Things we found that changed a decision

Platform build

173 tables · 128 routesbuilt in 15 months

We built the platform a four-department revenue org now runs on

It started as an ROI calculator for the sales team at a logistics SaaS company. Fifteen months later it was the system sales, account management, marketing and finance work in every day. Every external system mirrored into Postgres, one screen per deal instead of six tools, and an AI layer where every one of 157K+ annual model calls is costed.

• No third-party API on any read path. Nothing waits on a vendor when a page loads.
• One privileged credential, in one place. Writes go through a field allowlist.
• Renewal dates ranked by how sure we are of them, so leadership can see how much of the book is fact and how much is guess.

CRM cleanup

$13.78M → $11.21Mcorrected booked revenue

The booked revenue that was counted twice

An onboarding pipeline recorded the delivery of a sale, not the sale. 61 of its 67 wins duplicated a sales deal. Because onboarding lags sales by about a year, the problem grew every year: 0% in 2023 and 2024, 24.7% in 2025, 35.7% in 2026. Fixing it moved like-for-like growth from +16.5% to −20.9%. We had already presented the +16.5%.

• The rule we wrote afterward: always ask what's underneath a correction that improves the number.
• The dated correction went at the top of the page people had already quoted.

AI automation

78% of AI spendre-scoped to three bands

The flagship AI deal score was worse than a coin flip

The AI lead score ate most of the AI budget across three prompts. Back-tested on 1,171 closed deals it looked fine, AUC 0.642. Limited to deals that reached a demo, the ones reps actually triage, it dropped to 0.492. It had been detecting whether a deal moved at all, not whether it would close. What survived was coarse triage: the top band won 20.8% against a 13.8% baseline, the bottom band 3.9%.

• Re-scoped from ten tiers to three bands, with the caveat inside the API payload so nobody can read the number without it.
• We recommended against spending more to fix it.

RevOps analytics

24.7% → 9.1%real new-business win rate

The win-rate analysis nobody could reproduce

A 28-page win/loss deck, still being quoted seven months on, reported a 24.7% win rate on 174 deals that didn't exist in any reading of the CRM. It had counted expansion, renewal and onboarding wins against new-business losses. Rebuilt from 1,304 closed deals down to 1,006 real new-business deals, the number was 9.1%.

• The useful output wasn't the number. It was a gate that refuses to cut a rate by any field that's filled in differently for wins and losses. Nine fields got refused.
• Subscription ARR was filled on 96.2% of won deals and 15.3% of lost ones. The original deck's whole velocity section rested on a field like that.

RevOps analytics

763 company-yearsback-tested

The churn signal every CRM encodes, that had no signal

A back-test over 763 company-years and 124 churn events showed engagement recency, worth 30 points in the renewal-risk model, predicted nothing. Churning accounts had been contacted more recently, not less. We rebuilt it around the four signals that held up, with coverage and confidence stated on every response.

• Account tier, ARR and ticket counts were each skewed by reverse causality or mapping bias.

AI automation

56% of spendin two workflows, re-tiered

Treating model spend like cost of goods

Ten weeks auditing 157,000+ model calls a year. 99% of the volume cost under 1% of the money. Two workflows carried 56% of spend and moved to smaller models with no quality drop we could measure. Unit costs came straight off the ledger: $0.018 per task audit, $0.23 per partner classification, $0.32 per deal analysis.

• Rate limits, concurrency caps and kill switches on autonomous spenders came out of the one loop that nearly took the database down.

RevOps

83%annual spend cut

Cut one platform's spend 83% and got five times the coverage

Ran selection and negotiation across the sales tooling stack. Same period: rewrote lifecycle stages with enforced criteria, rebuilt lead scoring with ICP, behavior and decay, and built partner attribution across six ERP ecosystems so partnership decisions stopped running on anecdote.