I build the systems that find the revenue.

Make the data exist at all. Most revenue questions are unanswerable not because the analysis is hard but because nothing is recording the thing being asked about — the pipeline lives in one system, the fulfilment data in another, and the join between them is nobody’s job. I build the system of record and wire it to the systems around it: a business development CRM architected from scratch and integrated with the company ERP to carry pipeline, outreach status, opportunity scoring, and estimated ACV; a Postgres schema with row-level security behind an account portal; an outbound console consolidating four separate tools into one workflow.

The usual substitute is a spreadsheet that gets rebuilt every week, and it fails quietly — not by breaking, but by being three days stale in a way nobody can see. By the time the data is trustworthy enough to act on, the window it described has closed.

Make movement visible while it is still cheap to act on. That means a live view rather than a quarterly deck: channel-sourced ACV trending against target, account activation as it happens, and a standing intelligence program tracking twenty competitors and seven platform surfaces weekly, mapped to product line so a finding lands on someone’s desk with a decision attached to it.

A manual quarterly reporting cycle tells you what happened one quarter after you could have changed it. The failure mode is worse than lateness: a metric nobody watches between reviews is a metric that drifts for eleven weeks and then gets explained rather than fixed.

Find what is underperforming, and rank it by what it actually costs rather than by how visible it is. This is the stage that pays for the other four. A carrier and logistics spend analysis surfaced an annualised late-fee exposure nobody had aggregated because it arrived as line items across separate invoices. A profitability suite ranks a catalogue by margin contribution rather than by revenue, which reorders it substantially. A deal analyser screens acquisition candidates against fixed criteria so the weak ones are cheap to reject.

Ranking by revenue puts the loudest line at the top and buries the one quietly costing the most. Averages do the same thing to a heavy-tailed distribution: the mean describes the winners and says nothing about the long tail where the recoverable money usually is.

Prove why, with a method that survives being argued with. Dependence modelled explicitly with a Gaussian copula rather than assumed away, because correlated line items widen a forecast interval far more than independence implies. Bayesian updating on activation rates, because an account that activated one of two times is not a fifty per cent account. Regression for elasticity where the relationship is worth quantifying. Every output carries its interval and the conditions that would make it wrong.

A single-point estimate invites the reader to argue with the number instead of the decision. Showing the distribution moves the conversation to what to do about the downside — and it is also the honest reporting of what a model actually knows.

Ship the thing that closes the gap, rather than filing a request for it. A finding that ends in a recommendation is half a job. The recovery wizard turns the analysis into a guided flow an operator can actually run; the enablement portal does the account’s onboarding work for it instead of scheduling another check-in call; the prompt tooling puts the qualification framework where the conversation happens. Same week, against real accounts, while carrying a number.

The standard path is a ticket, a roadmap, and two quarters. The cost is not only the delay — it is that the person who understood the problem is not the person who eventually builds for it, and the specific thing that made it worth solving does not survive the handoff.