Governing Braze AI: Permissions, Limits and Approvals

Braze AI governance comes down to four controls: who can use each feature, what it sees, how much it runs, and who approves its output. Operator and the Model Context Protocol server act as the signed-in user, so you govern them through people. Agents carry their own permissions and daily limits, so you govern the agent itself.

Key Takeaways

  • Operator and the Model Context Protocol server act as the signed-in user, so govern them through user permissions and View PII.
  • Braze Agents carry their own permissions, a daily invocation limit and credit use, so cap and test them before scaling.
  • Turn on the approval workflow for campaigns and Canvases before widening Operator access, because it's off by default.
  • Write a policy for Operator's auto-approve toggle, because it's a per-session choice that resets on refresh, a new tab or logout.
  • Compare each surface's data retention with your own requirements before rollout, because Operator and Agents use different model providers.
  • Audit Operator use through the "Requested BrazeAI Operator Response" event in the Security Event Report.

What does Braze AI governance need to cover?

Every BrazeAI feature you switch on raises the same four questions. The table maps the three surfaces teams usually start with to the control that answers each one.

AI surfaceWho can use itWhat it can seeWhat limits itWho approves its output
BrazeAI OperatorUsers with the Use BrazeAI Operator permissionWhatever the signed-in user can see, plus what they typeThe user's own permissions, and approval pauses for costed actionsThe user, through action cards
Braze MCP serverThe signed-in dashboard user, through OAuthTools that don't return user-level PIIThe user's own dashboard permissionsThe user, in their AI client
Braze AgentsUsers with Edit Canvas and Data AgentsThe context passed in at runtime, up to 25 KB per requestDaily invocation limit and creditsThe builder, through test previews before launch

Our suggestion is to separate inherited access from owned access. Operator and the Model Context Protocol server act as the person using them, so their governance is your user management. Agents run inside Canvases, so their governance sits in the agent's settings and the Canvas approval.

Which permissions control access to Braze AI features?

Braze manages AI access through the same granular permissions as the rest of the dashboard. That's useful, because AI access can go straight into the permission sets you already maintain.

  • Use BrazeAI Operator: access to Operator for questions, setup, troubleshooting and ideas.
  • View, Edit and Archive Canvas and Data Agents: who can see, build and archive agents.
  • View, Edit and Archive Agentic Standards: who can read and change your automated campaign QA rules.
  • View Decisioning Studio Agent and View Decisioning Studio Audience: read access to decisioning agent configuration.
  • Approve and Deny Campaigns, Approve and Deny Canvases: who can pass the launch gate.

Braze states that Operator fully inherits the permission profile of the logged-in user. The Model Context Protocol server mirrors dashboard permissions in the same way. Permission changes apply on the next request, and deactivated dashboard users lose MCP access.

We recommend treating Use BrazeAI Operator as its own grant rather than a default in every builder's set. Start with people who already hold edit rights on the objects Operator will touch, since Operator can't go beyond them anyway.

How should you limit what Braze AI can see?

For the inherited surfaces, visibility is the control you hold. Braze states that Operator has no automated PII redaction layer in its request pipeline. A request can carry the typed message, tool outputs and content scraped from the active dashboard page.

Braze advises disabling the View PII setting for Operator users and not opening Operator on a user profile page. Put both on your rollout checklist, because they're decisions your admins make about people.

The Model Context Protocol server takes a different route, because tools that expose user-level PII aren't available. That makes it a sensible place for performance questions, while Operator stays the place for building.

Retention differs by surface, so put the figures in front of your security team. Operator requests go to OpenAI, which keeps them for 30 days for abuse monitoring.

On the US-02 cluster, OpenAI retention is zero. The Braze-provided model for Agents uses Google Gemini, and Google may keep inputs and outputs for up to 90 days.

Braze also states that data sent to OpenAI's API through Braze isn't used to train OpenAI models. It documents that EU data routing is not currently implemented for Operator. Compare that with your own data residency commitments before rollout.

Which limits keep Braze AI spend and volume predictable?

Agents are the surface where volume grows without anyone typing a prompt. Each agent has a default daily invocation limit of 250,000 runs. It can be raised to a maximum of 1,000,000 runs per day.

Timeouts consume credits, while rate limit errors and daily limit failures don't. By default each run must complete within 20 seconds. A slower run returns a null response where the agent is used.

Input is capped at 25 KB per request, and longer inputs are truncated. When an agent reaches its daily limit, Braze applies fallback values if you've configured them, and otherwise the output is null. Configure fallback values in Agent Console for Canvas Step Agents, so users still receive output when an invocation fails.

Operator handles cost at the moment of action. It always pauses for approval before it generates an image or runs a report, query or data sync. Those can carry a usage-based cost.

Braze's own guidance is to test with an Experiment Paths step. It also advises raising the daily limit before a full send, so invocations aren't capped mid-send. Our addition is to raise it only after you've read the run logs and agreed the credit estimate with the budget owner.

Where should a human approve Braze AI output?

Approval happens at two layers, and a governed rollout uses both. The first sits inside the AI feature, and the second is the launch gate on the campaign or Canvas.

Operator presents its changes as action cards, and auto-approve is off by default. Auto-approve resets when someone refreshes the page, opens a new tab or logs out and back in. We recommend writing a policy about who may turn it on.

Even with auto-approve on, Operator won't make a campaign active unless someone is there to confirm. Approved actions can't be undone through Operator, so the careful review has to happen before the click.

For the Model Context Protocol server, Braze recommends against an AI client's auto-mode for automatic actions. Campaign edits made through it are attributed to the signed-in user and appear in campaign edit history.

The launch gate is the approval workflow for campaigns and Canvases, which is off by default. Turn it on before you widen Operator access. Any saved change resets every approval status to Pending Approval.

Approve and Deny permissions also allow self-approval, so we recommend giving them to someone other than the builder. Campaign approval isn't supported for API campaigns or Transactional Email campaigns, so govern those routes through who can trigger them.

How do you audit Braze AI use after rollout?

Download the Security Event Report on a schedule rather than after an incident. The "Requested BrazeAI Operator Response" event gives an audit trail of the exact inputs provided to Operator. That shows you whether anyone pasted personal data into a prompt.

The manual download holds only the most recent 10,000 security events. You'll find it under Settings > Company Settings > Admin Settings > Security Settings. A busy account can pass that between reviews, and Braze documents an Amazon S3 export without the row limit.

For agents, the run logs show the input and output of each run. The version history shows inline diffs of every instruction change. Review an instruction change the way you'd review a Liquid change to a live template.

In what order should you roll out Braze AI governance?

This is the order we'd follow for a workspace switching on its first AI features.

  1. List every AI surface you plan to switch on, with one named owner for each.
  2. Build a permission set for AI builders holding Use BrazeAI Operator and Edit Canvas and Data Agents, and keep it out of general marketer sets.
  3. Remove View PII from Operator users who don't need it for their daily work.
  4. Turn on the approval workflow for campaigns and Canvases, and give Approve and Deny to people other than the builders.
  5. Set each agent's daily invocation limit, then test it on part of the audience with Experiment Paths.
  6. Compare the sub-processor, retention and data routing details with your security requirements, and schedule the Security Event Report review.

Steps three and four come before wide Operator access for a reason. Once Operator is part of daily work, removing a permission feels like taking a tool away. Setting it first feels like onboarding.

Which Braze AI governance mistakes should you avoid?

Name these in your governance review so they're quick to spot. The inherited admin happens when an admin tries Operator first. The first AI actions then run with the widest permissions in the account.

The sticky toggle is a policy that says auto-approve is off, when it's a per-session choice any user can make. The bypassed gate assumes approvals cover every send, while API campaigns and Transactional Email campaigns sit outside it.

The profile-page prompt happens when someone opens Operator on a user profile and the page content goes with the request. The uncapped agent gets its daily limit raised to fit a full send before anyone has read the test logs.

Where does this connect to the rest of the programme?

Start with the feature order in which Braze AI features to adopt first. Operator's action cards and data handling are covered in what to delegate to BrazeAI Operator. The permission model for AI clients is set out in connecting Claude or ChatGPT to Braze.

Decisioning agents have their own controls, covered in guardrails for Braze AI decisioning agents. Automated QA before launch is the subject of Agentic Standards, and customer-facing agents are covered in Braze conversational agents.

Our case study on unifying guest profiles for a global hotel group covers a related control. It describes approval workflows embedded in the Braze interface.

As a Braze implementation partner, CustomerIK recommends agreeing permissions and approvals before the first AI feature goes live. They're far easier to agree while nobody depends on the tools yet.

Frequently Asked Questions

1. Who can use BrazeAI Operator?

Only users granted the Use BrazeAI Operator permission. Operator then inherits that user's permissions, so it's bound by what that user could do directly.

2. Can Braze AI features see personal data?

Operator sees what the signed-in user can see, so remove View PII for Operator users who don't need it. Model Context Protocol server tools that expose user-level PII aren't available.

3. Is there a limit on how often a Braze Agent runs?

Each agent has a default daily invocation limit of 250,000 runs, which can be raised to a maximum of 1,000,000 runs per day.

4. Does the approval workflow cover campaigns built with AI?

It applies to all campaigns and Canvases in the workspace once you turn it on. It's off by default, and campaign approval isn't supported for API campaigns or Transactional Email campaigns.

5. How do I audit what people asked Operator?

Download the Security Event Report and look for the "Requested BrazeAI Operator Response" event. It gives an audit trail of the exact inputs provided to Operator.

Sources

CustomerIK is a Braze implementation partner working across onboarding, technical integration, marketing operations and customer data management.

If you're switching on BrazeAI features and want the permissions and approvals designed before launch, let's talk.

Preferences

Privacy is important to us, so you have the option of disabling certain types of storage that may not be necessary for the basic functioning of the website. Blocking categories may impact your experience on the website.

Accept all cookies
Accept all cookies

These items are required to enable basic website functionality.

Always active

These items are used to deliver advertising that is more relevant to you and your interests.

These items allow the website to remember choices you make (such as your user name, language, or the region you are in) and provide enhanced, more personal features.

These items help the website operator understand how its website performs, how visitors interact with the site, and whether there may be technical issues.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.