
By Siddharth Borderwala
Published
11 min read
AI governance for marketing: permissions, approvals and brand checks for agents
What marketing agents may read, draft, change and publish in your CMS, ad accounts, email platform and CRM, which assets need whose approval, what to record and disclose, and where brand checks fit. With a proposed one-page policy for one workflow, and a run that shows a stale brand edit being refused.
TL;DR
Govern marketing agents by what they can do in each system: let them read widely and draft freely, change things carefully and publish rarely. Match approvals to the risk of the asset, give agents their own accounts so every change is logged, follow disclosure, data and rights rules, and check each asset against the brand before a person approves it.
What it covers
Justworks’ ad for a Senior Manager, Growth Marketing Automation & AI lists the job in one sentence: “Establish Marketing-specific governance standards for agent permissions, human approvals, brand consistency, customer data usage, monitoring, and quality, aligned with broader GTM technical and security standards.” That sentence is a fair outline of marketing AI governance.
Company-wide AI policy usually covers which tools are approved and how data is protected, and content governance covers who may publish what. The marketing layer is narrower and more practical: which marketing systems an agent can touch, what it may do there, who approves which assets, what gets recorded, and what has to be disclosed. Brand governance, the rules and owners behind the brand itself, is one part of it; brand governance for AI agents covers that part in detail.
The people doing it are a mix. Databricks wants a Principal Brand Content lead to own “the brand context, target prompts, and writing guidelines that guide agents” and to review “drafts from agents”. Code and Theory, an agency, is hiring an architect to build “approval chains, and role-based governance frameworks that enforce brand consistency and compliance” for clients.
Permissions: read, draft, change, publish
Four verbs cover almost everything an agent does in a marketing system. It can read what’s there, draft something new, change something that exists, or publish, send or spend. Risk rises with each one, so a sensible default lets agents read widely and draft freely, change carefully and publish rarely.
| System | An agent may, by default | Needs a person |
|---|---|---|
| Website and CMS | Read pages; create drafts and staging pages | Publishing, and editing live pages |
| Email platform | Read templates and past results; build draft campaigns; send tests to an internal list | Scheduling or sending to customers, and changing segments |
| Ad accounts | Read performance; create new ads paused | Turning ads on, and changing budgets or bids |
| Social scheduler | Draft posts in the queue | Approving and scheduling them |
| CRM | Read the fields a task needs | Writing to records, and exporting contacts |
| Brand guidelines and assets | Fetch guidance and files for a task; propose a change | Applying a change to the brand |
This table is a starting point, not a rule. A team that has checked a hundred templated social posts may let an agent schedule the hundred-and-first. A regulated business may not let it draft a claim at all.
Put each limit in the tool, not only in the prompt. An instruction such as “don’t publish” is guidance an agent can misread; an account that can’t publish is a control. In practice:
- Give each agent its own account or API key with the least access its task needs, rather than a person’s login.
- Remove the tools it shouldn’t have. In Claude Code, a deny rule that names a tool “removes the tool from Claude’s context entirely, so Claude never sees it.” In the session we used to research this article, our project’s deny rules for another service’s write tools were in force: that service’s read tools were in the agent’s tool list and the denied ones weren’t.
- Know what each setting covers. Codex runs commands in a sandbox with network access off by default, but OpenAI’s security guidance says those command-network controls “do not, by themselves, disable hosted web search, apps, or MCP.” Connector and MCP tools have their own approval rules, based on what each tool says about its side effects.
- Don’t give an agent a shell and a key that can do what you’ve denied. A blocked tool is no barrier if the agent can make the same API call from the command line.
- Keep publishing, sending and spending behind a person’s account until a workflow has a record you trust.
Approvals by risk
Approving every asset by hand cancels out the reason for using agents. Approving none is how an unreleased price ends up in a public ad. Tie the approval to what could go wrong:
| Risk | Examples | Approval |
|---|---|---|
| Internal only | Research summaries, internal drafts, briefing notes | None; spot-check a sample |
| Public, from a template | Social posts and emails built from approved templates and approved copy | A brand check, then one owner |
| Claims, prices, spend | Ads with an offer, comparative claims, pricing, anything with a budget | The owner plus product marketing or whoever owns the claims |
| Regulated or sensitive | Health, financial or legal claims, testimonials, real people, crisis replies | Legal or compliance, every time |
Name a person for each approval, not a team. The approver should see the asset, the sources behind its claims and the check results together, so approving takes minutes rather than a meeting.
Brand checks are one control
A brand check answers one question: does this asset follow the brand? It doesn’t decide whether a claim is true, whether the offer is approved or whether the asset should go out at all. Run it before the person approving sees the asset, so they spend their time on the decisions only they can make.
In Rainbrand, which gives AI agents a brand’s guidance and files, a Review checks a captured image’s colour, typography, logo and copy against the brand and returns a verdict and a reason for each. It reports; it doesn’t edit or publish anything, and it records which brand version it checked against. What is brand compliance? covers what to check and how. The creative QA checklist covers the checks beyond the brand, from claims to rights and disclosure, and brand compliance monitoring covers checking assets once they’re live.
Keep a record you can audit
When something goes wrong, you need to know which agent did it, with what, and who approved it. For every asset that goes out, keep:
- The task, the workflow and the person who started it.
- The sources and inputs the agent used, including the brand context: for Rainbrand, the Brief.
- The checks it passed or failed, and how each failure was resolved.
- Who approved it and when it was published, sent or launched.
- Any change the agent made in a live system.
Many marketing platforms already keep part of this. Google Ads’ change history, for example, shows “the email address of the person who made the change” for changes made in its interface, going back two years. It can only tell agent changes apart if the agent has its own login, which is one more reason not to run agents on a person’s account.
Disclosure
Disclosure rules depend on the platform, the place and the kind of asset. Three that affect marketing teams now:
| Rule | What it requires |
|---|---|
| YouTube’s synthetic content policy | Creators must “disclose AI-generated or altered content that seems realistic”, such as a real person appearing to say something they didn’t |
| EU AI Act, Article 50(4) | Since August 2, 2026, deployers must disclose deepfakes, and AI-generated text published to inform the public on matters of public interest unless a person reviewed it and someone holds editorial responsibility |
| FTC rule on reviews and testimonials | Bans fake reviews, including reviews that claim to be “by someone who does not exist, such as AI-generated fake reviews” |
Sources: YouTube Help, the AI Act Service Desk (with the application date in Article 113) and the FTC, checked November 15, 2026. Ad platforms have their own rules, especially for political and social-issue ads, so check each one you buy on. This is a summary, not legal advice; your legal team decides what applies to you.
Data and rights
Justworks’ ad puts “customer data usage” in the same sentence as permissions, and it belongs there. Give an agent the customer fields a task needs, not the whole CRM, and only through tools your company has a data agreement with. Check each tool’s training terms for the plan you’re on: Anthropic, for example, says it won’t use inputs or outputs from its commercial products for training by default.
Rights need the same care. Don’t generate images of real people, customers or employees without their consent, and don’t present a generated product image as a photograph of the real thing. Use licensed photos and fonts as references only where the licence allows it. And know what you own: the US Copyright Office’s January 2025 report concluded that prompts alone don’t make someone the author of what an AI generates, so a fully generated image may not be protected the way a designed one is.
A proposed policy, and the control we tested
Write a one-page policy for each recurring agent workflow, such as weekly social posts, webinar emails or ad variants. Here is a proposed one for editing a brand to update its logo files from its official brand page. Changing a brand is the riskiest thing an agent can do in Rainbrand, because every later Brief, the guidance and files Rainbrand prepares for one task, uses the result.
| Proposed policy field | This workflow |
|---|---|
| Workflow | Update a brand’s logo files from its official brand page |
| Owner | The brand owner, by name |
| Agent and its access | A drafting agent whose client denies the apply tool, with no Rainbrand key in its shell |
| Reads | The brand page, the official logo pack, the current brand version |
| Drafts in | The brand’s library (uploaded files) and edits awaiting approval |
| May change | Nothing active: an edit changes the brand only when applied |
| May publish | No. Applying an edit is the publish step, and only the owner does it |
| Owner applies from | Their own connected-agent session, or an API client they control |
| Approver sees | The proposed changes, file by file, and the source each one rests on |
| Check after approval | A fresh Brief, to confirm agents now get the new file |
| Records | Each edit’s ID, the version it started from, its status and the result |
| Stops and asks when | The edit asks a question, or the source doesn’t support a change |
Rainbrand API keys aren’t limited to drafting, so the drafting limit in that table has to come from the agent’s client and from who holds a key.
What our run tested: stale-edit protection
On November 16, 2026, we ran the drafting half of this workflow in our own Rainbrand account, with Attio’s brand. Attio, a CRM company, offers a logo pack on its brand page “in both light and dark variants”, and the copy of its brand in our account had no primary logo set. The run went like this:
| Step | What the system recorded |
|---|---|
| Prepare edit A: black lockup as the primary logo | 47e90421, from version 24a548a2, awaiting approval |
| Prepare edit B: white lockup for dark backgrounds | 7df8c839, from 24a548a2; asked for the exact file |
| Answer with the file from the logo pack | 7df8c839 awaiting approval |
| Inspect A, then apply it | Applied; the brand is now version d8e585c0 |
| Try to apply B | Refused: “Brand Edit cannot continue from stale” |
Prepare B again from d8e585c0, inspect, apply | b00b87e7 applied; the brand is now 4c4a1f5a |
| Request a fresh Brief for a dark-background image | 3eaf9495 returns the white lockup |
The refusal is stale-edit protection. Edit B was prepared and inspected against a version of the brand that no longer existed once A was applied, so Rainbrand marked it stale and wouldn’t apply it. Nobody had to remember to check. The same thing protects two people, or two agents, preparing changes to one brand at the same time.
It’s a version check, not an authority check. This run doesn’t implement or test the owner gate in the proposed policy. The same agent, holding a Rainbrand API key that could apply any edit, applied edit A and the re-prepared edit B. It followed a written rule to apply only what it had inspected and only what the official page supports, and nothing in the system enforced that rule. That’s acceptable for a test and wrong for a team.
To put the proposed policy in place, deny the apply tool in the drafting agent’s client, keep any key that can apply out of that agent’s environment, and have the owner apply from their own connected-agent session or through an API client they control. Then test the gate: have the drafting agent try to apply a valid edit, and record that it couldn’t.
The same fields work for any workflow. Start with one, run it for a few weeks with every change approved, then loosen the approvals the record shows are safe to loosen.
FAQ
Who should own AI governance in marketing?
Split it. IT or security sets the access boundaries and approves the tools; marketing operations owns the workflows, permissions and approvals within them; the brand owner owns the brand context agents use. Justworks puts the marketing standards with a marketing automation lead, “aligned with broader GTM technical and security standards.”
How do we keep humans in the loop without slowing everything down?
Approve by risk rather than by asset, as in the table above. Internal drafts need only a spot check of a sample. Public assets built from templates get a brand check and then one owner’s sign-off. Claims, prices, spend and regulated topics get a named approver every time.
Should AI agents be allowed to publish?
Start with no. Let agents publish only low-risk, templated assets to your own channels, and only after a run of checked and approved examples shows they get them right. Paid spend and anything with claims stays behind a person.
More in Review
All 9 Review articles15 min read
How to run brand compliance monitoring across partners and channels
Brand compliance monitoring checks your brand after it ships, on your own channels and in the assets partners, resellers, franchisees and affiliates publish. What to sample and when, how to check what you capture, how to escalate and what to measure, with a monitoring log to copy.
8 min read
What is brand intelligence? The monitoring meaning and the production meaning
Brand intelligence usually means data about how a brand is seen and discussed: listening, sentiment, share of voice, reviews and now AI answers. Vendors also use it for applying and checking brand knowledge while assets are made. What each meaning covers, the kinds of tools, and which one you need.
12 min read
What is content governance? A framework, a plan template and the AI version
Content governance decides who creates, approves, publishes, maintains and retires content, and by what standards. A six-part framework, the NHS website's approval chain as a worked example, a plan template to copy, and what changes when AI agents draft and publish.

