Agentforce governance, also called agent-aware governance, is Salesforce governance set up for AI agents that act without a human in the loop. It covers how you release Agentforce agents, what they're allowed to touch, how you trace what they did, and what you show an auditor afterward.
Most organizations we see designed their controls around human workflows, with a review before a change ships and an audit to catch what slipped through. An agent left to handle customer chats at 3 a.m. gets past both.
This page is a practical overview, and each of the four areas below has its own guide. Read this first, then go to the guide that matches your biggest gap.
What is agent-aware governance?
Until recently, AI in Salesforce mostly made suggestions, and a human made the decisions. An Agentforce agent chooses its own actions from the context of the conversation, and it can execute flows, run code and update records.
So oversight can't stop at approving a change up front. Our framework for agent-aware governance rests on three jobs your Salesforce environment has to keep doing:
- Enforce control: the same policy applies at the moment of action, whoever or whatever started it.
- Observe behavior: you can see changes, access and activity across orgs, agents included.
- Recover quickly: when an AI agent gets something wrong, you can put it back fast, without guesswork.
You won't get any of these from policy documentation. They live in your tools: the release pipeline, the permission model and the logs you actually keep.
Why do controls built for people fall short?
Break a validation rule and someone sees an error message straight away. Give an AI agent a bad instruction and it can hand out the wrong answer all night without anything failing.
A weekly change review won't catch that, and neither will an audit next quarter. Backups let you restore data, but they don't tell you whether a human or an agent changed it.
Salesforce secures the platform and draws the line plainly. Under its shared responsibility model for Agentforce, keeping privileges and actions tight and auditing activity is something you're responsible for, not Salesforce.
Getting it wrong has a business cost as well. Gartner forecasts that over 40% of agentic AI projects get canceled before the end of 2027, and inadequate risk controls are one of the three reasons it gives.
The four parts of Agentforce governance
We break the work into four areas, and each one has a full guide. The table lists the question each area answers and the Salesforce controls you'd use for it.
| Area | Question it answers | Main Salesforce controls | Full guide |
|---|---|---|---|
| Release control | Which agent version is live, and who approved it? | Version control, sandbox testing, agent versions, activation | Deploying Agentforce agents |
| Security | What can each agent reach and do? | Agent user, permission sets, sharing, Trusted URLs | Agentforce security practices |
| Audit and observability | What did the agent see, decide and change? | Session Tracing, Audit and Feedback, Field History Tracking, Setup Audit Trail | Agentforce audit trail |
| Compliance | Can you prove the controls work? | Change records, access reviews, BAA coverage, exported logs | Agentforce compliance |

1. Release control
Agents are metadata you can deploy, so you can build and ship them through the same pipeline as your other Salesforce metadata. Salesforce's architects suggest keeping the definition in Git and building and testing in a scratch org or sandbox, with sign-off before anything is deployed to production.
Versioning helps a lot. Only one version can be active at a time, and on API version 67 and earlier a committed version is locked against edits.
So if version 7 starts offering discounts it shouldn't, you go back to version 6 and activate it, which Salesforce says happens with zero downtime. Our guide to deploying Agentforce agents safely covers the full release path, and you can read how Flosum approaches agentic DevOps for rollouts at scale.
2. Security and access
Think of each AI agent as a non-human identity, the way you'd treat an integration user or a connected app. A service agent runs as its own agent user by default, but an employee agent works as the person using it and gets that person's permissions.
Agents follow the platform's usual access controls like permissions, field-level security and sharing, so one on a broad profile can see a lot. The same page says Trust Layer masking is off for agents, which leaves tight access as your main protection.
Our Agentforce security best practices post goes through ten controls, from least privilege to prompt injection, and names the Salesforce setting for each one.
3. Audit and observability
Say a case is closed overnight and no one on your team did it. Field history gives you the field, the time and the user, but not the customer request or other context that led the agent there.
You get that part from Session Tracing. It logs every turn of a session in Data 360 (formerly Data Cloud), with the reasoning, the actions it executed, any errors and its final response.
Permission changes sit in Setup Audit Trail, which only goes back 180 days. Export it if your audit period is longer.
Our Agentforce audit trail guide follows one record change back to the session behind it. For the flow, Apex and API events an agent triggers, our Event Monitoring post has the detail.
4. Compliance
SOX never mentions specific tools, so start with what the agent can reach. If it creates or edits records that end up in financial reports, count it as in scope.
PCAOB standard AS 2201 names three general IT controls: program changes, access to programs and computer operations. An agent that edits financial data runs into all three, since someone changes its instructions, it holds access, and it runs every day.
For HIPAA, check Salesforce's BAA restrictions first. They list which Agentforce services the Business Associate Addendum covers, and every product that could touch PHI should be on that list.
For Agentforce, the BAA asks you to use Trust Layer masking to the extent it's within your control. Salesforce has switched masking off for agents, so keep PHI out of an agent's reach through its access settings instead.
Our Agentforce compliance guide maps SOX and HIPAA controls to what agents do and lists the evidence auditors ask for. For the wider SOX toolset, see our roundup of SOX compliance tools.
Who owns governance for your agents?
Nobody owns all of it, so put a name against each area. A split we'd suggest is admins and architects on releases, the security lead on access, and compliance holding the evidence.
Keep the person who edits an AI agent away from the button that switches it on. Segregation of duties still applies when the change is a sentence of plain English rather than code.
Data needs an owner too. An agent can pass every test and then pull answers from the wrong data once it's live, so bring whoever owns that data into the group.
A readiness check for your agents
Before you scale an agent out to more teams, see if your team can say yes to each of these. A no shows where your issues are and which guide to read next.
- Can you name the version live in production, who approved it and who activated it?
- Could you roll back to the last good version today, along with the automation it calls?
- Does each agent run with only the objects, fields and actions it needs?
- Can you trace any record change an agent made back to the request that caused it?
- Are your audit logs kept, or exported, for the full length of your audit period?
- Is each Agentforce product that could touch PHI covered by your BAA?
Most organizations already keep this evidence for the rest of their org, and our Salesforce audit trail guide covers that side.
Where does Flosum fit?
Flosum DevOps puts every Salesforce change through one set of version control, approvals and audit trail, and Agentforce is on its list of integrations. Flosum Trust Center tracks who changed permissions or configuration and what users did, and it adds access analysis, audit logging and automated remediation.
For the wider program, see Risk & Governance. To go deeper on the framework, read our white paper, Governing Autonomous Salesforce Agents.
FAQs
What is agent-aware governance?
It's Salesforce governance built for AI agents that act on their own. Policy applies at the moment an agent acts, its activity is logged, and you can roll back fast when it gets something wrong.
What does Agentforce governance cover?
Four areas. Release control and agent access come first, then the audit trail of what agents did and the compliance evidence for SOX and HIPAA.
Who is responsible for governing Agentforce agents?
Mostly you. Salesforce secures the platform, but its shared responsibility model leaves agent privileges, action limits and activity audits with the customer.
How do you roll back an Agentforce agent?
Reactivate the last version that worked. Only one version can be active, so switching back takes the newer one out of service without a rebuild.
Does the Einstein Trust Layer govern agents for you?
No, but it helps: it adds prompt defense, and third-party models don't store or train on your data. Masking is off for agents, though, so access, change control and audits are still up to you.
Book a demo to see how Flosum tracks agent changes, permissions and actions across your Salesforce orgs.
Thank you for subscribing




