AI in sales
Before your sales agent hits send
I’m happy to have AI do the first draft. I’m less interested in letting it make promises I have to explain later.
Once an agent can research an account, draft a message, and suggest the next action, it is tempting to connect everything and let it run. Before doing that, I’d write down exactly which decisions it gets to make.
This is part of the sales playbook. The team needs to know who owns a conversation, what can go out under their name, and when the system has to hand the work back.
Separate preparation from permission
My starting point would be to let the agent prepare research, opening lines, meeting briefs, and follow-up drafts. A seller approves anything going to a buyer. Pricing, commercial terms, delivery dates, and claims about what the product can do stay with the people authorised to make those commitments.
The distinction should hold even when the wording sounds harmless. “We can support that” can become a very awkward sentence if implementation has never seen the requirement.
Give the agent approved product information and examples of claims it must not make. Keep an accessible record of the source material and the final message. If someone asks where a statement came from, the answer should be available.
Have a plan for the reply
A polite response does not mean a qualified opportunity. A request to reconnect later does not mean a meeting is booked. And a question about security is a question that needs a real answer.
I’d have the agent sort replies and propose next steps, with the owner reviewing anything ambiguous. Objections, complaints, requests to stop contact, and questions outside the approved material need clear handling. They should never sit in an automatic follow-up sequence because the system failed to recognise them.
For Canadian outreach, have the appropriate person review your contact permissions and applicable requirements before automating distribution. An agent finding an email address does not settle whether or how you should use it.
Test the awkward cases
Use examples that make the process uncomfortable: two companies with similar names, a contact who changed jobs, a buyer asking for an unsupported integration, a reply that says “not me,” and an old announcement presented as current.
Ask whether the system identifies the problem and hands it back. A smooth demo is easy to like. The useful test is what happens when the information is incomplete.
Anthropic’s guidance on agent tools argues for systematic evaluation of how tools perform. For a sales team, I’d make the evaluation concrete: wrong account, unsupported claim, missed reply, unnecessary follow-up, and time spent correcting the output.
Keep the first release small enough to inspect
Choose one segment, one owner, and a limited batch. Review the drafts together. Track useful conversations and correction work alongside activity. Sending more messages is easy to measure; it is not enough to justify expanding the system.
Agree on who can pause it and what would cause them to do so. Add more autonomy only where the observed results justify it. Some decisions can stay with a human indefinitely.
The experience I want for the buyer is simple: someone understands why they reached out, remembers what was said, and follows through. The machinery behind that can be clever. The conversation should still feel like a conversation.
Start upstream with account research in an agentic sales playbook, or bring me the sales motion you’re trying to improve.