A GTM system for fintechs.

Experiment loop: what the team does

Every bet starts as a hypothesis and the smallest test that could prove it wrong. Read the result. Then scale it or kill it.

Learning loop: what the team knows

An uncaptured correction is a repeated correction. So every shipped piece leaves one line of learning, and it goes back to whoever owns the fact.

Adoption loop: what the customer uses

Each product makes the next one more useful. The core product builds a balance and a history. The balance earns the right to be where the money sits, holding funds and paying suppliers and payroll. Spend tools put every line of spend in one place. The history qualifies the customer for credit when they need to grow.

Every step deepens revenue per customer. The next product is offered when behavior says it’s needed, not when a campaign is due.

So every play is written trigger first. It names the behavioral event and the next product. It says whether that product is available. And it names the holdout that proves the play caused the lift.

The trigger is the asset. The copy is replaceable.

Referral loop: who the customer brings

Referred customers arrive with trust already earned. Customers who get value tell customers like them. A referral program makes that word of mouth measurable. It runs on four parts.

• Tracked invites.
• A reward sized to pay back inside the payback window.
• Referred customers tracked as their own cohort. They usually retain differently.
• Customer stories that turn one result into word of mouth at scale.

It’s measured by referral rate and referred-customer share of sign-ups.

Written cadence: the clock

The cadence is the clock the loops run on. Launches go out in three tiers, each with one brief, named owners and fixed lead times. Business reviews run monthly. Operating plans get written for the year and again at mid-year.

Attribution

1. One event stream from first touch to revenue, joined to the customer record.
2. CAC by channel on both blended and paid bases, refreshed monthly.
3. Incrementality over last-click. Holdouts and geo tests decide budget.
4. Revenue by acquisition cohort. Lifetime value gets measured per channel instead of averaged.
5. Payback by channel, checked in the monthly business review. The threshold is set before the spend.
6. At scale, a marketing mix model. It allocates across what tracking can’t see, like word of mouth, PR and brand.

Attribution models describe. Experiments decide.

The correctness bottleneck

Generative tools collapsed the cost of a first draft. They did nothing to the cost of a wrong claim.

In payments, one wrong line about fees, licensing or settlement times is a compliance event, a support ticket and a lost merchant, all at once.

So the scarce resource moved. Throughput is solved. Correctness at throughput is the constraint. Every design choice below follows from it.

When output gets cheap, design for the error, not the draft.

Pointer, never copy

AI-assisted marketing fails by default through fact drift. Twenty prompts each carry their own copy of the pricing, the positioning and the product status. Each copy is a little more out of date than the last.

Canon fixes drift by structure. About 22 documents hold everything the team asserts, from audience and positioning to voice, brand, product taxonomy and policy. Every fact has one home, a named owner and a last-validated date. Every skill and every draft reads canon at generation time. None stores its own copy.

Two details do most of the work.

• Status flags decide what’s sellable. Every product carries a status. Roadmap items are visible to the team and blocked from campaigns. Beta products can only be sold with beta framing.
• A drift sentinel scans the workspace for canon copied outside canon.

If you can’t name an owner for a fact, it isn’t canon yet.

Same-shape workspaces

Same shape means a person or an agent can move between functions without relearning where anything lives. There are five workspaces, one per function. They cover acquisition, lifecycle, ops and analytics, brand production and executive comms. Each holds instructions, context, skills, workflows, outputs and a gaps log.

Defects travel upstream. Fixes travel down. Every workspace loads canon before it does anything else, and none can write to it. A workspace that finds a missing or wrong fact logs a gap for the owner. It doesn’t patch the fact locally.

Friction where it pays

Gates add friction back exactly where a mistake is expensive. The rest of the system exists to remove it.

• Brand review on everything that ships.
• Compliance precheck returns clear, edit or hold.
• Claims as code. A style guide is a request. A lint rule is a guarantee. Regulated claims live as data. The build fails when a page asserts something the data doesn’t allow.
• Website QA in three tiers. Automated checks fail CI. Agents audit. A human signs off.

The proof ladder

Every claim in a draft gets one of four dispositions.

1. Verify: source it to canon or a public document.
2. Soften: scope it to what is provably true.
3. Cut: remove it.
4. Escalate: the fact’s owner decides.

No speculative numbers. No invented quotes. The team’s first run of the ladder resolved six borderline claims in about ten minutes.

Earned autonomy

Every agent loop starts by proposing. It gets promoted only after a human reviews real sample output and signs off. Each loop runs observe, decide, act, verify, at an explicit autonomy level. Spending money and publishing never go autonomous. Paid media is the hard line.

Evaluation comes before autonomy. Every skill runs against golden briefs and a claim-accuracy check. That lets a promotion rest on evidence instead of a good week.

Agents propose. People approve.