Playbook

The same loop, every time.

Every GTM system I put into production goes through the same six steps: CRM changes and data hygiene, workflows and integrations, AI agents. It's slower on day one and much faster by day thirty, because nothing has to be redone.

The loop

  1. 01MeasureStart with a baseline: what's broken, how often, and what it costs. Every system here has a before number.In practiceBefore the inbound lead desk, leads waited a median 44 hours for a first touch.
  2. 02PlanWrite down the rules, the edge cases and what done looks like, and get them reviewed before any code.In practiceCall intelligence started as a plan: 75 entities, a scoring rule, and the bar the judge had to clear.
  3. 03Build safelySandbox first, with tests and evals on every change. Anything customer-facing gets a human gate and a kill switch.In practiceThe knowledge layer can't merge unless critical evals pass 100% and standard ones 98%.
  4. 04Roll outA dry run that writes nothing, then a small pilot, then everyone. Every bulk data change keeps a rollback file.In practiceThe account matcher ran on 200 leads before the 7,175-lead backlog.
  5. 05VerifyRead the result back from the live system, whether CRM records, sync status or model output, and compare it with the plan. Trace every AI call and its cost.In practiceThe inbound desk reads Salesforce back five minutes after it writes, to check the values landed.
  6. 06GuardData-quality checks, monitors and alerts, so drift or a regression shows up within days and someone is told.In practice56 SQL checks read a nightly copy of Salesforce and post new problems to Slack every Monday.

Principles

Habits that hold up.

Durable over clever

Every workflow runs through a Postgres-backed queue with checkpoints, so a crash resumes instead of starting over.

Humans gate anything customer-facing

Automation prepares the change; a person approves anything a customer or prospect will see.

LLMs write, code decides

Segmentation, scoring, subjects and links are deterministic code. The model only writes prose.

Fail closed on quality

If a quality check can't be parsed, it's a fail. Copy that still trips the spam gate after retries is dropped.

Measure before you trust

Models are picked by blind judges, retrieval is gated by evals, and every number on this site has a source.

When it went wrong

The loop is also how you recover.

An automatic sync between the sequencer and Salesforce got into a loop and marked 150 leads Lost. Because every field change is tracked, all 150 were restored from field history. The bad suppression records it created were deleted, and the design changed so Salesforce can no longer unsubscribe anyone in the sequencer.

Incident
Sync loop marks 150 leads Lost
Recover
Restore from field history · delete 741 bad records
After
150 of 150 restored
Guard
Design change: the loop can't recur
the suppression engine →

Want systems like these at your org?

Message me on LinkedIn