The first real task kicked off with the enterprise client today. I can't say much about who they are or exactly what we're doing — that's the deal with enterprise work. But the shape of the problem is worth writing about, because it's the kind of problem fractional AI ops is actually for.
Their CFO spends twenty hours a month reconciling across multiple accounting systems in multiple jurisdictions. Different software that doesn't talk cleanly to each other. Different tax regimes. Different chart-of-accounts conventions. Different reporting cadences. By the time the numbers agree, a full working week is gone every month, every month, forever.
Twenty hours at CFO rates is not a small line item. Twenty times twelve is two hundred and forty hours a year of senior finance time spent producing nothing new. Just the same answer, recomputed, over and over.
This is the exact shape of problem I want on my plate.
The Instinct to Overbuild
My first instinct, is to go with something light in my own environment, massage and mix the data, and then simply port it to their environment. With so many unknowns, that's the closest I think of solving this simply. What we don't want to do is get caught in the trap of thinking, let's solve this buy building something new. I've written about this exact trap enough times in the past week that I recognised the smell before it got anywhere.
The right move is not to build the big thing. The right move is to build the lite thing — the thinnest possible reconciliation that handles the 80% case. A script that pulls from each source, normalises the fields that matter, flags the deltas that don't zero out. Nothing fancy. Nothing clever. Just the smallest thing that makes the CFO's Monday morning shorter than it was the week before, so she can actually have time to drop her children off at school, return home for a quiet cup of tea before she starts her week. If I can give her just a bit of that feeling, it's what I want.
Codify, Close the Loop, Verify
Once the lite version runs and produces something useful, the discipline is what you do next.
First, codify it as a skill. Not a one-off script hidden on my machine. A markdown skill with the logic, the field mappings, the known edge cases, the jurisdiction quirks. Written so the next person to run it — or the next version of me three months from now — doesn't have to re-learn what I already figured out. That's the fat-skills rule from earlier this week, applied to a client engagement instead of my own stack. I'm still figuring how we can do that in their environment but that's a problem for a different task. I know that we'll get there.
Second, close the loop. Every time a reconciliation line fails, the failure gets logged with a reason. That reason becomes a new rule in the skill. The next run catches it. The time from "discovered edge case" to "handled automatically" collapses from a quarter to a day. The system gets smarter with every month it runs, not staler.
Third, verify when the source systems change. Accounting software updates. Jurisdictions change reporting rules. Schemas drift. If the reconciliation isn't re-checked against each source every cycle, it rots quietly and you only notice when the auditors show up. So the skill has a verification step that fires when a source version bumps or a column header moves — it flags what to update before the next run drifts out of sync.
Build lite. Codify as a skill. Close the loop on rejections. Verify it stays in sync with the sources. Four moves, repeated every engagement. This is a new daily skill that I'm rotating in with my other ones, including the one task at a time, even if I have three queries going on at once. They run as a single task, work through to completion and the loop gets closed in the same way. Codified everytime to keep the systen in sync.
Why This One Matters
A lot of AI ops work looks impressive on a slide and never actually saves anyone twenty hours of anything. The test I use — and I'm going to keep using it — is whether a real human is currently spending a real, painful block of time on the thing. Twenty hours a month of CFO time is that block. The first CFO who saves fifteen of those hours doesn't need a pitch from me for the next engagement. At this point, I don't know if I can do it, but I'm sure going to try and even if it takes me all month, the 20 + 20 + 20... seems like it'll definitely be worth it.
That's the flywheel. One boring, well-built, well-codified, well-verified reconciliation.
Revenue: $0. Clients: 2. Prospects: 5.
Day 21 of 365.