Before You Deploy AI, Record These Three Numbers

You bought the AI accounting tool. Invoices go out faster. Your bookkeeper seems less buried. Late payments feel like they're down. But when your CFO asks for the ROI number, you go quiet — because you never wrote down what "before" looked like.

That silence is the actual problem.

The measurement failure happens before deployment

Thongprim, Tulacharatkul, and Sincharoonsak (2025) surveyed 372 SMEs on AI-based accounting adoption and found improvements in cost information quality and decision-making efficiency. The "Human + AI in Accounting" study (2026) tracked 79 SMEs on a GenAI-enabled platform and documented productivity gains alongside a reallocation of accountant effort away from data entry. Both studies show real operational change. Neither of those gains is provable at your firm level if you walked into deployment without recording what you were starting from.

The three numbers you need to record before you flip the switch are days sales outstanding, total weekly bookkeeping hours, and your late-payment rate across your receivables portfolio. Not after. Before.

What each baseline actually measures

Days sales outstanding is the average number of days between issuing an invoice and receiving payment. Pull twelve months of invoices, calculate the average collection period, write it down. One number. It takes under an hour if your accounting software exports invoice and payment dates.

Bookkeeping hours require more honesty. Log every hour your team spends on data entry, reconciliation, categorization, and exception handling for four consecutive weeks. The 2026 "Human + AI in Accounting" study found that after GenAI adoption, accountants shifted effort toward business communication and quality assurance rather than eliminating hours outright. If you don't know what your exception-handling time looks like right now, a post-deployment dashboard showing fewer data-entry hours will tell you nothing useful about whether total labor cost changed.

Late-payment rate is the share of invoices in a given month where payment arrived after the agreed terms. Count them. The research scope underlying this diagnostic is direct: e-invoicing and digital tools improve processing speed but do not automatically fix strategic late payment behavior driven by buyer-power dynamics. A large customer who pays late because they know you need them more than they need you will keep paying late after you deploy AI. A drop in your late-payment rate post-deployment means something only if you know what the rate was before — and only if you can separate AI-driven improvement from a seasonal shift in your customer mix.

What the platform dashboard won't tell you

Here's the counterargument worth taking seriously: AI platforms generate their own comparison data. The 2026 study pulled proprietary transaction-level records from the platform itself, not from pre-deployment logs the founders kept. If researchers built a before-and-after picture from internal platform data, why do you need your own baselines?

The answer is that platform analytics measure platform activity. They track invoice send times, payment receipt timestamps, and ledger account volume. They do not capture whether a customer's payment behavior changed or whether it was already changing before you went live. [Inference: A platform showing faster average payment receipt after deployment cannot separate AI-driven improvement from a large buyer settling an old balance in the same quarter.] The platform also does not log the staff hours spent correcting AI miscategorizations or chasing edge cases the tool mishandled — hours that disappear from the data-entry column and reappear as "quality assurance," which the platform may not track at all.

Platform data is a partial substitute for one baseline. It is a poor substitute for the other two.

Record the numbers this week

Open your accounting software. Export twelve months of invoice and payment data. Calculate your current days sales outstanding. Start a four-week time log for bookkeeping tasks — every team member, every task category. Pull last quarter's invoices and flag every one where payment arrived after terms.

Write those three numbers down somewhere you will find them in six months.

The Thongprim et al. (2025) study shows AI accounting adoption improves decision-making efficiency when firms adopt it thoughtfully. "Thoughtfully" includes knowing what you measured before you started.