The highest-leverage rate in your business is being run on a usage dashboard and an expiry email.

Who this is for

Post-seed, pre-CMO, product-led SaaS with a live trial motion. Concretely: customers on the books, the product works, roughly $30–150K MRR, seed capital about to go into acquisition for the first time, a Series A nine to fifteen months out — and nobody in seat yet whose job is the funnel. If you have already hired a VP Marketing, this is a system they should inherit rather than one I should install; talk to me anyway and I will tell you honestly which.

It is not for pre-revenue products (nothing to instrument yet) or for funnels without a trial or freemium step.

What it is — the Sensor / Brain / Executor pattern

Your product is a Sensor. It emits raw, append-only events with server-side timestamps, plus the few facts only it knows: subscription status, trial expiry, which channels are connected. It computes nothing.

Kyber runs the Brain. Every metric, the activation score, the conversion matrix, the orchestration decisions and the measurement all run on Kyber’s infrastructure. Every threshold change and every new intervention is a data operation on my side; your team is never in the loop after initial instrumentation.

Your product is also the Executor. It exposes one authenticated endpoint. The Brain composes every message and calls it; your infrastructure delivers under your own identity. No SMTP keys, bot tokens or OAuth grants cross the boundary in either direction, you can rate-limit it, and you can audit every message it ever sent.

The system is self-contained. It depends on no other analytics platform, and if you run a broader marketing-intelligence layer it can feed one through a single, optional export.

The two things a usage dashboard cannot see

Whether the product worked for them. A user can hit seven of eight activation milestones and still not have achieved the thing they signed up to do. On your dashboard they are green. In reality they are your highest-risk account: they have used the product enough to judge it, and their private verdict is “not yet”. They either decline to convert while looking healthy, or convert on momentum and churn in month two after CAC is spent. So the system asks — once at signup (“what do you want to accomplish in the next 30 days?”) and once at 70% of the trial (“you said you wanted this; where are you?”, one tap). That answer is the second axis.

Who is about to buy cheaply. The mirror image: a light user whose goal succeeded. Every score-only system marks them at risk and sends them activation checklists — the fastest way to irritate a qualified buyer who got what they wanted. The matrix converts them on the outcome and expands later.

Cross the two axes and every trial lands in exactly one of 21 cells, each with a pre-decided response and no fall-through. Only one cell — heavy usage, failed outcome — routes to a human, and that exclusivity is what keeps the human queue small enough to serve consistently.

What changes, week by week

Weeks 1–2 — Discovery and design. I send a nine-item request (product flow, stack, volumes, every message currently reaching trial users, honest human capacity). We make twelve decisions together: trial parameters, the first-value moment, the eight activation milestones, the goal question, the escalation gate, channels, orchestration defaults, measurement mode, the message library’s voice, the calibration date, the operating model, and whether to switch on the integration export. Every decision is written down.

Weeks 3–4 — Build. Your engineers instrument the events, add the daily export, backfill 30 days of history and expose the send endpoint: four to seven developer-days in total. In parallel I configure the Brain against your backfill and draft the message library — two variants per message from day one, so A/B testing starts with the first send.

Week 5 — Shadow mode. The full system runs for one to two weeks with delivery switched off. Every message it would have sent is logged with its reason. We review the log together: matrix distribution, volume per account under the caps, false-positive flags. Nothing reaches a user until that review passes.

Week 6 onward — Staged activation. Conversion-critical messages first week, risk interventions second, everything else third. Your existing drip sends are merged or retired in the same window so no user is ever double-messaged. Week-four checkpoint against indicators we committed to in writing before launch; recalibration at day 60–90, then quarterly.

What you get

• +The data contract — three tables and one endpoint, every field justified by the metric that consumes it and the action that metric drives. No consumer, no capture.
• +The metric set — sixteen metrics across five lifecycle stages, each with three exhaustive bands and a defined action; time-to-value estimated properly (Kaplan-Meier, so the accounts that haven’t activated yet are counted, not hidden).
• +The activation score and the matrix — eight independent, verifiable milestones; the 21-cell total function with its seven intervention classes; the escalation gate configured to your real capacity.
• +The orchestration layer — at most three non-critical messages per person per week, no repeats inside 72 hours, one message per underlying cause, silence on any account with a human conversation open, and a log of every message the system decided not to send.
• +The message library — every intervention class × goal-aware and generic variants × channel, in your voice, loaded with variant IDs.
• +Measurement you can put in a data room — intervention-level A/B on every send, pre-registered leading indicators and failure conditions signed before go-live, a framework holdout that activates only when your volume makes it honest, and a weekly report that shows every decision the system made.
• +The standing loop — weekly copy review, monthly cohort report, quarterly recalibration of the milestones against actual purchases.

Two ways to run it

Either way, when the system is running, your first CMO inherits an engine, not a rebuild.

How we know whether it worked

Before-and-after comparison is not accepted as a lift claim — the acquisition mix changes at the same moment the system launches, and any delta is unattributable. What is measured instead: every automated message carries a variant and its outcome joins back on a correlation ID, so copy quality compounds from the first week; activation rate, time-to-first-value, check-in response and matrix distribution are pre-registered with thresholds; a randomised holdout activates automatically when the volume forecast clears the threshold for the split you can live with; and below that, Bayesian grading gives honest graded evidence instead of a p-value that cannot be reached. The distinction is written down before launch, because after launch everyone’s incentives favour forgetting it.

The evidence this is built on

No one has published a causal study of a system like this end to end, and the page will not pretend otherwise. The components rest on peer-reviewed work: the only large randomised SaaS trial experiment (337,724 users) found that treating trial users differently on observed behaviour beat the best uniform policy on subscriptions, subscription length and revenue together; field experiments across sectors show that targeting retention effort by responsiveness beats targeting by risk score, and that unsolicited “proactive” interventions can raise churn rather than lower it; and trial-acquired customers have been shown to carry materially lower lifetime value than other customers while being more responsive to usage and communication — the gap this system is built to close. The white paper grades every source it uses by provenance; vendor benchmarks are shown for contrast, never as targets.

Questions founders ask