The general question of whether AI replaces salespeople produces a general answer that helps nobody. The useful version is stage by stage, because the answer changes at every stage and the changes are not subtle.
The process, stage by stage
| Stage | AI handles | Human owns | Why the split falls here |
|---|---|---|---|
| Market and account targeting | Building and filtering lists, matching against fit criteria, enrichment | Defining the criteria and approving exceptions | Filtering is mechanical; deciding what "good fit" means is strategic |
| Research and briefing | Assembling company context, structure, developments and history | Reading it critically and checking sources | Assembly is fast and verifiable; confident inaccuracy needs a human check |
| Prioritisation | Scoring and ranking against agreed signals | Overriding the score where local knowledge contradicts it | Scores encode the past; local knowledge covers what the data cannot see |
| First contact | Drafting grounded, specific opening messages | Editing, sending and owning the relationship from message one | The draft is a starting point; the credibility attaches to a person |
| Discovery | Suggesting question sets and capturing the conversation | Conducting it, hearing what is unsaid, adapting live | The valuable information is the part the buyer has not articulated |
| Solution design | Assembling components, precedents and comparable scopes | Deciding what to propose and what to exclude | Trade-offs depend on delivery reality and relationship, not templates |
| Proposal | First draft from approved components | Commercial terms, positioning, final content | Contractual language carries consequences a draft cannot own |
| Objection handling | Surfacing relevant evidence quickly | Reading the real concern behind the stated one | Stated and actual objections frequently differ |
| Negotiation | Modelling scenarios and flagging precedent | Every decision and every concession | Requires authority, judgement and accountability |
| Close and commitment | Administration, sequencing, reminders | Asking, and judging readiness | Commitment is a relationship act |
| Onboarding handover | Summarising context and transferring records | Setting expectations and introducing the delivery team | Continuity of a named relationship is the point |
| Account management and renewal | Usage signals, renewal timing, comparable accounts | Judging when to raise expansion and when not to | Timing depends on service reality the data may not reflect |
What the pattern shows
Three things. AI concentrates in stages where the input is documented and the output is checkable. Humans concentrate where the relevant information is unstated and the consequences are commercial. And the boundary is not a line through the middle of the process — it alternates, stage by stage, which is why "AI-first" and "human-first" are both wrong as organising principles.
How the role changes
| Previously | Increasingly |
|---|---|
| Source of product information | Interpreter of fit and risk for this specific buyer |
| Manages activity volume | Manages fewer, more complex relationships |
| Prepares extensively by hand | Reviews prepared material critically and quickly |
| Competes on responsiveness and diligence | Competes on judgement and commercial credibility |
| Measured on activity | Measured on progression quality and outcome |
Implications for hiring, training and structure
- Hire for judgement and comfort with ambiguity, not for administrative diligence — the latter is now largely supplied.
- Train on the hard stages explicitly. Discovery, objection handling and negotiation used to be learned incidentally through volume; with volume compressed, they must be taught deliberately.
- Protect deliberate practice for junior staff. Someone who has never built an account picture from scratch cannot spot when a generated one is wrong.
- Restructure teams around fewer, deeper relationships rather than territory coverage arithmetic.
- Change the measures. Activity counts mean less when activity is cheap; measure progression and outcome quality instead, using the framework in how to measure success in AI-enabled sales.
A worked example: one stage in detail
Take discovery, because it is the stage most often mislabelled as automatable. A generated question set is genuinely useful: it ensures the commercial basics are covered and frees attention for listening. But the value of a discovery call comes from what happens when a buyer answers a question in a way that quietly contradicts something they said ten minutes earlier.
Recognising that contradiction, deciding whether to surface it now or later, and choosing language that does not embarrass the person in front of their colleagues is the whole skill. No transcript summary produces it, because the decision has to be made in the three seconds before the conversation moves on. The correct design is therefore assist-only: prepare with AI, capture with AI, decide as a human throughout.
The same test applies to every stage in the table. Ask what the stage is actually for, not what activity it contains. Research is for knowing things, and knowing things is transferable. Discovery is for understanding a situation nobody has fully articulated, and that is not.
Caveats
This map describes considered B2B purchases with multiple stakeholders and meaningful consequences. High-volume transactional sales sit much further toward automation at almost every stage. And the boundary moves — some stages currently marked human-owned will shift, though the ones involving accountability for a commitment will move last, if at all.
Key takeaways
- The division of labour alternates by stage; neither "AI-first" nor "human-first" describes it.
- AI dominates documented, checkable stages; humans own unstated information and commercial consequence.
- Removing the routine stages leaves a job made almost entirely of the difficult parts.
- Train discovery, objection handling and negotiation deliberately now that volume no longer teaches them.
- Measure progression and outcome quality rather than activity, which is now cheap.