At the start of every month, it starts the same way. There are 5 different bank accounts, multiple currencies, and the manual downloading to transfer over bank files and credit card statements starts off the new month accounting cycle. All done manually, and done by the most senior accounting principal in the company, the CFO. Why? Because these are bank files, sensitive, and only she can handle them.
Eight CSVs, one for each entity, each named whatever the bank's export tool felt like calling it that week. The CFO would spend two hours renaming them every quarter to match what the accounting package needs. Eight hours a year, typing the same words in the same order.
According to the COO, this is clearly a task that any 16-year-old average maths student should be doing, but the CFO is the only one who can do it. Tasks she is way over qualified to be doing, but a new hire is out of the question because it's not needed. Just enough hours to drain the CFO's day and cause havoc and playing catch up the rest of the week. Not enough hours to justify a new body, training, or onboarding.
Working on this fix is one of two I shipped for the same enterprise client before lunch today. The other is a buy decision on Lightyear, a piece of accounting SaaS. Both get measured the same way, and most pitches for AI for KPI dashboards get this part wrong. Hours given back from a named human, on a specific kind of work. Not tasks completed. Not tokens consumed. Not agents running. Just hours saved on work that bored the human doing it. Not only bored, but work they shouldn't be doing because it can be automated by AI. No one in the company knows how to do this work though, and it's not an easy workflow to accomplish.
The twelve-agents tweet
Earlier this morning, before any of that, I was on X. Some guy with a header full of laser eyes posted: "Bro, I have 12 agents running 24/7 while I sleep, no one is prepared for what's coming."
What are the agents doing?
"Bro, don't get left behind. You only have two months."
The AI hype runs on this. Imply motion, get likes. The post never has to describe what the motion is. You can run a content business on it if you do not mind sounding the same as everyone else doing the same content business.
My youtube feed is smothered with it. Mo Gawdat, ex-Google X, on Diary of a CEO saying AI will erase professional jobs by 2027. His own startup runs three people doing what 350 developers used to do. Meta announced 8,000 layoffs on May 20 as an AI restructuring. A Louisville professor predicts 99% unemployment by 2030. The hype is loud. Some of it is happening. Engineers got walked out of Meta this week.
If this X poster does have something running, he cannot tell you what hour of human time he is taking off the calendar. Because if he could, that would be the post.
Build one, the rename script
Another task came in yesterday on a Wednesday call. Two hours of her quarter, every quarter, retyping a naming convention by hand. I asked when she had to do it next. End of June. I will have a way to automate this by mid next week. That will be the last time she has to spend two hours manually renaming hundreds of files in the wrong format. An hour or two of my time to write the script, then tracked on the hours-saved KPI line.
It saves her eight hours a year. It runs every Monday for the rest of the engagement. The build was included in the fixed monthly rate. No line-item bill, no scope-creep negotiation, no proposal round.
The math the client sees: my hourly does not matter, the script exists now, the CFO has eight hours a year back. They keep the script forever and I maintain it to make sure it runs. I maintain something they don't need to maintain.
Build two, the Lightyear evaluation
The accounting package they use is fine for almost everything. What it does not have is an AI PDF and image reader for receipts and invoices, and the mobile app is borderline unusable. Lightyear has both. The CFO currently spends roughly five hours a month coding invoices the bookkeeper has manually retyped out of PDFs. Five hours of a CFO who bills at roughly $100 to $150 an hour.
Lightyear's lowest tier costs about one hour of that CFO's time per month. The math is one line: pay one hour, get five hours back, integrate Lightyear with our existing automation scripts. Approved before the second coffee. The premise for the chat I have with the CFO is this:
Listen, the way I see it, you won't probably ever even login to this software. If they have a free trial (I need to check if they do) and the onboarding is easy, we buy it. They claim 100% accuracy on reading PDF invoices and 50% on hand-written invoices. Since we get no hand-written invoices, we are looking at 100% accuracy on PDF invoices. If the onboarding is setting up a dedicated email inbox for all invoices and pointing the software at this email, and it does the rest, we're done.
No sales call, no explanations, no caring about special reports or features. It's this one thing: an AI image reader that is 100% accurate on PDF invoices. It's something I don't need to build or rewrite, and it saves the CFO five hours per month. Additional time she gets back to do other more creative and we hope productive things, rather than fiddling around with plus and minus signs or known vendor GL codes.
Where the hours go
Both builds, custom script and SaaS swap, get reported to the client the same way. Hours back to a named human. The dashboard is one column wide, for now. This is an idea I am proposing to this client, but if they like it, it becomes part of the sales pitch and copy on next month's home page rewrite.
Where the hours land matters too. The CFO's job is to plan across multiple entities, decide which of two competing growth bets gets the cash this month, sit with the founders and tell them whether the headcount they want is fundable. We pull hours back so she can do that part of the job. Not so she can take a longer lunch. So she can do the work a fifteen-year-old maths student cannot do. The hours also come back so she can do a pre-school drop off with a little more peace, and enjoy a cup of tea a little longer, because the first 10 days of her month are not scrambling with manual fiddling of files, and the back half of the month is not spent playing catch up.
That's the theory at least. We go live with the first live run on June 1.
The other chap's eight hours a year they gained back? They are not the moonshot. They prove the measurement system is real. If the rename script saves eight hours and the dashboard records eight hours, then the Lightyear sixty-hours-a-year claim is credible too. You earn the right to make the big claim by being honest about the small one.
The prospect on Tuesday
A prospect asked me on Tuesday how I measure AI success. He had been told the answer was "tasks automated" or "uptime" or "tokens." I told him the answer is hours, on a named human, on a specific kind of work that bores that human, freeing them for the work only humans can do well.
He went quiet. I do not know if I sold him on it yet. I will know in a week.
The X guy and his twelve agents will be on the timeline tomorrow. Probably louder.
---
Monthly Revenues $11,800 | Clients 2 | Prospects (AI marketing employee live next week) | Team: Me + Jan (CTO)