The failure mode with AI in a sales team is not that it produces bad work. It is that it produces plausible work very quickly, and nobody agreed in advance who checks it.
By the time you notice, four hundred messages have gone out.
Draw the boundary first
Write down which steps AI may touch and which stay human. One page, agreed before rollout.
| Step | Reasonable position |
|---|---|
| Company and prospect research | AI-assisted, verify specifics |
| Angle generation | AI-assisted, rep chooses |
| Writing the outbound message | Human |
| Reply handling | Human |
| Summarising a call | AI-assisted, rep corrects |
| Deciding who to contact | Human |
The line most teams should hold is that outbound copy stays human. Not for principle, for results. Generated outreach is recognisable because so many senders are generating it with the same models, and the pattern is what gets ignored.
Fabrication is the real risk
A model will state a funding round, a headcount or an org structure that is wrong, in the same confident register as one that is right.
Generic openers are forgettable. Confidently wrong specifics end conversations and occasionally do brand damage. Anything specific enough to be worth including is specific enough to need checking.
Make that someone’s job explicitly, or it becomes nobody’s.
Measure the right output
AI makes volume nearly free, which is exactly why volume is the wrong metric to celebrate.
Track positive reply rate and meetings held. If those hold steady while output triples, the campaign is producing more noise, not more pipeline, and the account risk is rising underneath it.
Where the tooling sits
Wherever you land on the boundary, the campaign itself should hold the human side of it.
Doreach provides merge tags with fallbacks and spintax, so copy a rep wrote goes out at volume without arriving identical, plus validation that blocks a launch on broken personalization. Research your team does upstream lands in the campaign as text a person chose to send.
Roll it out narrowly
One team, one workflow, four weeks. Compare against the team that did not change anything.
Broad rollouts of tools nobody has tested produce a lot of activity and very little evidence, and they are hard to walk back once the habits are set.
What to do next
Write the one-page boundary document and get the team to disagree with it out loud. The steps people argue about are the ones where the policy actually matters.
