---
title: Instrument before you optimize: what a zero taught me about GTM
description: Six months running go-to-market for an AI-native data startup. The most valuable finding was a zero — and the instrumentation that made the zero trustworthy.
date: 2026-08-31
tags: [go-to-market, sales ops, ai agents, instrumentation]
techStack: [AI Agents, CRM Automation, Email Sequencing]
canonical: https://harrison.build/blog/instrument-before-you-optimize.html
---

# Instrument before you optimize: what a zero taught me about GTM

I just wrapped six months of GTM consulting for an AI-native data platform coming out of stealth. The engagement produced pipeline systems, an agent-driven prospecting desk, and a live revenue forecast — but the most valuable thing it produced was a zero. Cold outbound, measured honestly, had converted exactly zero companies to a real conversation. Not a low rate. Zero.

That number was almost impossible to see. The raw dashboard said the motion was fine: a healthy-looking reply rate, opens accumulating, campaigns segmented by persona and vertical. Every one of those signals was lying. Most "replies" were out-of-office autoresponders. Open tracking was disabled on two campaigns, which made their engaged companies look dead and made opens appear meaningless everywhere else. Link tracking was off across the board, so "nobody clicked" measured the instrumentation, not the buyers. A company's history lived in whichever CSV it last appeared in, recoverable only by diffing files.

So before diagnosing anything we rebuilt the measurement layer: one canonical ledger with a single record per company across the CRM, the email tool, and the ticketing system; reply classification that separates a human answer from an autoresponder; tracking actually turned on. Only then did the funnel tell the truth, and the truth was clarifying in the way only a clean zero can be.

The zero settled an argument that could have burned months. If cold outbound converts at some low-but-nonzero rate, you can always argue for one more copy rewrite, one more segment, one more sequence. A measured zero — with deliverability proven by the autoresponders themselves, personalization verified by sampling, and targeting confirmed against the ICP — rules all of that out. Good copy, delivered, aimed correctly, and nobody cared. That is not a wording problem. It is an offer problem, and rewriting the email would have reproduced the result exactly.

It also revealed where the actual signal was: every closed deal had come through a warm path. The highest-expected-value work in the entire system turned out to be a short list of companies where a founder connection already existed — free touches nobody had systematically worked, invisible until the ledger existed to surface them.

The other half of the story is how the work got done. The prospecting system ran as AI agents with human approval gates: agents did enrichment, validation, and daily review queues, metered to the humans' real decision capacity; humans made every judgment call. The agents were also confidently wrong three times — mixing per-lead and per-company denominators, trusting replies that were autoresponders, tuning a rule against a meaningless outcome variable. Each time, the fix was the same discipline: check the denominator, check the outcome variable, read the underlying records. Agents multiply your legwork; they do not substitute for reading the data yourself.

The lesson I keep re-learning is old and unglamorous: instrument first, then optimize. Optimization applied to unmeasured systems does not just waste effort — it manufactures confident wrong answers. The zero was worth more than any plausible-looking positive number, because it was the only number in the system we could trust.

This engagement — like most of my work now — was built with AI agents doing the legwork. The judgment calls, the pushback, and the mistakes worth owning were mine.