Unseen

Framework — AI Readiness Series

Governance Model For Enterprise AI Agents In Salesforce

This governance model defines how an enterprise should govern AI agents operating inside Salesforce environments. It is intended for CIOs, enterprise architects, platform owners, security leaders, and governance bodies responsible for approving AI deployment.

Framework Governance Model

Series AI Readiness

Published May 2026

Purpose

The model assumes a simple truth:

If the institution cannot explain how the CRM behaves, it cannot responsibly delegate action authority to an AI agent inside it.

Governance Objectives

The model is built to ensure that every AI agent is:

bounded by explicit authority
observable after every action
interruptible when risk rises
reviewable by humans
attributable to named owners
governable as the Salesforce environment changes

Governance Layers

Layer	Primary question	Control objective
Policy	What is the agent allowed to do?	Define allowed action classes, data boundaries, and approval requirements.
Control	What stops misuse or drift?	Enforce permissions, human checkpoints, testing, and rollback paths.
Operations	How is behavior observed day to day?	Maintain logging, alerts, reviews, and issue handling.
Oversight	Who is accountable?	Assign clear ownership and decision rights for deployment, incidents, and change.

1. Agent Action Logging And Audit Trail Requirements

Every production AI agent must generate an audit trail sufficient to answer:

what the agent saw
what instruction or policy it used
what decision it produced
what action it took
what changed in Salesforce
who owned the agent and who approved the action class

Minimum logged events

Event type	Required fields
Agent invocation	Timestamp, agent name, version, use case, initiating trigger, business owner
Context retrieval	Objects and records accessed, source systems consulted, retrieval query or context policy, data sensitivity classification
Reasoning artifact	Prompt or instruction version, policy version, decision rubric version, confidence or rule match indicator
Output	Recommendation, classification, generated content, or action proposal
Executed action	Created or updated record, field-level changes, downstream automation triggered, approval path used
Human intervention	Reviewer identity, approval or rejection, comments, override reason
Exception or failure	Error type, affected record, escalation destination, recovery action
Model change	Model version, prompt revision, policy update, date released, approver

Audit trail requirements

Retain logs for at least the same period as equivalent change-governance and audit records.
Store enough context to reconstruct why a business record changed.
Separate agent identity from human identity, but link the two when a human initiated or approved the action.
Preserve prior values and post-change values for any write-enabled workflow.
Log policy references, not only raw prompts, so governance can trace decisions back to approved control logic.

2. Human-In-The-Loop Trigger Conditions

Human review must be mandatory when any of the following conditions are met:

the agent is about to change owner, stage, forecast category, contract status, entitlement, or other decision-critical field
the agent is about to trigger a downstream external action such as provisioning, billing, routing, or service escalation
the agent encounters missing, contradictory, or low-confidence context on a required record
the action touches a regulated data class, sensitive customer commitment, or financial-impacting process
the workflow includes an undocumented exception path
the agent recommends a policy exception or a reversal of prior human judgment
the record has been modified by an integration or privileged user within the defined conflict window
the action would cross business-unit, territory, or segregation-of-duties boundaries

Review modes

Mode	When to use it	Human role
Recommendation only	Early pilots, high-risk processes, or unclear environments	Human executes the final action manually
Approval gate	The agent can propose and pre-fill a change	Human approves or rejects before write-back
Post-action review	Lower-risk, repeatable actions with proven controls	Human reviews samples and exceptions after execution

3. Permission Boundary Definitions For AI Agents

AI agents should be governed by explicit authority classes rather than generic service-account access.

Authority classes

Class	Allowed behavior	Typical use
Observe	Read data and summarize state only	Search, briefing, anomaly detection
Recommend	Produce a recommendation or draft, but do not write	Next-best action, draft updates, triage suggestions
Execute bounded	Write within pre-approved objects, fields, and conditions	Data normalization, low-risk status updates, case classification
Execute elevated	Perform material state changes under explicit approval rules	Ownership reassignment, lifecycle progression, exception handling

Boundary rules

Create a dedicated principal for each production agent or agent family.
Deny broad admin-style permissions by default.
Define allowed objects, allowed fields, prohibited fields, and maximum action classes.
Require separate approval for any expansion from Recommend to Execute bounded, or from Execute bounded to Execute elevated.
Treat field-level security, approval rules, and external-action permissions as part of the agent boundary, not as afterthoughts.
Re-certify agent permissions on the same cadence used for privileged human access, and after any major process change.

4. Escalation Protocol When The Agent Encounters Undocumented Logic

Undocumented logic is not a minor inconvenience. It is a governance event.

Required protocol

Pause the action.
Log the triggering record, automation path, or conflicting rule.
Route the issue to the named system owner and business owner.
Classify the issue:
- undocumented rule
- conflicting automation
- permission ambiguity
- data ambiguity
- integration side effect
Determine whether the issue is:
- isolated and correctable
- systemic and deployment-blocking
Update the documentation, policy, or control design before resuming similar actions.

Escalation timers

Severity	Example	Response expectation
`P1`	The agent reaches a privileged or financially material action with unclear rule basis	Same day review; similar actions paused immediately
`P2`	The agent hits an undocumented exception path on a critical workflow	Review within `1` business day; affected use case remains constrained
`P3`	The agent surfaces a noncritical documentation gap or edge case	Review within `3` business days; tracked into documentation backlog

5. Change Governance For AI-Modified Records

Any record that can be modified by AI must be governed as part of the enterprise change system.

Required controls

versioned definition of the agent's business purpose
test evidence before production release
approval for changes to prompts, policy instructions, retrieval strategy, action scope, or model version
rollback plan for output quality or behavior regression
sampling and post-change review cadence
communication plan when the agent materially changes user workflow

Record-level expectations

AI-modified records must be distinguishable from human-only updates.
Field history or equivalent change records must capture before and after states.
Where possible, store an agent decision code or reason reference, not only free-text output.
High-impact updates should require dual evidence: machine action log plus record-history change trail.

6. Accountability Framework

If AI agents act inside Salesforce, ownership cannot be vague.

Required roles

Role	Accountability
Executive sponsor	Approves the business case and accepts residual risk at the program level
Salesforce platform owner	Owns the system surface the agent operates within
Business process owner	Owns the business rule set the agent is expected to follow
Security or risk lead	Owns access review, control design, and incident escalation standards
AI product owner	Owns the agent backlog, performance metrics, and release decisions
Operations reviewer	Owns day-to-day exception review and human-in-the-loop handling

Decision-rights matrix

Decision	Primary owner	Must be consulted
Approve first deployment	Executive sponsor	Platform owner, security, business owner
Expand action authority	Platform owner	Business owner, security, AI product owner
Change prompt or policy logic	AI product owner	Business owner, platform owner
Approve model version change	AI product owner	Security, platform owner
Pause deployment after incident	Platform owner	Executive sponsor, operations reviewer
Accept unresolved risk for continued use	Executive sponsor	Security, platform owner, business owner

Operating Cadence

The governance model should be run on a fixed operating rhythm:

weekly exception review for active agents
monthly control review for logs, escalations, and approval adherence
quarterly permission recertification and policy review
event-driven review after incidents, process redesign, major releases, or model changes

Minimum Go-Live Standard

No production AI agent should go live in Salesforce until the enterprise can show:

a named owner for the agent and the business process it touches
a documented action boundary
an auditable logging model
a human escalation path
a rollback plan
a change-governance process for future modifications

What This Model Protects

This model protects against the most common enterprise failure mode in CRM AI:

The organization thinks it is deploying intelligence, but it is actually scaling undocumented local logic, unclear authority, and invisible side effects.

Governance is the difference between those two outcomes.

Ready to assess your Salesforce org?

Email hello@unseen.so to request an AI Readiness Audit.