The Forward Deployed GTM Engineer

My biggest takeaway from talking to companies standing up their GTM engineering orgs is that the unicorn hire is the Forward Deployed GTM Engineer.

These companies keep running into software engineers who don’t understand sales teams, or salespeople who have never coded.

An FDE typically embeds into a customer’s org and builds what they need. This one embeds into your sales floor.

The major change is that GTM teams are increasingly choosing to build rather than buy. And they want one central team building the AI rather than each rep building their own.

I read a job post for an FDGTME this month and recognized the last few years of our work.

What this person does:

1. Learns the alpha from the SDRs and AEs first (one team watched their top rep for a week before writing any code)
2. Builds the agent workflows
3. Builds the evals
4. Captures feedback from reps and the market
5. Ties agent behaviour to revenue

Most teams can build the workflows, but steps 3-5 are especially hard.

Karpathy’s line: “models can automate what you can verify.”

Coding got the best agents first because verification is immediate. Tests pass or fail.

GTM feedback is sparse, delayed and noisy.

We saw this ourselves this year. Two outbound variants tied in the first batch. In the next batch from the same list source, one produced a 3.4% positive-reply rate and the other 1.7%.

At a 3.4% baseline, detecting a 50% lift requires ~2,200 sends per variant. Detecting a 20% lift takes 12,000. Most teams declare a winner after 200 sends.

Waiting for revenue to grade every iteration makes the learning loop impossibly slow.

So the best GTM AI teams build a verification stack:

1. Evals on your own unique GTM tasks
2. Structured rep judgment on agent output
3. Live market and revenue outcomes that validate whether the evals and rep judgment predict pipeline

Despite what it sounds like, the entire process still isn’t automated (this will change soon, though). Even OpenAI is building an entire team to make the most of agentic systems for its GTM team.

This person consistently thinks through these four tasks:

1. Rules vs. model. Does this step need judgment, or would a rule do?
2. Data. What data does it read, where does it write back, who can see it, and where is it stored?
3. Evals. What does good look like, and how do we re-check it for new tasks and new model releases?
4. Oversight. Do we need a human in the loop for this?