Each one started as a personal annoyance and ended as a repeatable system. Some live, some portfolio.
Automations fail silently — you only notice when something downstream breaks days later. So I built two checks on top of the work: a daily heartbeat that confirms every job actually ran, and a weekly pass that cross-checks whether my own notes, plans and status still agree. Blind spots surface as alerts, not nasty surprises.
Read the build note →Letting an AI build and change things unsupervised is how quiet mistakes reach production. So I borrowed the control every finance team already trusts — four-eyes review — and wired it into the machine: every piece of automated work runs plan → build → an independent fresh-eyes audit, and nothing ships without a human approving it.
Read the build note →A newly merged, private-equity-backed group with no shared system and board reporting stitched together by hand each month. Over a year I designed the data model, automated the reporting, and stood up the first standard workflows — turning thirteen-step manual processes into one. The lesson: AI didn't win on model quality, but on ownership, process and clean data.
Ask me about it →Map the manual steps and lost context before writing a line of code.
One automation that removes one step, wired into existing tools, documented to hand over.
Schedules and alerts that only fire on failure. Silence means healthy.
I let an AI system build things for me most days. What lets me sleep on that isn't the model — models sound confident whether they're right or wrong. It's that I wrapped it in two controls any finance team would recognise on sight. One: the thing that builds never approves its own work — maker and checker are never the same actor. Two: assume most failures are quiet, so the system checks its own health and is built to make a sound when something breaks.
A system that checks its own health: a daily heartbeat that confirms every job ran, and a weekly pass that catches when its own records drift apart — plus the day it found its own blind spot.
The hours didn't go on judgment — they went on work that shouldn't have been manual. The gap in audit isn't technology. It's execution.
Nine years ago I took a career break to study data science. What actually prepared me for AI in finance wasn't AI — it was data literacy.
An AI agent — not an RPA bot — processed a backlog of scanned contracts across eight languages overnight. The difference wasn't speed. It was judgment.