Let the model draft, never let it send

There is exactly one rule that separates AI that saves a rep an hour a day from AI that costs a company its domain reputation, and it is about send authority, not about model quality.

A model may draft anything. A model may send nothing.

That sounds absolutist and it is, deliberately, because every real disaster in AI-assisted outbound has the same shape: somebody removed a human from the last step to gain throughput, and the failure that followed was not a bad sentence but a bad sentence delivered four thousand times before anybody looked.

Where autonomy is fine

The rule is narrower than it sounds, because most of the work is not sending.

Research is fully automatable. A model reading a company's postings, filings, and recent news and producing a two-line summary is doing work that is verifiable at a glance and carries no external consequence if it is wrong. If the summary is nonsense, the rep notices in three seconds and discards it. Let it run unsupervised.

Drafting is fully automatable, with the same logic. A first draft that the rep edits is strictly better than a blank page, and the editing step is the review. The failure mode of a bad draft is a mildly annoyed rep.

Classification and routing are fine. Scoring a lead, tagging a reply as interested or not, deciding which sequence somebody belongs in. These have consequences but they are internal and reversible, and errors surface quickly because a misrouted lead behaves oddly.

Summarizing a call and writing it back to the record is fine, and is probably the single highest-value automation available to a sales team, because it fixes the thing reps genuinely will not do by hand.

What all of these share is that a mistake is caught by the next human who looks, and no external party has been affected in the meantime.

Where it is not

Sending an email to a prospect. Posting anything publicly. Replying to an inbound message. Committing to a date, a price, or a scope. Anything that touches money.

The distinction is not about how good the model is. It is that these actions are irreversible and externally visible, which means the error budget is zero and the review has to happen before rather than after. A model that is right ninety-eight percent of the time and sends two thousand emails a week is sending forty embarrassing emails a week, and embarrassing is doing a lot of work in that sentence, because the two percent failures are not evenly distributed. They cluster on unusual records, which are often the largest accounts.

There is a common counterargument, which is that humans also make mistakes at some rate and nobody proposes reviewing every human email. This is true and it misses the mechanism. A human making an error makes one error. An automated system making an error makes the same error systematically until somebody notices, and the noticing is the bottleneck, not the erring.

The rule per step

Prospect selection: automate, review the list once at the start.

Research: automate, no review needed.

First draft of the opener: automate, rep edits the first line by hand. This is the step where the human minute buys the most, because the observed detail is what determines whether the email works at all.

Send: human, always. In practice this means a queue the rep clears rather than a scheduler that fires, and the queue should show the whole message, not a preview.

Reply handling: model classifies, human responds. Auto-replying to a prospect is where reputational damage lives, because a reply arrives with context the classifier did not have and the response lands in an existing conversation with an existing tone.

Follow-up timing: automate freely. Timing has no content to get wrong.

Call summary and write-back: automate, rep corrects. The correction rate is low and the value is high.

Forecast and pipeline hygiene suggestions: automate the suggestion, human accepts. A model marking your deals closed-lost on its own is a fast way to lose trust in the system entirely.

In practice the failure is rarely a model saying something outrageous. Models are reasonably well-behaved now. The failure is subtler and more corrosive: the model produces forty emails that are individually fine and collectively identical in rhythm, and the prospect who receives two of them, or the prospect who compares notes with a peer, recognizes the pattern instantly.

This is why the human step cannot be a rubber stamp. A rep clicking approve forty times in ninety seconds has automated the sending and merely added latency. The review has to include at least one edit, and the discipline that works is to require the first line to be typed rather than accepted, which is both the highest-value sentence and the one a model is least equipped to write, since it depends on judgment about what is interesting rather than what is true.

Draft everything. Send nothing. Type the first line yourself.

— DealArena Team