How it works
A daily brief with 3 to 5 recommendations. You decide which to take. 30, 60, 90 days later we tell you whether the metric actually moved, with citations to the records that prove it.
HubSpot OAuth: contacts, companies, deals, emails, meetings, calls, deal-stage history. Optional: Google Ads, GA4, Search Console, Stripe.
Every contact gets a source class (outbound, paid, organic, referral, conversation) and an attribution confidence. The classifier owns up to OFFLINE traffic instead of guessing.
Daily brief: 3 to 5 ranked recommendations with citations to specific deals or segments. Plus tactical claims: at-risk deals, champion signals, stage drift, unowned high-fit accounts.
Each recommendation has accept / dismiss / snooze on a single click. A pipeline snapshot is taken the moment you decide. Deep-links open the source record directly in HubSpot.
30, 60, 90 days later, an outcome cron re-reads the same metrics and writes a verdict. Worked / didn't work / neutral / inconclusive. The math is pre-registered, identical across customers.
Brief reads the ledger before it writes. Claims you already actioned do not get repeated. Past wins get cited as evidence. A claim whose verdict came back didn't_work gets suppressed by a deterministic ranker before the model sees it.
Concrete deliverables on a fixed cadence. No vague promise of insights.
The first brief lands within 48 hours of HubSpot connect. Alongside it: a replay of the last 12 months showing what Recon would have flagged, when, and what actually happened. The proof artifact ships before any verdicts are due.
The first cohort of accepted recommendations comes back with a verdict. Each one cites the specific deal or segment whose metric moved, the baseline it moved from, and the directional change vs the workspace's prior drift.
Cumulative correlated pipeline movement. Take rate. Worked / didn't work split. Inconclusive rate (we expect 25 to 40 percent and disclose it). The CFO appendix: every row clicks through to the brief that wrote it.
The math is identical across customers. No per-account dials.
| Verdict | Definition |
|---|---|
| worked | Primary metric improved by 10 percent or more, and the change is at least 1.5x the segment's prior 30-day natural drift. |
| didn't_work | Primary metric degraded by 10 percent or more, and drift is at least 1.5x prior. |
| neutral | Metric moved less than 10 percent in either direction. |
| inconclusive | Sample size below the floor (n less than 5 closed events for win-rate, n less than 3 deals for ARR-movement), or underlying records changed materially between t=0 and t=30d. |
Inconclusive is a feature, not a failure. On a mature pipeline we expect 25 to 40 percent of verdicts to land inconclusive. The track record page shows that rate prominently. If a vendor shows you 95 percent "worked" verdicts, ask them to define their thresholds.
No holdout. V1 verdicts are pre/post against a frozen baseline with a natural-drift coefficient as the guard against false positives. We do not run a randomised holdout against your live pipeline. A causation-defensible holdout tier is on the roadmap.
Where the brief says +£68,000 correlated movement, that means the metric directionally moved in the window after acceptance, correlated with the claim. Not caused by it. The buyer infers causation; Recon never asserts it.
Recon ingests email body text and classifies sentiment via Claude Haiku. That is a real trust commitment, so the posture is documented in full.
Regulated EU buyers (financial services, health, public sector): an enterprise security review has not yet been completed. Get in touch before signing. Honest about where the bar is not yet.
Full sub-processor list: /sub-processors. Privacy policy: /privacy.
From the desk of Recon AI
Connect HubSpot. First brief lands within 48 hours, every recommendation cited to your real deals. Verdicts get tracked in the ledger. The model sharpens on your data, not someone else's.