The GTM Engineer Agent-Deployment Loop
Turn a sales function into an AI agent by shadowing your best rep, then redeploy the humans up-market
- Difficulty
- Advanced
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 95%
A repeatable method for converting a repetitive go-to-market workflow (inbound triage, outbound prospecting, install-base upsell) into a supervised AI agent. A GTM engineer shadows the highest performer, documents the legible workflow, builds an agent that makes the judgment calls but keeps a human on final send, tracks the same KPIs the humans were held to, and only cuts over once conversion holds flat. Freed headcount moves to higher-value selling.
Origin
Developed by Jeanne DeWitt Grosser at Vercel; a direct descendant of 'Project Rosalind' she attempted at Stripe in 2017 (a company-universe database for targeted outbound that failed pre-AI because the false-positive rate was too high). Named after Rosalind Franklin, who mapped DNA.
Core principles
- 01Start with the most 'legible' workflow — one you can write down, that is replicable and mostly deterministic — because AI does those well
- 02Model the agent on your best performer's actual behavior, not a generic playbook
- 03Always keep a human in the loop reviewing and hitting send until reject/edit rates fall
- 04Redeploy freed humans up the value chain, not out the door
- 05Measure the agent against the exact KPIs (lead-to-opp conversion, touches, time-to-convert) the humans were accountable to
How to run it
- 1
Pick the most legible workflow first
Choose a function you can literally write down step-by-step and that is mostly deterministic. Grosser started with inbound lead triage before outbound because inbound is more replicable; deep-enterprise prospecting comes last because it spans multiple org-chart layers.
Pro tip Lower-market segments are easier to automate first because there's usually a single decision maker and less customization needed.
Watch out Don't start with large-enterprise prospecting — you may use an agent for research but not to send messages there for a long time.
- 2
Shadow the highest performer
Have the GTM engineer sit with the best individual in that function and watch the real workflow: the seven browser tabs, the LinkedIn lookup, the company research, the database attribute pulls, the ChatGPT step. This informs the initial agent workflow.
Watch out Model on someone with real expertise. Grosser found the agent sending things she wouldn't have, because it was modeled on a rep with two years of experience, not her twenty.
- 3
Let the agent make the judgment calls, human hits send
For inbound, the agent decides (a) is this lead likely qualified and (b) what to say. It does deep research, pulls from internal databases, and drafts the response — but a human reviews every one and actually clicks send.
Pro tip Feed every human reject/edit back into the agent so it keeps improving.
- 4
Track the same KPIs and hold conversion flat before cutover
Monitor lead-to-opportunity conversion, number of touches, and time-to-convert throughout. Only cut over when the agent holds conversion flat versus humans. At Vercel this took six weeks; the agent held conversion flat and cut the number of touches (it responds instantly instead of leads sitting in a queue overnight).
Pro tip The agent's edge is speed of response, not just cost — it kills the dead time of leads waiting in a queue.
- 5
Redeploy the freed headcount up the value chain
Vercel went from 10 inbound SDRs to one agent-QA, moving the other nine to outbound. The lead agent runs full-stack for ~$1,000/year versus over $1M in SDR salary — a 90%+ cost reduction on that function.
Pro tip The average SDR could have gone straight into outbound or SMB closing anyway; you're just using more of their capacity out of the gates.
In the wild
A single GTM engineer spent ~25-30% of his time over six weeks building a lead agent on Vercel's own workflow SDK and AI gateway. It triaged inbound leads and drafted responses with a human reviewer.
→ Went from 10 SDRs to 1 QA-ing the agent; conversion held flat while touches-to-convert dropped; nine reps redeployed to outbound; agent runtime cost ~$1,000/year.
Common mistakes
Modeling the agent on an average performer
If you encode a mediocre or junior rep's behavior, the agent inherits their gaps — Grosser caught the agent sending messages she, with 20 years of experience, would never have sent.
Trying to automate the hardest workflow first
Starting with deep-enterprise prospecting (multi-layer org charts, business-line triangulation) fails because the workflow isn't legible or deterministic; begin with inbound or low-market outbound instead.
Is it for you?
Best for
Founders and revenue leaders at companies with ~10+ people in a repetitive GTM function who already have a documented, replicable sales process
Not ideal for
Early startups with no written playbook, or pure deep-enterprise motions where every deal is bespoke
From the transcript
“we have them shadow the highest performing individual in that function”
“we had 10 SDRs doing this inbound workflow and now we just”
“have one that is effectively QAing the agent. The other nine we deployed on outbound”
“we let the agent make a”
“It ultimately then does some deep research, pulls in a bunch of information from our databases and crafts a response. But we have a human…”
“we felt confident going from 10 to one”
From the episode
What world-class GTM looks like in 2026
Jeanne DeWitt Grosser (Vercel, Stripe, Google)