Writing
How a bookkeeping firm finds a client's AI spend
From the general ledger, the same way it finds any other recurring vendor. AI tools bill on a cadence like everything else, and the ledger holds every payment however it was made. What makes AI spend different is where it lands: split across a company card, a reimbursed personal card, and an annual invoice paid by bill pay, so any one of those views shows a fraction of it.
Ask a client what they spend on AI and you will hear about the ChatGPT charge on the company card. The books can hold a different answer, and the difference is where the tools were bought rather than what they cost.
Where does AI spend show up in a client's books?
In three places at once. A team plan on the company card, which is the only one a card tool sees. A personal-card subscription that comes back through an expense report and posts as a reimbursement to a person rather than as a payment to the vendor. And an annual invoice for the seats a team standardised on, approved like any other contract and paid by bill pay or ACH. The pattern has a name: shadow AI.
The sample ledger on the demo page has two of the three: $150 a month for Anthropic Claude Team on the company Visa, and a $3,000 annual invoice for OpenAI ChatGPT paid by bill pay, which the engine reports at $250 a month so it can be added to everything else.
Why does a card tool undercount it?
Because it can only see its own card. In the sample that is $150 of $400 a month of AI spend, about 38 percent, and the larger charge is the one that never touches a card. This is the ordinary card-blind pattern, and AI tools are unusually prone to it: individual seats get bought on whatever card is closest, and the moment a team standardises, the vendor moves the account to an annual invoice.
What does the audit report about AI vendors?
Every AI vendor the engine recognises is grouped under one category, so a firm can total what a client spends on AI across every rail and put that figure in the report. The catalog names the chat assistants, the image and video generators, the coding tools and the model APIs, and matches them the way it matches any vendor: from the descriptor in the ledger, by rule, with the charge history shown for every match. A language model is used only to name a descriptor the rules could not, and its matches are held for review rather than trusted.
The category is deliberately broad, which is why it is excluded from duplicate detection. A chat assistant, an image generator and a transcription API are not substitutes, and a report that flagged three AI tools as redundant would put exactly the wrong figure in front of a client. Renewals, price increases and charges that have run unchanged for months are reported for AI vendors like any other. In the sample report the first finding on the cover is a Claude renewal nineteen days out.
What can the ledger not tell you about AI tools?
Who is using them, or whether anyone is. A ledger records what was paid, not who logged in, so the audit does not report seat usage for AI tools any more than for anything else. It can show that a team plan and an annual invoice for a second assistant are both being paid, which is the conversation worth having. Whether the seats are idle is a question for the client.
How does a firm turn this into an advisory line?
By putting one number and one list in front of the client: what they spend on AI a month across every rail, and which of those charges renews in the next ninety days. Few clients have seen the first figure, because nobody adds up a card charge, an expense report and an invoice. The renewal list is where the decision sits, since an annual invoice is where a tier change or a seat count nobody revisited gets paid for another year.
The audit runs from a CSV export, with no client login required, and the report carries the firm's own name. See how firms run it across a client book.