Solutions
Start where you actually are.
Nobody wakes up needing a platform. They wake up with a must-win account and no way in, a champion who stopped replying, or a forecast they cannot defend on Monday. Pick the moment you are in, or the seat you sit in. See what Adrata reads, the move it recommends, and the model that produced it.
How the team is modelled
Same evidence. Five different altitudes.
A rep and a CRO are not looking at different data, they are asking different questions of it on different clocks. A reply matters to a seller this afternoon and means nothing to a board in March. Adrata classifies every recommendation by the layer it belongs to and the horizon it pays back on, so an answer arrives at the altitude of whoever asked.
| Layer | Typical titles | What they control | Judged on |
|---|---|---|---|
| Seller | Enterprise AE, Account Executive, SDR / BDR, Strategic Account Director | Next account, next person, next message, next call, next deal move. | Closed revenue |
| Manager | Sales Manager, Regional Director, Team Lead, First-line Manager | Coaching focus, rep behaviour, pipeline inspection, forecast risk. | Forecast accuracy |
| Leader | CRO, VP Sales, VP Revenue, General Manager | ICP, segment, territory, capacity, strategic bet, market learning. | Durable revenue growth |
| Revenue Operations | RevOps, Sales Operations, Deal Desk, Revenue Systems | Routing, hygiene, SLA, stage definition, attribution, workflow constraint. | Less process drag |
| Enablement | Enablement Lead, Sales Trainer, Onboarding, Programs | Playbook, skill gap, content, objection handling, onboarding path. | Faster ramp |
The models underneath
You should be able to argue with the recommendation.
A move you cannot interrogate is a move you will ignore the first time it is wrong. So each model below states the question it answers, what it actually computes, and what the naive version of it gets wrong.
Causal uplift, not correlation
Did the action cause the outcome, or did the good deals just get more attention?
Treatment effects are estimated with propensity weighting and doubly-robust estimators, adjusting for deal size, buyer-group size, whether the deal was intro-sourced, whether it was an existing customer, and known competitive pressure. Estimates that have not been adjusted for confounders are quarantined so they can never be presented as a cause.
The naive version. Win-rate-by-activity charts are the most confidently wrong artefact in revenue reporting. Reps spend more time on deals that were already going to close, so every activity looks like it works.
Break-in probability
Of every possible way into this account, which one actually opens?
Each candidate route is scored on seven components: relationship strength, how relevant the connector is to the target, freshness, consent safety, social cost, evidence quality, and expected lift. A calibrated model converts them into the probability that the route breaks in, with each component’s contribution visible.
The naive version. A fit score tells you the account is worth pursuing. It does not tell you the door. Most tools stop at the account and leave the hardest decision, who to touch first, to a guess.
Off-policy evaluation with a safety gate
Would a change to the recommendation logic actually have done better?
Every recommendation is logged with the probability it was selected, which makes it possible to score a proposed new policy against history using inverse-propensity, self-normalised, and doubly-robust estimators. A new policy is promoted only if its lower confidence bound beats the incumbent, with effective sample size and reward-hacking guards checked first.
The naive version. Most systems ship a scoring change because it demos well, then measure it by whether the quarter was good. That is not evidence, and it cannot be rolled back on evidence either.
Thompson Sampling on what to send
Which version of this actually earns replies, for this kind of buyer?
Message decisions (template, tone, call to action, send time, subject style, opening hook) are treated as arms with Beta priors and sampled rather than fixed. Where a workspace has thin data for a specific context, the estimate falls back through archetype and industry to a global prior instead of pretending to know.
The naive version. A/B tests need traffic most enterprise teams do not have, and they stop learning the moment you declare a winner. Buyers change; a frozen winner decays quietly.
Stable matching
Given limited attention, who should reach whom?
Assigning outreach across a buyer group is a matching problem with preferences and capacity on both sides. Adrata solves it with deferred acceptance (the Gale–Shapley algorithm), so the assignment is stable rather than greedy, and no two touches quietly compete for the same person.
The naive version. Assigning by rank order sends everyone to the same few names, burns the highest-value contacts first, and leaves the room unevenly covered.
Calibrated timing
When will this actually close, and how sure can you be?
Deal timing is modelled with competing risks: won, lost, and stalled are different fates, not one binary. The intervals come from conformal prediction, which produces ranges with a stated coverage guarantee rather than a single confident date.
The naive version. A close date is a rep’s guess rendered as a fact. A probability with no interval is the same guess with a decimal point.
Priors learned from your own wins
In your business, who actually turns out to be the decision-maker?
Closed-won history records which title was genuinely the economic buyer or champion. That becomes a per-workspace prior, smoothed so one deal cannot become a rule, weighted by how much history stands behind it, and isolated so a workspace only ever learns from itself. It carries its own provenance: "Economic Buyer in three of your last eight wins in Fintech".
The naive version. A generic methodology tells you an economic buyer exists. It cannot tell you that in your segment it is usually the VP of Finance, because it has never seen your wins.
The right to say nothing
Is there enough evidence here to justify a recommendation at all?
Learned components stay inert until they beat the simple heuristic on held-out data, and every model has a floor below which it abstains rather than guessing. On a new workspace the honest answer is often "not yet", and the system is built to say so instead of manufacturing confidence.
The naive version. A system that always has an answer is not confident, it is unfalsifiable. The first wrong high-confidence call is the one that ends the trust.
Bring a real account.
The fastest way to judge any of this is to point it at a deal you already know well and see whether it tells you something you did not.