The Lock Is the Trigger: When You Commit, the AI Does the Homework

There is a moment in every marketing engine where uncertainty collapses into commitment. You stop debating who the ideal customer is and lock the ICP. You stop reshuffling growth bets and lock the matrix. In T2D3 OS that lock is a first-class act — it freezes the decision, snapshots it, and makes it the ground truth every downstream module builds on.

Which made the next thing embarrassing to admit: for most of our modules, the most information-rich moment in the whole product used to produce… a link. Lock your personas and a card would appear: "Value Props unlocks automatically — start here." The system knew exactly what had just become possible. It had the grounding, the prompts, the evidence. And it waited for a click.

We call that the click-inviting anti-pattern: the loop is closed on paper, open in practice. This week we closed it in practice. The lock is now the trigger.

What happens now when you commit

Every example below is from the live acceptance runs we shipped with — real locks, real model calls, real data landing in the running product.

Lock your Growth Matrix → your GTM plan drafts itself. The matrix lock is the moment your 90-day plan becomes maximally groundable: the bets are chosen, their revenue estimates committed. So the lock now drafts the full 30-60-90 plan as editable cards — in our test run, nine initiatives, each with an owner role and an evidence-citing rationale — and pre-computes the north-star goal recommendation with its arithmetic shown: "the $1.2M → $3M target requires roughly $150k of net-new ARR per month, and your matrix puts $1.05M of budget behind two motions — so the 90-day goal has to prove those motions can carry that run rate." The goal field itself stays empty. Adopting it is your click, not the machine's.

The same lock derives your budget targets. The spend plan's four required numbers — pipeline target, ARR target, target CAC, monthly budget — used to be bare fields defaulting to nothing. Now they arrive pre-derived, each with a one-line "why" ("new revenue goal $1.8M × a 4x pipeline-to-closed-won ratio = $7.2M pipeline target"), cached and waiting when you open the page. And when the evidence isn't there, the system refuses rather than invents: in our first test run, with no Growth Calculator data, the recommender declined outright — "run the Growth Calculator or set a revenue goal, then try again." An honest refusal is a feature. We test for it.

Lock your ICP → your ABM account list builds itself. In the acceptance run, the ICP lock kicked off a real market search and AI fit-scoring with no one watching: one hundred accounts landed tiered and ranked, each carrying its rationale — "a US medical-billing software vendor with ~$78M revenue implying a few hundred employees, squarely in your 50–500 band; denials reconciliation is its core domain." The work self-drains in the background, a hundred accounts per leg, narrating as it goes ("Scoring 100 of 2,867…"), then hands off to building the buying-committee contacts — best-fit accounts first.

A locked Website Audit now watches your live site. A locked audit is a snapshot of a moving target, and knowing it had gone stale used to be one hundred percent human-remembered. Now a weekly watch re-fetches the audited pages and computes what actually moved — deterministically, no model involved: "1 audited page no longer exists." It proposes the re-run; it never spends it. A full re-audit crawls your site and replaces a report your team signed off on — that stays your call.

The fences: initiative without trespass

"The AI acts on its own" is exactly the sentence that should make you nervous, so the engineering here is mostly fences — and they are checkable properties, not policy hopes:

• First draft only, ever. A background draft fires only into empty space. Typed a goal? The recommendation stays out of it. Started your plan, your account list, your relay kit? The trigger skips — a background job must never extend, rearrange, or "improve" work a human began. The buttons remain for that, because re-running over your own work is a decision, not a default.
• It drafts into your modules; it never creates them. Opening a module is the human's declaration that this work is a thing. No lock ever conjures one.
• Gated as you, not as the system. The trigger checks your plan's feature access as the person who locked — background work your subscription can't see simply doesn't run.
• Human gates hold under automation pressure. The ABM chain will happily build accounts and contacts all night, but it stops cold at the money and judgment boundaries: revealing contact data (spend), sending anything (outbound), and the compliance jurisdiction — a decision only a human may supply. In the live test we verified the negative: with the jurisdiction unconfirmed, the chain drafted zero sequences. The proof of a gate is what doesn't happen.
• Everything says why it exists. Every proactively-created draft carries its provenance — drafted because your ICP locked — and narrates while it runs. Work that appears without explanation is indistinguishable from a bug, even when it's right.

One more fence proved itself mid-test, unprompted: the relay-kit generator produced a draft that failed its own quality validation twice — and the system refused to save it, leaving existing content untouched and surfacing an honest error instead. A colleague who says "that draft wasn't good enough, I didn't file it" is worth ten who file everything.

Why the trigger matters more than the model

Most AI products take initiative on a timer, a login, or a guess — which is why "proactive AI" usually means notification spam. The insight we'd offer any team building agentic software: tie initiative to conviction, not to time. A lock is the strongest signal a human emits in our product — this decision is final enough to build on. Triggering work there means the AI acts exactly when its grounding is best and its output is most wanted, and never before.

Four questions to ask of any tool that claims its AI "works proactively":

1. What triggers the work? A schedule and a guess produce noise; a human commitment produces signal.
2. What can it never touch? If the vendor can't name the fence between drafting and overwriting, assume there isn't one.
3. Can it refuse? A system that always produces something is inventing some of it.
4. Does the work explain its own existence? Provenance and narration are the difference between a colleague and a poltergeist.

Last week's essay ended with the human's role: every judgment call gets a recommendation, a devil's advocate, and your final word. This week completes the shape from the other side. You bring the conviction — the lock. The system brings the homework, already done, sitting in your review queue, touching nothing that was yours. That division of labor is the whole thesis: agents lead, humans steer, and the steering wheel is never wrestled over.