Sales: AI Agents
Lead generation, enrichment, scoring, sales intelligence and proposal generation, with the scoring model shown and the limits on anything sent to a human being.
Markdown. No sign-up, no email.
Sales work divides cleanly. Finding, researching and preparing are pattern work at volume, and machines are better at them than tired people. Judging, promising and relating are not, and the damage from automating them is disproportionate to the time saved.
Lead generation#
Finding companies matching the ideal customer profile: sector, size, geography, and the observable signals suggesting the problem we solve is currently expensive for them.
The profile has to be written before this runs. An agent given a vague profile returns volume, volume feels like progress, and a pipeline full of poor fits takes months to work through before anyone concludes the targeting was wrong.
Enrichment#
For each candidate: size, sector, technology stack, likely problems, and who decides.
| Enriched | Useful because |
|---|---|
| Company size and structure | Decides whether we are relevant at all |
| Technology stack | Tells you the actual problems, not the generic ones |
| Likely problems | Inferred from stack, sector and observable signals, and labelled as inference |
| Decision makers | Who signs, who blocks, who feels the pain. Rarely one person |
Every inference is labelled as one. "Probably running a legacy integration layer, inferred from job postings" is useful. The same sentence without its basis becomes a fact by Tuesday and is repeated to the prospect on Thursday.
Scoring#
Enterprise AI opportunity, Score 87/100 Fit 32/35: sector and size match, stack indicates the integration problem we solve. Timing 28/30: two relevant hires this quarter, published initiative. Reachability 15/20: decision maker identified, no warm route. Budget signal 12/15: comparable spend observable.
The breakdown is the deliverable, not the number. A score of 87 tells you to call; the reachability line tells you what to do first, which is find a warm route.
Sales intelligence#
Before any conversation: company brief, prospect briefing, likely pain points, suggested solution, proposal strategy.
This is where the hours go in real selling, and it is the least controversial thing to automate. It also has a failure mode worth naming: a briefing full of confident inferences produces a first meeting where the seller explains the prospect's business to the prospect and gets it slightly wrong. Separate what is known from what is inferred, visually, on the page.
Proposal generation#
Drafts of proposals, SOWs, presentations and capability documents.
Constrained by what we can actually deliver. A generated proposal is a plausible-sounding document, and plausibility is the danger: it will happily promise a capability we do not have, in a timeline nobody costed, because those sentences are common in proposals it has seen.
Every generated proposal is checked against three things before it leaves: can we deliver this, have we costed it, and does it commit us to a date nobody has agreed?
What stays with a person#
- Any first contact with a human being. Automated outreach at volume is how a brand becomes spam, and the damage lasts far longer than the pipeline it generated.
- Every price and every discount.
- Every commitment: scope, date, capability. A proposal that leaves the building is a promise.
- Qualifying out. Deciding a prospect is not a fit is a judgement with revenue consequences.
- Anything said about what we have delivered before. Reference claims are checked against what was actually delivered, because the one that is wrong is the one that gets checked.
The rule that governs the whole function#
A machine may prepare anything and promise nothing. Preparation is reversible and unattributed. A promise is neither, and this is the function where the two are easiest to confuse, because a proposal looks like a document right up until the moment it is a contract.