Finance on AI employees.
One seat. The numbers desk pulls Stripe, NMI and the ad accounts, puts revenue against spend for the window, stops a send when it crosses the line it was given, and flags a campaign that does the same.
Finance is the department most $5-50M founders have delegated and then re-grabbed. A bookkeeper runs the books. A fractional CFO shows up quarterly. The daily number is still something the founder calculates in their head, and the daily number is the one that decides what gets built this week. The numbers desk fixes that.
What the numbers desk does
Pulls the sources.
Stripe. NMI. The Meta ad account. The Google ad account. The TikTok ad account if it is on. The affiliate ledger if there is one. Each source is a known API with a known schema; the agent reads them on a schedule.
Puts revenue against spend.
By window: yesterday, this week, this month, this quarter. By source: cold ads, retargeting, email, affiliates, direct. The output is one table the owner can read in thirty seconds.
Stops a send when it crosses the line.
The owner set the kill bars. The numbers desk watches them. A campaign at twice the kill cost per lead triggers a pause recommendation and, if the rule allows, an actual pause. A broadcast that triggers a spam complaint rate over threshold triggers a hold with a decision brief.
Flags the campaign that is drifting.
Not every change is worth paging the owner. The desk uses a baseline: the moving average of the last fourteen days. A one-day swing is noise; a three-day trend above or below the band is a signal. The agent writes the signal into the morning briefing with the number attached.
What the morning briefing looks like
Three paragraphs. One number per paragraph. First paragraph: yesterday’s revenue, by source, with the delta against the baseline. Second paragraph: yesterday’s spend, by source, with the delta. Third paragraph: anything that crossed a line and anything that is on track to. No charts. No prose. Numbers.
What stays human
- Tax filings. A human accountant files, with the numbers desk feeding them clean data.
- Pricing changes. No agent raises or lowers a price. The agent models the impact; the owner decides.
- Chargebacks and disputes. A human handles the message to the processor, with the agent assembling the evidence package.
- The quarterly review. The CFO or the owner reads the full ledger, finds the thing the agent did not see because it was not in the baseline, and writes the one change for next quarter.
What this unlocks
Three things a finance department on an AI employee gives you that most $5-50M businesses do not have.
- A daily number. You know yesterday’s revenue by source by 7am every morning, without opening a dashboard.
- Live brakes. A broadcast that would have cost you a thousand dollars gets stopped at a hundred.
- A one-page quarterly. The CFO opens one table, not seven.
A real number
Live from Brad’s business: $326,700 in revenue (May 1 to Sept 29, 2026), on under $20,000 of ad spend, in a window where every job was done by an AI employee. The numbers desk is one of them.
“The job the numbers desk takes off your plate is the one where you look at Stripe at 11pm and feel the day on you. You get it back as a paragraph in the morning.”
Where to start
If finance is your lowest-scoring department, start with the morning briefing. One agent. One schedule. Three paragraphs. Everything else — the brakes, the drift flags, the quarterly roll-up — compounds from there.
Frequently asked questions
Does this replace my bookkeeper?
No. Your bookkeeper books. The numbers desk reports. The two feed each other: the agent flags any entry it cannot classify; the bookkeeper handles the exceptions.
Does it have access to my bank account?
Read-only, usually through the payment processor and the ad accounts, not the bank. No agent moves money in Brad’s rules.
What about forecasts?
The agent can produce a straight-line forecast from the trailing data; it does not make a strategic forecast. The CFO or the owner builds the strategic one, using the agent’s backdrop.
Can it issue refunds?
No. Refunds are a human decision. The agent assembles the facts and writes the suggested message; the human issues the refund in the processor.
Score your own company
Twenty-four questions, six departments, 0-24 score. The lowest-scoring department is the first seat to fill with an AI employee. Free, one email.
Send me the scorecard