Measuring AI in Finance: Close Time, Cost per Invoice and Forecast Accuracy

By Jane Ha · Updated Oct 11, 2026

Measuring AI in finance means picking a small set of operating metrics that finance already trusts (days to close, cost per invoice, forecast error), recording each one before the tool goes live, and comparing the same definition after. The tool itself is rarely the hard part. The hard part is proving, to a CFO or an auditor, that the number moved because of the AI and not because of a quiet quarter, a new hire or a changed definition.

This guide is for controllers, finance directors and FP&A leads who have one or more AI pilots running, or about to start, and need to report results. It builds on the process guides for the month-end close, accounts payable and FP&A. If you're earlier in the journey, the broader question of why AI ROI is hard to prove is a useful companion.

Key Facts

• APQC's General Accounting benchmarking survey, reported by CFO.com in 2018, put the top 25% of organizations at 4.8 calendar days or less to close, the median at 6.4 days and the bottom 25% at 10 days or more, across 2,300 respondents (CFO.com).
• In that measure, cycle time runs from running the trial balance to completing the consolidated financial statements, and counts both active work and time spent waiting (CFO.com).
• Ardent Partners' 2025 State of ePayables report, based on 204 AP and finance leaders, put best-in-class AP teams at $2.65 per invoice against a $12.42 average (Ardent Partners).
• Forecasting textbook guidance defines MAPE as the mean of absolute percentage errors, and warns that percentage measures are infinite or undefined when an actual value is zero and unstable when it's close to zero (Hyndman and Athanasopoulos).
• The same text says a test set should ideally be at least as large as the longest forecast horizon you need, and describes rolling-origin evaluation, where each forecast uses only earlier data (Hyndman and Athanasopoulos).

Why Finance Teams Need a Measurement Plan Before the Pilot

Most AI pilots in finance get judged by feel. The AP clerk says it's faster, the controller says the close "felt smoother," and six months later nobody can say by how much. That's a problem for two reasons.

First, budget. A tool that can't show a result gets cut at the next planning cycle, even if it was helping. Second, controls. If an AI step touches financial reporting, you'll need evidence that it performs inside a tolerance, which the internal controls guide lists as backtest and accuracy results. Measurement does double duty.

The good news is that finance already has well-understood metrics. You don't need a new AI scorecard. You need the old scorecard with clean baselines.

The Three Anchor Metrics

Close time

Close time is the number of calendar days between running the trial balance and completing the consolidated financial statements. That's the definition behind the APQC measure above, and it includes waiting time, not just hands-on work.

The benchmark spread is useful context. A median of 6.4 days with the top quartile at 4.8 or under and the bottom quartile at 10 or more means the gap between a good close and a slow one is several days. But it's a 2018 report of one survey question, so use it to place yourself roughly, not to set a target. Your own trend matters more than anyone's median.

What AI should move: reconciliation and matching hours, the number of manual journal entries, the time spent chasing schedules, and post-close adjustments. What it shouldn't move on its own: the sign-off steps, which are still done by people.

Cost per invoice

Cost per invoice is the fully loaded cost of running the AP process in a period divided by the number of invoices processed in that period. "Fully loaded" is where teams disagree. Some count only AP staff salaries, others add software, bank fees, overhead and the time of approvers in other departments.

Ardent's figures give a sense of the spread: $2.65 for best-in-class teams against a $12.42 average. Don't copy those into your business case. The report's sample, invoice mix and cost definition won't match yours. If a vendor quotes a benchmark, ask what's inside the number. The cost components are yours to define, and the only rule is to use the same ones before and after.

What AI should move: touch time per invoice, the share of invoices that need manual coding, exception handling time and duplicate payments caught. Volume matters too. If invoice count doubles after an acquisition, total cost may rise while cost per invoice falls, which is fine, but say so in the report.

Forecast accuracy

Forecast accuracy is how close the forecast was to what happened, measured on data the model hadn't seen. The textbook measures are mean absolute error (MAE), root mean squared error (RMSE) and mean absolute percentage error (MAPE). MAPE is popular in finance because it's a percentage, but it breaks down near zero, which is why it suits revenue better than net cash flow. For a line that crosses zero, use MAE or a scaled measure instead. The cash flow forecasting guide covers that choice in more detail.

Two habits make forecast metrics honest. Evaluate on held-out periods, never on the history the model was fitted to, because fitted accuracy always looks better than real accuracy. And compare against a naive baseline, such as "next month equals the same month last year." If an AI model can't beat that on your data, it hasn't earned its keep. The FP&A guide explains why simpler methods often hold their own.

What AI should move: error by line and horizon, bias (whether you consistently over- or under-forecast), and the days from period end to a finished forecast package.

The Measurement Table

Use this as the skeleton for a one-page scorecard. Add rows for your own processes.

How to Set a Baseline, Step by Step

1. Pick one process and two or three metrics. Resist the urge to measure everything. Close time and manual journal count for the close, cost per invoice and touchless rate for AP, error and bias for one forecast line.
2. Write the definition down. Where does the clock start and stop? Which costs are in? Which invoices are excluded? A paragraph now prevents an argument later.
3. Pull the history. Use at least six periods where you can, so one odd month doesn't set the baseline. Note anything unusual: an acquisition, a system migration, a holiday-heavy month.
4. Capture it before anyone changes the process. Pilots tend to start with enthusiasm and no snapshot. The data readiness guide is the place to check that the source data for each metric is clean enough to trust.
5. Add a naive comparison. For forecasts, score the "same as last year" or "last value" forecast on the same periods. For the close, the comparison is simply last year's trend.
6. Run in parallel where you can. For a few cycles, let the old process and the AI-assisted one run side by side. Compare line by line. This is also what your auditors will want to see.
7. Hold out the test. For forecasting, keep recent periods unseen by the model, at least as long as your longest horizon, and use rolling-origin evaluation so each forecast only sees earlier data.
8. Report on a fixed schedule. Same metrics, same definitions, every cycle. Put the baseline next to the current number every time.

Examples by Function

Controller, month-end close. Picture a mid-size company whose controller records calendar days and manual journal entries for the last eight closes, and notes the range. After adding automated matching on bank and intercompany accounts, the controller tracks the same two numbers plus post-close adjustments. The goal is days down, manual entries down and adjustments flat. Flat adjustments are what make a faster close believable.

AP manager. A team counts its monthly invoice volume and writes down every cost it's including: AP salaries, software, bank charges and a share of procurement time. After rolling out document capture, it reports cost per invoice, touchless rate and reviewer correction rate. The correction rate is the early warning. If the touchless rate climbs while corrections climb with it, the tool is pushing errors downstream.

FP&A lead. The team scores its current revenue forecast, a naive seasonal forecast and a machine learning model on the last 12 months of held-out data. Say the model beats the naive baseline on some product lines and loses on others. The team deploys it where it wins and keeps the existing method elsewhere. A mixed result honestly reported is worth more than a blanket "AI improved accuracy."

Common Pitfalls

• No baseline. The most common failure. Without a before number, every result is an anecdote.
• Changing the definition mid-project. If "close" starts at a different step after go-live, the days saved are partly an accounting trick.
• Measuring speed without quality. Faster close with more post-close adjustments, or faster invoices with more duplicate payments, isn't improvement. Always pair a speed metric with a quality metric. The accuracy controls guide covers the quality side.
• Scoring a forecast on training data. Fitted accuracy overstates real accuracy.
• Using MAPE where actuals hit zero. Net cash flow and small cost lines produce misleading percentages. Pick a measure that handles them.
• Borrowing a benchmark without its definition. A vendor's "average cost per invoice" may count different costs than yours. Compare like with like or don't compare.
• Crediting AI for everything. Seasonality, volume changes, new staff and process cleanups all move these numbers. Where possible, keep a control group, such as one entity or product line without the tool.
• Averaging away the story. Report by line, entity or invoice type. A strong average can hide a segment where the tool does badly.
• Ignoring the cost side. Software, implementation, review time and rework belong in the ROI.

What to Do This Quarter

Choose one process. Write the definition of two metrics in plain language, pull six periods of history, and score a naive comparison. That's a week of work, and it turns the next pilot from a story into a result. Then agree, before launch, what outcome counts as success and what would make you stop.