Every agent should ship in review mode

A common way to ship an AI feature is to launch it running, then add a pause button once something goes wrong. That ordering puts the burden of proof on the customer. They have to notice a mistake before they get a way to stop it. It also means the first version of the agent anyone uses is the least tested one, running unsupervised from day one.

SalesCrew's agents ship the other way around. Every agent has four possible modes: off, draft, review and auto. Off means it does not run. Draft means it writes something, an email, a summary, a next step, and stops there without entering a queue. Review means it finishes the work and puts it in a queue, where a person sees the artefact, the evidence behind it, a confidence score and the cost before approving, editing or rejecting. Auto means it runs and logs the result, with a fallback to review if its own confidence drops below a threshold.

Every agent SalesCrew ships starts in review. Not draft, because draft produces work a person still has to go and find. Not auto, because nothing has earned that yet. Review is the default because it is the only mode where the agent's output is seen by a person before it has any effect, every time, while the agent is still new.

A single on/off switch for an agent is too coarse, because one agent usually does more than one kind of thing. The inbox agent can classify a reply, and it can draft a response to send. Those carry different risk. A wrong classification mislabels one thread. A wrong send reaches a real person's inbox. Treating both the same, either both auto or both review, over-restricts the safe action or under-restricts the risky one.

SalesCrew's default policy table sets the mode per action class instead of per agent. Research, enrichment, tagging, scoring, summaries and internal tasks default to auto, because a wrong internal note is cheap to fix and nobody outside the CRM sees it. Classifying a reply, and suppressing a contact on unsubscribe or bounce, also default to auto for the same reason. The blast radius of a mistake is contained inside the CRM.

Anything that leaves the building defaults to review: an email, a LinkedIn message, an SMS, a marketplace proposal. So do stage changes and contact merges or deletes, because those alter records other people rely on. Anything that spends money, whether credits, connects or ad budget, is human-only. No mode ever lets an agent do it unattended, not even auto. That category has no ramp. Review is not a stepping stone to auto for spend decisions, because SalesCrew's guardrails treat money as categorically different from everything else an agent can touch.

The two actions that do move from review to auto in the default table, creating a deal from a positive reply and booking a meeting, move only after a clean week: seven days with no item in that queue rejected. That is how trust in a specific action is earned by evidence, one action class at a time, instead of assumed for the whole agent at launch.

Review mode has a real cost: someone has to look at the queue. SalesCrew's approval queue is built to keep that cost low. One queue grouped by agent. Each item shows the artefact, the evidence it drew on (a thread, a score, a knowledge-base source), a confidence number, and what it would cost to run. Approving takes seconds. The reviewer never has to rebuild the agent's reasoning from scratch.

What it buys is a mistake caught before it reaches a customer instead of after. An agent that sends 200 emails overnight and gets ten wrong has sent ten wrong emails to real people by morning. An agent that drafts 200 emails and queues them has sent zero wrong emails, and a person can reject the ten bad ones in the time it takes to skim a queue. The difference is not whether the agent makes mistakes. Every agent does. It is whether a mistake becomes an artefact a person can catch, or an action that already happened.

If you are evaluating any AI feature that touches customers or money, ask what mode it ships in and who decided that. "It can be turned off" is not the same claim as "it starts off, or in review, until something has earned auto". The first puts the burden on the customer to notice and react. The second puts it on the product to prove itself first, one action class at a time, with a record of what it got right before it runs alone.