Writing an AI business case that survives the board
"AI will save us loads of time" will not get past a decent finance director. A business case for a Business Central agent needs volume, current cost, expected accuracy, credit cost, risk, and an answer to "what if the price doubles?". The structure, the numbers, and a worked example.
Most AI business cases fail at the board for the same reason: they're written as enthusiasm, not as a case. "It will free up the team", "it's the future", "everyone else is doing it". A finance director will ask three questions: how much, how sure, and what if it goes wrong. Answer those and it'll get approved.
The seven numbers
Volume. How many transactions a month (supplier invoices, sales orders, expense claims). From the system, not a guess.
The routine share. What proportion are routine enough for an agent: known supplier, PO-backed, within tolerance. Sort a real month. (See what the Payables Agent actually does.)
Current time per transaction. Time a sample. Keying, matching, coding, chasing.
Expected time with the agent. For routine items it's the review time, not zero. Assume a minute to review, not a second.
Agent cost. Volume multiplied by credits per transaction at today's rate. (See the real cost of AI agents.)
Set-up cost. Configuration, permissions, supplier defaults, clean-up of master data, training, a pilot month.
The error cost. What a mistake costs (a duplicate payment, a wrongly coded invoice), times the expected error rate after review.
The structure
The problem in the business's terms: "the AP team spends 60% of its week keying invoices, and month-end accruals are late because of it".
The proposal: one agent, one process, a pilot, with a person approving every posting.
The numbers above, with sources.
What the time is used for. Redeploying people to supplier reconciliation, disputes and early payment discounts usually beats "headcount reduction", and is more honest about the risk.
Risks and controls: accuracy, fraud, outage, cost increases, data protection, and what each control is.
Sensitivity: the case if the credit price doubles, if accuracy is 20% lower than hoped, if volume halves.
The decision gate: after the pilot, continue, extend or stop, based on agreed numbers.
A worked example
A distributor receives 1,200 supplier invoices a month. A sample month shows 75% are routine. Keying and matching takes about four minutes each. With the agent, routine invoices take one minute to review; the rest still take four minutes but arrive pre-keyed, say three.
Time today: 1,200 × 4 min = 80 hours a month.
Time with agent: 900 × 1 + 300 × 3 = 1,800 min = 30 hours. Saving: 50 hours a month.
At a loaded cost of £25 an hour, that's £1,250 a month.
Agent cost: about 65 credits a typical invoice, so roughly 78,000 credits a month. At $0.01 a credit that's around $780, or about £600 at current rates.
Net benefit: around £650 a month, plus faster month-end and fewer late-payment penalties.
If the credit price doubles: roughly break-even on time alone, still positive on the softer benefits.
That's a case a board can make a decision on. It may also be one they decide to wait on, which is a perfectly good outcome; an honest case that says "not yet" is worth more than an optimistic one that goes wrong.
All figures are illustrative; use your own volumes and Microsoft's current pricing. If you'd like help building the case for your business, book a 30-minute call.
Email this post as a PDF
A branded PDF of this post, sent to your inbox to keep, print or share. Copy in colleagues if you like.
We only use these addresses to send this PDF. Nobody is added to a mailing list. Mike at Business Boxed receives a copy. See our privacy notice.
