The situation
An enterprise B2B SaaS business running go-to-market across the Americas, EMEA and APAC, building an account-scoring model on a modern stack: CRM as revenue truth, marketing automation as behavioural truth, an enrichment and orchestration layer computing across both, plus paid, a site router, a sequencer and conversation intelligence.
The audience and signal layers were funded and under construction. Three things were not: what a score has to cross to become qualified and who agreed that line, what happens the moment it crosses, and how a rep tells the model it was wrong.
That split is the whole diagnosis. The first two layers are tooling problems and money solves them. The last three are agreement problems and no tool solves them. Most scoring programmes fail in the second group while budget continues to flow into the first.
The architectural decision
The score lives on one object, computed once, mirrored read-only everywhere else.
Account-level scoring belongs on an account record. The sequencer triggers natively on field changes there, with no middleware. The site router reads ownership and record fields from the same place. Computing a second copy inside the marketing platform creates a number that drifts from the first, and then nobody can say which is right.
Everything else in the design follows from that one field existing in exactly one place.
Score to play, with nobody in the middle
The orchestration layer computes and writes the field. The sequencer fires on the field change and enrols. The tier syncs into a company-matched paid audience. The site router reads the same field and puts a scored visitor in front of a person. A computed artefact — a diagnosis assembled for that account — attaches to both the sequence and the ad, so the rep and the advertising say the same thing. Conversation intelligence and the rep’s reject reason return to the model.
The design rule we enforced: if a human has to look at a dashboard for the play to start, the play will not start. A prioritised list that still requires manual review captures a fraction of its value.
The model
Two numbers, not one. Fit is who and is stable. Likelihood is when and decays weekly. Collapsing them into a single score destroys both, because a perfect-fit account that is dormant and a poor-fit account that is active become indistinguishable.
Decay tuned to the sales cycle. Signal half-life set at roughly a quarter to a third of the median cycle, decaying exponentially, with recency carrying a substantial share of composite weight in a long enterprise cycle.
Co-occurrence, not counts. Funding plus frontline hiring plus a leadership change in the relevant function, inside the same thirty days, is a read. Any one of them alone is noise. Counting events rewards noisy accounts.
Recalibrated quarterly against closed-won. Static models rot, quietly, and the first symptom is reps ignoring them.
The feedback layer
Four reject reasons, and only four: not our ICP, right company wrong person, bad timing, already engaged. Each maps to a specific correction — firmographic weighting, contact targeting, decay length, suppression. Four, because a rep will use four and will not use nine.
This is the cheapest layer to build and the one most often skipped. It is also where the return lives: the published evidence on scoring programmes consistently shows the model being right and the value appearing only where reps actually acted on it. Adoption is the multiplier on everything upstream.
The choreography
Three tiers, where what changes by tier is computed versus written, not effort. One-to-one for a small set of named accounts, bespoke and human-written. One-to-few for clusters, modular copy plus the computed number. One-to-many at volume, computed artefact and peer benchmark, nothing hand-written.
For the top tier we wrote the cadence as a dated sequence running from ten days before the first dial to three weeks after: audience synced and suppression checked, air cover live at industry level, the computed artefact landing with the rep, first touch referencing that artefact, connection, peer benchmark, a site visit triggering proactive chat, multi-threading to the second and third persona, the call, and an outcome logged as accepted, rejected with a reason, or recycled with a date.
Paid runs before the rep deliberately. Buyers identify most of their shortlist months before they engage a seller, a far larger share of accounts show active buying signal at any moment than folklore suggests, and a large proportion of B2B purchase attempts are abandoned because the buying group could not agree. The rep is rarely the first touch and should not be designed as one. The biggest addressable problem is helping a buying group reach agreement, which is why the artefact is a diagnosis rather than a pitch.
What gets measured, in what order
Adoption before accuracy. Weeks one to four: sales acceptance rate against an agreed band, and whether reps are actually using the artefact — that is the read on whether marketing built something useful. Weeks four to eight: reject reasons by category, the first real signal on what is wrong. Weeks six to twelve: meeting rate for scored against unscored, with a holdout, for the first causal read. Months three to six: qualified-to-opportunity by score band, calibrated against closed-won.
Attribution goes last, on purpose.
The operating agreement
Because the engine crosses four teams that do not report to each other, we wrote down what each side owes rather than assuming goodwill. Scored list out the same day each week, with a reason per account rather than a score. One computed artefact per top-tier account. Air cover live seven to ten days before the dial. A rep flag acknowledged within one working day. Reject reasons reflected in every recalibration and reported back. A suppression path so marketing never contradicts a live conversation. Signal to action inside forty-eight hours, because reply rates on an acted-upon signal decay sharply over the first week and converge on the cold baseline after about three.
On credit economics, honestly
Enrichment platforms bill in two currencies — data credits for provider calls and actions for orchestration steps — and a waterfall across a large account list burns a great deal before any scoring runs. Four levers in order of size: gate expensive providers behind a cheap fit filter, use your own provider keys where the platform allows it, scope agent prompts tightly, and stamp every enriched record with source and timestamp so you can see what is actually being used.
This is the least well-documented part of the stack, so we recommended a pilot on a few hundred real accounts to measure burn rather than offering a confident estimate. Starting narrow is the right answer twice over.
Status
Delivered inside the last twelve months; the engine is in flight. We are not publishing outcome figures — the results are not in, and we do not publish numbers we cannot stand behind. The benchmarks cited above are public industry data, not the client’s. No client figures, account data or identifying details appear on this page.