Agentforce
The Agentforce Approval Agent: what it routes, and what it will never decide
Salesforce built a prebuilt Approval Agent into Agentforce Revenue Management: reps submit and track approvals in conversation, reviewers get AI summaries and approve from Slack. It routes and summarizes. It does not make the call. The real project is the Advanced Approvals engine underneath it, and that is where the gotchas live.
A seven-figure deal sits in a VP’s approval queue, and the VP is on a plane, out of office.
Every RevOps team knows the next three days. The rep pings Slack, someone forwards the record, a second approver asks what changed since the last version, and the quarter closes on a deal that was technically approved on the thirtieth and unblocked by a hallway conversation.
Salesforce’s answer in Winter ‘27 is a prebuilt Approval Agent inside Agentforce Revenue Management. The first thing to be clear about is what it does and doesn’t touch.
It routes the request, searches for the right records, and writes the reviewer a summary of what they are approving. It does not make the approve-or-reject call. That boundary is the whole design, and it is the correct one.
The work the agent removes is the chasing, the forwarding, and the “what am I even looking at” that eats the days around a decision. The decision itself stays with a person.
So the real question is not whether to trust an agent with approvals. The question is whether the approval engine underneath it is set up well enough for the agent to help.
What the agent does
The Approval Agent sits on top of Salesforce’s approval machinery and gives it a conversational front door. A rep can submit an approval, recall one, and check status by asking, rather than clicking through records.
A reviewer gets an AI-written summary of the request and can act on it without reconstructing the context by hand. Underneath, Agentforce routes each request to the right approver through the rules you have configured.
Salesforce splits the agent into three named subagents, and the names are the clearest public guide to the split: Approval Management, Search Approval Records, and Summarize Multiple Approval Work Items.
The shape reads as one subagent to act on and move a request through its approval, one to find the relevant approval records in conversation, and one to digest a stack of pending items into something a reviewer can clear in a sitting. Treat the exact per-subagent capabilities as something to confirm in your own org against current release notes, because the public detail stops at the names.
The capability that changes the felt experience most is the summarization paired with Slack. With the Winter ‘27 dynamic approval notifications turned on, a group or queue approval sends every active member a Slack message carrying an AI-generated summary of the quote.
The approver can approve or reject from Slack on desktop or mobile, drilling into line items without opening the org. For an executive who rarely lives in the CRM, that is where days come off the cycle.
It is the same human-in-the-loop pattern that works everywhere else in Agentforce: the agent assembles and presents, the human decides.
The line it holds, and why approvals are the right place to hold it
A lot of agentic-AI messaging blurs toward full autonomy. Approvals are exactly where you want the blur to stop, and Salesforce drawing the agent’s scope at routing and summarizing rather than deciding is a feature, not a limitation.
Two reasons. The first is the audit trail. An approval exists to record that a named person with authority looked at a specific request and accepted the risk. If an autonomous process both approves a transaction and can alter the record behind it, you lose the traceable accountability an auditor relies on, and “the system approved it” is a poor answer in a controls review.
The second is segregation of duties. The person or process that requests a discount should not be the one that grants it, and an agent handed broad standing permissions collapses that separation unless someone designs against it.
Keeping the verdict human sidesteps both problems. The agent can prepare the decision, surface the right numbers, and chase the laggards, and the accountable person still clicks approve with their own identity on the record.
If you are rolling out agents more widely, the same discipline belongs in your agent-fleet governance: scope the Approval Agent’s permissions tightly, and never let its convenience erode the control the approval was built to enforce.
The real project is Advanced Approvals
Here is the part the keynote underplays. The Approval Agent requires Advanced Approvals plus Agentforce, and standing up Advanced Approvals correctly is most of the work. The agent is a thin, helpful layer over an engine that is not thin at all.
Advanced Approvals began life as a managed package on top of CPQ and now underpins approvals in Revenue Cloud. Its moving parts are worth knowing before you promise anyone a conversational approval experience:
- Approval rules evaluate a submitted record and decide who needs to approve. Every rule needs at least one condition, and the quality of your routing is entirely the quality of these rules.
- Approval chains send a request through a sequence of approvers in order, so a deal isn’t fully approved until each step signs off, which is how you model “manager, then finance, then legal.”
- Approval variables roll up data from child records, such as aggregating quote lines, so a condition can fire on totals the parent record doesn’t store directly.
- Smart approval and recall decide whether a changed quote needs re-approval and send recalls back to the original approvers rather than restarting the whole chain.
- Serial and parallel steps let one stage wait on a single approver while another fans out to several at once.
All of this runs on Flow approval processes, so the orchestration is the same Flow-based automation you already govern. Enabling it is a Setup step under Approval Settings, where you turn on Flow approval processes and Advanced Approvals together.
Two reported gotchas are worth checking before you flip anything in production. If the prerequisites aren’t met, the setup is reported to fail without a clear error, which turns a five-minute enablement into an afternoon of confusion.
Worse, enabling Advanced Approvals is reported to silently disable approver delegations configured on user records, which is precisely the out-of-office routing you were trying to fix.
Winter ‘27 does add first-class delegated approvers, period-bound to vacation dates or standing, so the capability moves into the engine itself. Verify how your existing delegations behave through the switch rather than assuming they survive it.
Licensing, cost, and status: verify before you commit
Three honest unknowns to close on, because this is new and the public detail is thin.
On licensing, the hard requirement is Advanced Approvals and Agentforce together. Confirm which Revenue Cloud edition and Agentforce entitlement your org holds before you scope a rollout, because the agent is useless without the engine license underneath it.
On cost, Agentforce actions meter on Flex Credits, and the published economics put a standard action around 20 credits, roughly ten cents, against packs listing near $500 per 100,000 credits. I could not find a credit figure specific to Approval Agent actions, so measure a realistic approval volume against your own consumption before you extrapolate.
The Flex Credits pricing breakdown is the model to start from, not a quote for this agent.
On status and timing, the Approval Agent rolled out on Agentforce Revenue Management’s own release cadence in September 2026, separate from the core Winter ‘27 go-live weekends. I would not state a single clean GA date for it.
Check the feature’s current status in your org and the live release notes before you put it on a go-live plan. A prebuilt agent that is still settling is something to pilot, not to bet the quarter-end on.
What to do
Start underneath the agent, not with it. Audit your Advanced Approvals setup first: are the approval rules routing correctly, do the chains model your real sign-off order, and what happens to your current delegations when Advanced Approvals is enabled.
Get that right in a sandbox and the Approval Agent becomes a real accelerator: reps submitting and tracking in conversation, reviewers clearing a summarized stack from Slack, days coming off the cycle. Skip it, and you have put a conversational interface on routing logic that was already sending deals to the wrong person, only faster.
The agent moves approvals. It is still your job to make sure they move to the right desk, and that the person at that desk is the one who decides. For where an approvals agent fits against the rest of the deal desk, the deal-desk and quote-approval playbook is the wider map.
Understanding the basics
What is the Agentforce Approval Agent?
It is a prebuilt agent in Agentforce Revenue Management that gives Salesforce’s approval process a conversational front end. Reps submit, recall, and track approvals by asking, and reviewers receive AI-generated summaries of what they are approving and can act from Slack. It routes and summarizes approval requests but leaves the approve-or-reject decision to a human.
Does the Approval Agent approve deals automatically?
No. The agent handles routing, record lookup, and summarization. The actual approve-or-reject decision stays with the designated human approver, which keeps the audit trail and segregation of duties intact. Autonomy applies to moving and explaining the request, not to the verdict.
What do I need to use the Approval Agent?
Advanced Approvals and Agentforce. Advanced Approvals is the rules-driven engine, originally a CPQ add-on and now part of Revenue Cloud, that decides who approves what through approval rules, chains, and variables, running on Flow approval processes. You enable it under Approval Settings in Setup, and the Approval Agent will not function without it.
How is this different from standard Salesforce approval processes?
Standard approval processes and Advanced Approvals do the routing and record-keeping. The Approval Agent adds a natural-language interface and AI summaries on top of Advanced Approvals, so people interact by asking and reviewing rather than clicking through records. It doesn’t replace the approval engine, it sits on it, which is why getting the engine configured correctly matters more than turning on the agent.
Standing up Agentforce Revenue Management and want the Advanced Approvals engine configured right before the agent goes on top of it? Talk to us. Getting the routing, the chains, and the governance correct, so the agent accelerates good approvals instead of fast-tracking bad ones, is exactly the work we do.
