumulative ceiling per campaign, such as a fixed daily maximum, that no chain of small changes can pass. This is why the checklist asks for a per-change limit and an absolute ceiling, and why the agent should not be able to edit either one. Our post on agentic ad spend guardrails for Meta and Google Ads goes further on pacing rules.
Step 5: Which actions need a human approval gate?
Action: list every action type the agent can take and mark each as automatic, approval required or forbidden.
Result: a one-page action matrix that fixes the default for the agent's first month.
Why: the cost of a mistake differs by action. Drafting is reversible; spending is not.
| Action | First month | After evidence |
|---|---|---|
| Read performance and conversion data | Automatic | Automatic |
| Create a campaign in paused status | Approval required | Automatic |
| Activate a campaign for the first time | Approval required | Approval required |
| Pause an ad or keyword | Approval required | Automatic |
| Add negative keywords or exclusions | Approval required | Automatic |
| Raise a budget | Approval required | Automatic within the cap |
| Change targeting or geography | Approval required | Approval required |
| Change conversion tracking or attribution settings | Forbidden | Approval required |
| Delete campaigns or assets | Forbidden | Forbidden |
| Change billing, users or ownership | Forbidden | Forbidden |
The table is our own recommended default, drawn from the controls BattleBridge and Supermetrics describe. Supermetrics offers four approval modes, recommends selecting its All live changes mode for a first live test, creates new campaigns paused and records changes in Campaign history. Ask any platform you evaluate for the equivalent, in writing.
Tracking changes sit in the forbidden column because they quietly rewrite what the agent optimises toward. If you want to check what an agent is rewarded for, read our guide on why Meta ROAS can look good while revenue disagrees.
Step 6: What should the audit log and alerts show?
Action: require a change log you can export, and set alerts for spend spikes, failed actions and changes outside working hours.
Result: every change can be traced to an observation, a proposal and a rule.
Why: the platform's own history shows what changed, not why. BattleBridge suggests recording what the agent observed, the action it proposed, the reasoning, the authorising rule, the exact API request, the platform's response, the performance impact and whether a human approved or overrode it.
The ad platforms' native change history is your independent second record. Compare the two on a schedule. If the agent's log says it made four changes and the platform shows nine, you have found a gap in the vendor's reporting before it becomes a billing question.
For what to watch each day, see ad account alerts on Meta and Google, and for a review method, how to audit AI-automated ad campaigns.
Step 7: How do I plan revocation before the first live change?
Action: write and rehearse a shutdown sequence: disable the agent's identity, revoke the OAuth grant or API token, stop scheduled runs, review pending jobs, and restore the last approved campaign state.
Result: a one-page procedure with a named owner and a tested time to complete.
Why: the moment you need to cut access is the moment you have the least time to work out how.
BattleBridge's list of revocation steps includes rotating related credentials where necessary and confirming that no secondary automation remains active. Add one more: confirm the access has vanished in the ad platform's own user list, not only in the vendor's dashboard. Platform users and partner connections can outlive a cancelled subscription.

The pre-connection checklist
Run this in order before you connect any agentic platform to a live ad account. Each line should end with a name or a screenshot.
- Confirm who owns the ad account, the business portfolio and the payment method. It should be you, not the vendor.
- Choose one ad account for the pilot and write down its ID.
- Request read-only access and record the exact role granted on each platform.
- Ask the vendor to list every API scope it requests and why.
- Confirm the vendor will not ask for your password, a recovery code or a shared login.
- Set a per-campaign daily ceiling, a per-change size limit and an account-level ceiling, and note where each is enforced.
- Fill in the action matrix from Step 5 and get a second person to sign it.
- Confirm that new campaigns are created paused and that first activation needs a human.
- Check that an exportable change log exists and compare it with the platform's change history after the first week.
- Rehearse revocation once, then confirm the access is gone in the platform's own settings.
- Set a review date for moving up one rung, with the evidence you will judge it on.
How do I limit what an AI ad agent can change in my account?
You limit an AI ad agent with five stacked controls: a low platform role, a narrow API scope, an account allowlist, spend ceilings and approval gates. No single control is enough on its own. The role stops the agent touching billing and users, the scope stops it making calls you did not enable, and the ceilings cap the damage if it behaves badly.
BattleBridge states the principle directly: an agent should have enough access to improve performance, not enough to own your account. Restated, the platform role defines the outer boundary and the policy layer defines the inner one. Ask vendors to show both.
What mistakes do teams make with agentic platform permissions?
- Granting Admin so nothing breaks. It removes the main boundary between the agent and billing, users and ownership.
- Connecting through a personal login. You cannot isolate or revoke it without locking out the person.
- Connecting the whole business portfolio. One pilot account is enough to judge an agent.
- Trusting a per-change cap alone. As the worked example shows, small steps compound.
- Letting the agent edit conversion tracking. It then controls the score it is graded on.
- Skipping the revoke drill. The first real test of the off switch should not be an incident.
- Ignoring the learning phase. Frequent edits can disrupt delivery; see what resets the Meta learning phase before letting an agent edit often.
How does Pulse help you stay read-first?
Pulse by Nurdd is a web dashboard that connects to Meta Ads and Google Ads through read-only OAuth. It explains campaign performance in plain language, diagnoses what changed and recommends specific actions, without requiring anyone to build reports. Because the connection is read-only, Pulse sits on the Observe rung of this ladder: it shows you what an agent would need to know before anything is allowed to act.
Pulse is paid with no trial, and it is in private access. Read more on the Pulse page, or see how it fits Nurdd's other products in How Nurdd works.
Book a demo at https://www.nurdd.club/demo to see Pulse connected to your own Meta and Google Ads data, with the diagnosis and recommended actions it would give your team, and nothing changed in your account.
Which permission tier fits which job?
Use this decision table to pick a starting rung for the task you want an agent to do. The mapping is our framework; the rungs follow the staged rollout BattleBridge and Supermetrics describe.
| Job | Start at | Move up when | Never delegate |
|---|---|---|---|
| Explain why cost per result rose | Observe | Recommendations proved right on past problems | Billing, users |
| Weekly performance reporting | Observe | Stays at Observe | Any write access |
| Build new campaign drafts | Draft | Drafts pass review repeatedly | First activation |
| Trim wasted spend with pauses and negatives | Act within caps | Pause decisions match your own reviews | Deleting assets |
| Reallocate budget across campaigns | Act within caps, small cap | Cumulative ceiling tested | Raising the account ceiling |
| Manage creator partnership permissions on Meta | Draft, with approval | Approval log shows no surprises | Granting access to new partners unreviewed |
If you also use creator content in ads, the creator permission row matters more now, given the Meta MCP connector change DesignRush reports. For tools that cover the whole paid media stack, see AI agents for paid media.
How should an agency or in-house team split responsibilities?
Decide who is the human owner of each control. The most common failure is two teams each assuming the other holds the off switch. Write the names on the action matrix.
- Account owner: billing, users, ownership and the account-level ceiling.
- Media lead: the action matrix, per-campaign ceilings and weekly review.
- Analytics owner: conversion tracking and attribution settings, which the agent may not edit.
- Security or operations: credential storage, revoke drill and vendor questionnaire.
Teams weighing a managed option over software can read our comparison of a paid social agency versus software, or see the paid social advertising agency service.
How do I judge results once the agent is on a live account?
Judge results against your own break-even, not the agent's own dashboard. Use the Google + Meta ROAS calculator to check each channel against break-even, and the marketing budget planner to sanity-check the split. For category context, the paid media benchmarks post publishes Triple Whale medians by channel, and the benchmarks hub shows each figure with its source and period.
Before pausing or letting a campaign run on an agent's advice, our pause-versus-optimise framework gives a second opinion, and the 30-minute Meta account audit is a quick way to check the starting state. For the wider category, see AI agents for marketing.
How we researched this
We reviewed vendor guides on AI ad agent access, platform-integration setup documentation from a marketing data vendor, trade news on Meta's agent connector, and practitioner guides to Meta and Google Ads roles and spend limits. We read the BattleBridge and Supermetrics guides in full and used search results for the others. Pages were checked on the publication date of this post. We did not test any agentic platform hands-on, and roles and labels change often, so confirm them in each platform's own help pages.
Conclusion and next step
Agentic marketing platform permissions come down to one rule: grant the least access that supports the current job, and enforce limits where the agent cannot change them. Start read-only, map each rung to a platform role, cap spend with both a per-change and a cumulative ceiling, gate launches and budget rises, and rehearse the off switch.
If you want to see what read-first looks like on your own accounts, book a Pulse demo. You will see your Meta and Google Ads performance explained in plain language, what changed and what Pulse recommends, with read-only access.
Frequently asked questions
What permissions does an agentic marketing platform need?
It needs read access to campaign, spend and conversion data first. Write access should follow only for specific actions such as creating paused campaigns, adding negatives or small budget changes. It does not need billing control, user management or account ownership to optimise campaigns.
How do I limit what an AI ad agent can change in my account?
Stack five controls: a low platform role, a narrow API scope, an allowlist of account IDs, spend ceilings enforced outside the agent, and approval gates for launches and large changes. Then rehearse revoking access so you can cut it quickly.
Does an AI ad agent need admin access?
No. BattleBridge says an agent should work through scoped access while a human keeps control of billing, users, security settings and ownership. On Google Ads, Exchangefour suggests Standard is usually enough for an agency and Admin stays with the owner.
What is the difference between ads_read and ads_management on Meta?
According to adsap.ai, an app that reads reports asks for ads_read, while an app that creates or changes anything asks for ads_management. If a reporting-only tool asks for ads_management, ask why before approving it.
Should an AI agent be allowed to launch campaigns?
Not at first. Let it create campaigns in paused status and have a human approve the first activation. Supermetrics creates new campaigns paused and offers approval modes, and recommends its All live changes mode for a first live test. Keep first activation gated even later, since launching means immediate spend.
How do I revoke an agent's access quickly?
Disable the agent's identity, revoke its OAuth grant or API token, stop scheduled runs, review pending jobs and restore the last approved campaign state. Then confirm in the ad platform's own user list that access is gone.
Is read-only access really safe?
It removes the risk of unwanted changes, but the data remains sensitive. Ask the vendor about retention, who can view the data and which sub-processors touch it. Read-only is the right first rung, not a replacement for those questions.
Where can I see read-first ad intelligence on my own accounts?
Pulse by Nurdd connects to Meta Ads and Google Ads through read-only OAuth and explains performance in plain language. You can see it on your own data by booking a walkthrough at https://www.nurdd.club/demo.