Four phases of AI extraction in go-to-market sales

Executive overview

Most GTM teams today have human sellers talking to human buyers, yet the average rep spends only 25% of their time actually selling. AI is beginning to strip away all the non-selling work — prep, CRM updates, forecasting, coaching — before eventually replacing sellers themselves, then buyers, and finally dissolving traditional functional boundaries altogether. Mark Roberge maps this shift into four distinct phases, each further removing humans from the transaction. The practical advice is to focus on phase one now, using free AI agent tooling, rather than waiting for the science-fiction end state.

The biggest near-term leverage is not replacing sellers but freeing them — pushing selling time from 25% to 75%.

The four phases of GTM AI extraction

• Phase 1: Human sellers and human buyers remain, but AI handles all surrounding work (prep, CRM, forecasting, coaching).
• Phase 2: The seller role becomes an AI agent; early experiments visible in PLG funnels and SMB motions at companies like HubSpot.
• Phase 3: Buyers also become AI agents — AI sells to AI, removing political and relationship bias from purchasing decisions.
• Phase 4: Organisational functional boundaries (sales, finance, product, HR) blur significantly as AI removes the human limitations that originally created them.
• We are at the very start of phase 1; 2026 is projected to be the inflection year.
• Phases 3 and 4 are conceptually important but practically distant.

The selling-time problem

• Industry average: reps spend only 25% of their time in actual selling interactions.
• Best-in-class teams (e.g. Vanta) have reached roughly 40% — still considered exceptional.
• Non-selling time is consumed by meeting prep, CRM updates, demo building, forecast reviews, and administrative tasks.
• Phase 1 target: get selling time from 25% to 75% through AI automation of surrounding tasks.
• Manager-to-rep ratios can expand from roughly 8:1 to 20:1 as AI takes over repetitive oversight work.
• No significant spend is required; free tiers of ChatGPT, Claude, or other LLMs are sufficient to start.

A sequence of agents for phase 1

• Pre-meeting agent: reads everything about the prospect company and contact, then builds a customised meeting plan using the team's full methodology, product data, and battle cards.
• Practice agent: stands up a simulated buyer so reps can rehearse before the real meeting.
• In-meeting coaching agent: listens in real time and prompts the rep on what to ask, which story to tell, or which case study to show — replicating what great managers already do.
• Post-meeting agent: handles all follow-up — CRM updates, forecast entries, demo scripts, follow-up emails, and account notes.
• Skill-development agent: tracks rep performance across meetings, scores specific competencies (e.g. sense of urgency, champion development), and auto-generates personalised training clips.
• Together these agents maximise close rates while dramatically increasing the proportion of time reps spend in genuine selling conversations.

AI versus human forecasting accuracy

• Current win/loss and churn reporting is a telephone game: buyer tells rep a polite lie, rep tells manager a filtered version, and so on up to the board.
• Portfolio companies are now running parallel reports: what humans say about lost deals versus what AI analysis concludes.
• AI is winning more than half those comparisons already.
• AI can analyse call recordings and behavioural signals across dimensions no human can track simultaneously.
• The implication is that AI forecasting will surpass human intuition well before phase 2 arrives.

What phase 3 and 4 mean for organisations

• Phase 3 removes relationship bias: purchases would be driven by fit and metrics rather than golf trips or dinner relationships.
• Volatility increases — AI buyers may switch vendors the moment key metrics shift, with less inertia than human buyers.
• Phase 4 argues that today's functional org-chart design is an artefact of human cognitive limits, not optimal structure.
• Precedent from previous technology waves (mass production, electricity, computing) suggests jobs transform rather than disappear.
• Humans move from doing the work to directing and governing the algorithms.