Articles
Field notes
Deep dives, playbooks, and working notes on GTM Engineering.
Warm the Domain for 6 Weeks or the AI SDR Torches It in 9 Days
A cold domain dies in days under AI-SDR volume; a warmed one lands. The ramp is 6 weeks from 10 to 50 sends per inbox, on a send subdomain, under the 0.30% cap. Here is the schedule, the DNS, and the pool.
Read the article →
Most AI-in-RevOps Pilots Die in 90 Days. Here's the Autopsy.
88% of teams use AI, only 38% scale past the pilot. The survivors are not running better models. They shipped four guardrails the corpses skipped, and shadow mode is the one that flattens the death curve.
Lint Your Outbound Before It Lints Your Domain
AI slop reads fine to the sender and burns the domain by the thousand. Watch reputation fall off a cliff as the complaint rate climbs, then build the regex lint gate that catches slop before it ships, tied to the 0.30% cap.
Your AI-Readiness Score Is a Data-Readiness Score in Disguise
Everyone runs the AI pilot. Almost nobody scales it, and the reason is not strategy. Here is the 6-component readiness ladder and the 12-check scored gate to run before you buy the agent, with the completeness and dedup thresholds that turn it into a gate.
We Deactivated 123 Flows and Nothing Broke
An org running 282 automations dropped to 82, and record saves got faster instead of breaking. Here is the five-rung build order that made a two-thirds cut safe: inventory, dedupe, merge, retire, govern.
Prove the 10 Hours: A GTM Engineer Automation Ledger
A VP asked what my automations saved and I had a vibe, not a number. Here is the four-rung ledger that tags every flow with hours saved, run count, and failure rate, plus the schema and rollup query that defend the function net of failures.
Roughly 90% of Lead Scores Still Route on a Job Title
A VP title and a company size predict almost nothing about who buys. What buys is behavior: product usage and intent. Here is the scoring rubric that weights signals, the SQL that computes it, and the routing rule that acts on it.
Build vs Buy: The $30K Tool That Cost Us $190K
A license price is the smallest of five numbers you will carry. Run the 3-year TCO and the maintained cost jumps an order of magnitude nobody prices at signing. Here is the model, the decision tree, and the weighted scorecard that decide it.
Before You Give an AI Agent Write Access, Build the Blast Radius
An LLM with a Salesforce login and no guardrails can rewrite 160,000 accounts before lunch. Here is the three-layer control stack that runs before the agent does.
A Data Contract, Because Hoping Every System Agrees Is Not a Plan
The CROs dashboard said $14.2M, the board deck said $12.8M, and both were right. A wiki page cannot stop that. A data contract enforces one definition across enrichment, routing, and reporting: schema, validation, ownership, monitoring.
Stop Building Dashboards. Nobody Opens Them.
My forecast dashboard got opened four times a month, three of them me. The same numbers in a Slack digest got replies inside the hour. Push beats pull. Here is the build order that turns 93 static rows into one decision.
Your Enrichment Vendor Is Wrong 30% of the Time
Vendors self-report 95 to 98 percent accuracy. Run an independent SMTP test and roughly half of what they found bounces, because B2B contact data decays about 30 percent a year. Here is the verify-before-trust build that measures today, not ingest day.
The GTM Engineer's Definition of Done
It fires in production is not done. Idempotent, observable, reversible, documented, handed off. Here is the checklist that separates a demo from a system reps can trust.
Your 400 Custom Fields Are Debt an LLM Will Expose
Every custom field was a workaround for a question the CRM could not answer. An LLM answers those questions from the raw record. The fields are now debt, and the debt is coming due.
Your Stage Logic Lives in Seven Places and No Two Agree
Everyone puts stage exit criteria on a slide, then scatters the enforcement across validation rules, flows, and Apex that drift apart. Here is the same governance as one metadata type, a generic service class, and a test that proves the gate holds.
GTM Engineer Is RevOps With API Keys
More than half the boards I track now list GTM Engineer, Clay shows up in 80% of the posts, and most hires have under five years of experience. The title is inflating faster than the craft. The role is RevOps that stopped filing tickets and started shipping.
GTM Alpha: Firmographics Are the Price of Admission, Not an Edge
Employee count and industry code are what everyone already bought. Alpha is the signal your competitor cannot buy off a list, and defensibility runs in tiers: public, inferred, proprietary. Here is the ladder and the recipes to climb it.
The Big-Bang Cutover Loses 8% of Your Records. Parallel-Run Instead.
The dedup rule you switch on to keep the data clean is the same rule that silently drops 8% of your records on insert. Here is the six-stage runbook: map, dedupe, dry-run, parallel, cutover, verify.
The Approve Button Is a Feature, Not a Bottleneck
An AI agent drafted 40 emails and shipped them before anyone read one. Here is the five-rung approval gate I put between the draft and the send, with the block-kit payload and the guardrail logic.
Your Enrichment Job Runs Twice and Corrupts Data Silently
One accidental re-run created 1,400 duplicate contacts and a second invoice, and no alert fired. Without idempotency, every retry compounds the damage. Here is the build: keys, locks, backoff, dead-letter, and the ten-minute test.
Routing Is Latency. Every Minute You Sit On a Lead, the Odds Decay.
Round-robin whenever someone gets to it treats routing as a fairness formula. It is not. Routing is latency, and contact odds decay by the minute. Here is the decay curve, the queue design that survives its failures, and a worked example that reconciles to the meeting count.
MCP for RevOps: Giving Claude Governed Access to Your CRM
An MCP server is a door into your CRM for an AI. Build the door with a lock, a log, and a scope before you hand out the key. Here is the permission model.
Two CRMs, One Revenue Number, and a Board Deadline
The board wants one CRM by quarter end. Big-bang the merge and you lose the deal history plus a rollback. Parallel-run with a checkpoint keeps both.
Retiring Apex With Flows: A Migration Runbook That Won't Break Prod
Deleting the old trigger is how you break prod. The safe path is a five-stage ladder: inventory, shadow, cutover, stub, retire, and it promotes only the cluster you tested.
Rule-Based vs AI Lead Scoring: Run Both for 30 Days and Let the Data Pick
Nobody cuts over to AI scoring on faith and keeps their job. Run the AI model in shadow mode next to your rules for 30 days, score the same leads twice, and let conversion decide.
Git Tracks Every Line of Code. Your Forecast Has No History at All.
A $420K deal slid out of the quarter overnight and nobody could say who moved it. Field history plus a nightly snapshot-diff gives the revenue number a commit log: who moved the deal, when, and what it did. Here are the queries.
The Afternoon Flow That Silently Corrupted 3,100 Accounts
An LLM-generated Flow shipped in an afternoon with no idempotency and no fault path. Six months later it had overwritten 3,100 accounts and nobody saw it. The incident report, the broken flow, and the hardened version.
You Bought Twenty Tools. Five Layers Run the Engine.
Programmatic outbound has exactly five jobs: data, enrichment, signal, orchestration, sending. You can run all five on about five tools for roughly $1,200 a month. Here is the stack as a build order, the credit line that quietly triples, and a cost model that reconciles to meetings.
One Vendor Finds 45%. Four in Sequence Find 68%.
Buy a single data provider and you inherit its blind spot as your ceiling. Chain four in sequence, pay only on the miss, and verified coverage climbs to about 68 percent found, 62 percent valid, not the folklore 92. Here is the build order by marginal recovery.
Signal-Based Outbound: 10,000 Names or 620 Reasons to Call
Spray a static list and reply rates sit at 1 to 3 percent. Wait for a real buying signal and they hit 10 to 20 percent. Here is the build order, the deliverability ceiling, and a worked example that reconciles to the meeting count.
SDR to GTM Engineer: The Skills Ladder, Not the Quota Treadmill
Nobody gets the GTM engineer title for hitting quota three quarters running. They get it for shipping a workflow that books meetings while they sleep. Here is the five-rung skills ladder, in the order that compounds, with the day-90 artifact that ends the argument.
The AI SDR Sent 42,000 Emails and Booked Zero Revenue
Point an autonomous AI SDR at a list and it sprays. Volume on a broken ICP fails faster and torches the domain on the way down. Here is the send math, the reputation curve, and what the 3 percent who got real revenue did differently.