CodelessOps · Track record
Finance automation I've built.
Fifteen years in month-end and FP&A, then eight taking the manual work out of it. Here is what I've built, the numbers it produced, and how I go about it. One process at a time, every figure computed by code, a person signs.
01 · What I've built
Six pieces of manual finance work, taken out.
Month-end close, integrated
NetSuite actuals, HR data and timesheets loaded into planning without re-keying. Selected months refreshed, closed periods untouched, every load reconciled to source.
Exscientia (biotech), 2022 to 2025. A build inside one company.
Close from 12 days to under 8 hours. Handed to the finance team to run.
Variance reporting and commentary
Every figure from the ledger by script. Narrative drafted only for the lines that moved, each citing its source. A reviewer gate before anything is sent.
Published as the audited close kit for Claude; the same pattern inside Grounded for board-pack questions. You can run it.
81 of 81 on a public benchmark I published; the raw models scored 66 and 71.
Forecasting that doesn't start from an export
Driver-based reforecasting with actuals loading themselves and run rates with the one-offs already named, so the team starts at the judgement layer.
Exscientia; Goldman Sachs (driver-based FP&A); Astellas (budgeting, forecasting and consolidation across EMEA, APAC and the US). Builds inside those companies.
Reforecast cycles that run monthly without a rebuild.
Cash flow forecasting
Cash-forecast adjustments and consolidation built into the planning platform, so the weekly view rolls without someone rebuilding it. Judgement stays with the person who signs it; the assembly doesn't.
Exscientia. A build inside one company.
Part of the same close that went from 12 days to under 8 hours.
Invoice processing
Invoices read, coded and queued for approval as they land. The approval stays with a person.
A build inside one company.
60-plus hours a month to about two.
Contract review with AI, under controls
Rules could not read the contracts, so AI extraction. The model never computes a figure; a person validates every record before it is used; calculations stay in code.
Fox (media), 2024 to 2026. A build inside one company.
Hundreds of contracts per run, review hours removed, no unchecked number reaches a report.
02 · The way I think about it
Most of these are pieces of one idea.
The FP&A side of the close run as small nightly processes, with people judging weekly and month end kept for the things that need a signature. I wrote it up in three parts, with the unknowns stated on every page.
03 · How I go about it
Document first. Sort like an accountant. Build what survives.
I document the process with the people who run it first. Then I sort what I found the way an accountant sorts it: hours removed, error risk, who owns it. Rule-based and high-volume gets automated; judgement stays with a person; messy gets fixed before anything is built.
Each item goes live with a check, a named owner and a one-page runbook, or it isn't live. AI goes where rules run out, never where the arithmetic is.
04 · The honest line
Not everything here is a product.
Some of it is a build I did inside one company and can describe but not show. Some of it is published and you can run it. Each entry above says which.
05 · Next step
If one of these is the thing eating your month.
Twenty minutes is usually enough to say whether it's worth doing and what I'd do first.