The default agent policy table
Every action class an agent can attempt, and what it defaults to. This is the actual seeded table, published so you can copy it into whatever tool you use instead.
- The real seeded defaults
- Four modes explained
- Copy it anywhere
The short answer
- Four modes per agent: off, draft (writes the artefact and does nothing with it), review (queues it for a person), and auto (executes and logs, falling back to review below a confidence threshold).
- Every agent ships in review. That is not a suggestion in documentation; it is the seeded state on a new instance.
- One line never moves: anything that spends money, and anything outside the CRM, is human-only in every mode. An agent that can move a budget is a bad design regardless of how good the model is.
Why this exists
Most AI features ship with a single on-off toggle, which forces one decision to cover a hundred different risks. Logging a call and sending an email to a customer are not the same kind of action and should not share a switch.
A policy table splits them by class. Low-stakes, reversible, internal work runs on its own. Anything a customer sees waits. Anything that spends money never runs. That structure is worth copying whether or not you use SalesCrew, so the table is published rather than described.
The policy table, as seeded
Editable per instance under Agents, Guardrails. These are the starting values.
| Action class | Default | Why |
|---|---|---|
| Research, enrichment, tagging | auto | Reversible, internal, and wrong answers are visible in the record rather than in a customer's inbox |
| Logging, scoring, summaries | auto | Nothing leaves the building and everything is overwritable |
| Internal tasks and suggested contacts | auto | The cost of a bad suggestion is a person ignoring it |
| Classify a reply | auto | A label is a filter for attention, not a decision |
| Pause a cadence on reply | auto | The failure mode of not doing this is worse than the failure mode of doing it wrongly |
| Suppress on unsubscribe or bounce | auto | Delay here creates a compliance problem, so it should never wait in a queue |
| Any external email send | review | A customer sees it. That is the whole test |
| Any LinkedIn message | review | Same test, and the platform account is at stake as well |
| Any SMS or WhatsApp message | review | Same test, plus consent obligations that belong to you |
| Any marketplace proposal | review | A proposal is a commitment in the buyer's eyes |
| Create a deal from a positive reply | review, then auto after a clean week | Cheap to reverse, but a wrong deal distorts the forecast until someone notices |
| Book a meeting | review, then auto after a clean week | It occupies a real person's calendar |
| Change a deal stage | review | The forecast is a number people act on |
| Merge or delete a contact | review | Destructive and awkward to undo |
| Spend money: credits, connects, ads | human only | No mode changes this. Not now and not later |
| Anything outside the CRM | human only | If the product cannot audit it, an agent should not do it |
Seeded defaults on a new SalesCrew instance, September 2026. Every row is editable, and the last two are refused rather than merely defaulted.
The guardrails that run underneath
Mode is only half of it. Guardrails apply regardless of what mode an agent is in: at least three days between touches per contact, a per-cadence cap, per-mailbox and per-domain daily caps, a business-hours send window, suppression re-checked at send time, and a confidence threshold that drops an action back into review when the model is unsure.
Above all of it sits the kill switch: one control on the Agents page that stops every agent at once, with an MCP twin so it can be hit from a script as well as a screen. Whoever owns that switch is a decision to make at setup rather than discover in an incident.
How to adopt this without SalesCrew
Take the three columns and apply them to whatever automation you run. Write out every action your system can take, sort each into auto, review or human-only, and be honest about which ones a customer sees. Most teams find the auto list is shorter than they assumed and the human-only list is not empty.
Then add the part people skip: a record of what happened. Every tool call in SalesCrew is written to the audit log with the actor named as a user, a token or an agent. Without that, a policy table is a statement of intent rather than something you can check after the fact.
Start with everything in review
The seeded state on a new instance, and the only sensible first week.