AI on client books: who calculates the figures in the report?

AI can do parts of bookkeeping well, such as suggesting a category or reading a messy vendor descriptor. The question for a firm is narrower: in any report that goes to a client under your name, which figures came from fixed rules applied to the ledger, and which came from a model? If you cannot answer that for every number, you cannot stand behind the report.

That is the question firm owners are asking now. Accounting Today covered two studies that compared accountants with AI models on month-end tasks, and reported that both still saw AI as no substitute for accountants (Accounting Today, October 16, 2026).

Judgment tasks and arithmetic are different jobs

Most of a client report is arithmetic: totals, monthly amounts, the date a charge last posted, how often it repeats. Arithmetic on ledger data has one right answer, and a fixed rule gets it the same way every time.

A smaller part is judgment: what a cryptic descriptor refers to, which account a new vendor belongs in, whether two products overlap. That is where a model can help, as a suggestion a person checks.

Trouble starts when the two blur. A model that both reads the descriptor and adds up the year can produce a total nobody can trace back to ledger lines.

Five questions to ask any tool before it touches client books

1. Which steps are rules, and which are model output?

Ask the vendor to name them. "We use AI" is not an answer. You want a plain list: this step is a rule, this step is a model suggestion.

2. Can every figure be traced to ledger lines?

Pick one number in a sample report and ask to see the transactions behind it. If the tool cannot show them, you would have to rebuild the figure yourself before sending it.

3. What gets flagged for a person to review?

A model suggestion should arrive marked as a suggestion, with a way for a person to confirm or dismiss it. A suggestion that lands in the report as fact is the risk.

4. What access does it need, and what does it do with it?

Read the vendor's security page and data processing terms before you connect a client file. Check that your engagement letter covers sharing client data with a third-party tool, and say how AI is used if your letter or policy calls for it.

5. Does the same input give the same output?

Run the same export twice. Rule-based figures should match exactly. If totals drift between runs, the arithmetic is not fixed.

How ShadowLedger splits the work

ShadowLedger builds a recurring-charge report from a client's QuickBooks Online or Xero books, or from a CSV export. The split in the report is stated in one line:

In the report, fixed rules calculate every figure. AI only suggests vendor names, flagged for review.

In practice, that means the totals, the charge history, and the flags for charges that repeat, possible duplicate vendors, amounts that changed and renewals you can see from the charges all come from rules applied to the payments in the ledger. When a descriptor does not match a known vendor, a model may suggest a vendor name and category. The suggestion is marked for review, and a person at the firm assigns it to a vendor or marks it as not software.

The report is built from payments in the ledger, no seat or login data needed. The firm reviews it and sends it under the firm's name.

That split means the firm can open the charge history behind each subscription and check the figures against the ledger before putting its name on the report. It is a working report for an advisory conversation. It is not an audit, a review engagement or any form of assurance.

Where the AI line sits in your own policy

If your engagement letter or privacy policy already has AI language, a tool like this fits a simple description: a rules engine calculates the figures, and a model suggests vendor names that staff review. Whatever you write, keep it literal. Name what the model does and what a person checks.

ShadowLedger has other AI features outside the report, and the security page lists each one. Data handling is on the same page and in the DPA.

Try it on one file

To see the split on a real export, run the free CSV audit on one company's file. It needs no card. For a book of clients, the Firm plan is $15 per client per month with a 4-client minimum ($60 a month), with a 14-day trial.

Related reading: what live connect reads, finding a client's AI spend and payments, not seats.

FAQ

Can AI do bookkeeping?

AI can do parts of it, such as suggesting categories or reading unclear vendor descriptors. A firm still needs a person to review the output, and any figure that goes to a client should trace back to ledger lines. Separate the judgment steps from the arithmetic before you rely on a tool.

Does AI calculate the figures in a ShadowLedger report?

In the report, fixed rules calculate every figure. AI only suggests vendor names, flagged for review. Each suggested name is marked for review, and a person at the firm assigns it to a vendor or marks it as not software. ShadowLedger has other AI features outside the report; the security page lists each one.

Is a ShadowLedger report an audit?

No. It is a working report on recurring charges in the client's books, built for an advisory conversation. The figures trace to ledger payments so the firm can check them, but the report is not an audit, a review engagement or any form of assurance.

What should a firm check before connecting an AI tool to client books?

Which steps are rules and which are model output, whether every figure traces to ledger lines, what is flagged for review, what access the tool needs, and whether your engagement letter covers sharing client data with a third-party tool.