Hours returned: the AI metric your board should actually ask for

Sooner or later the board asks the finance lead how the firm is getting on with AI. The usual answer is a list of activity: licences bought, a pilot running, most of the team trying the chat tool at least once a week. None of that is a result. If you want to measure AI ROI in a finance team, there is one number worth reporting, and it is simpler than most dashboards: hours returned, net, per named recurring task.

This post explains why activity numbers mislead, what hours returned actually means, and how to keep the record behind it. The examples and numbers are illustrative.

Why "we use AI" is not a result

Activity metrics are easy to collect and easy to like. Seats, logins, prompts sent, percentage of staff who say they use AI weekly. They all measure spend and enthusiasm. None of them says whether a single piece of work got lighter.

Seat-level numbers also hide where the value sits. A preparer who uses a chat tool to tidy an email has used AI. So has a team that hands the first draft of every supplier statement reconciliation to an agent and reviews it. Both count as one active user. Only one of them changed the shape of a recurring task.

Finance people know this pattern already. Nobody reports "we bought a new accounting system" as a saving. They report the days taken off the close. AI deserves the same standard.

What hours returned means

Hours returned is the difference between how long a recurring task took before a step was handed over and how long it takes now, measured per cycle and multiplied by how often the task runs.

Three words in that definition do most of the work:

• Recurring. One-off tasks do not add up. A monthly task that runs for 30 clients does.
• Measured. The before number has to come from the work itself, not from memory. Most people underestimate how long a familiar task takes, and overestimate how long an annoying one does.
• Net. The after number includes everything a person still does: reviewing the agent's draft, correcting it, handling the exceptions it sends back, and the time spent writing the instructions in the first place.

Net is where most AI claims fall apart. A draft that saves three hours and needs two hours of checking has returned one hour, not three.

Measure AI ROI with a handover ledger

The record behind the number is a handover ledger: one row per step of a task that has moved, fully or partly, from a person to software or an agent. It can start as a spreadsheet. The useful columns are:

1. Task and step. "Supplier statement reconciliation, matching step", not "AP".
2. Autonomy level. Agent drafts and a person approves, agent acts and is reviewed, or agent acts alone. The ladder is covered in drafts first, trust later.
3. Baseline hours per cycle. Measured over at least one real cycle before the handover.
4. Current hours per cycle. Including review, correction and exception handling.
5. Cycles per period. Monthly for one client, or across every client the task runs for.
6. Corrections. How many drafts needed changing, and why.
7. Exceptions returned. Items the agent could not handle and passed back to a person.
8. Where the hours went. What the team did with the time.

The last column is the one boards rarely ask for and should. It is also the one that keeps the exercise honest.

An illustrative row

Take a hypothetical firm running supplier statement reconciliations for 20 clients. Say a preparer measures the matching step at 90 minutes per client per month before any handover, so 30 hours a month. An agent now drafts the matches and a person reviews them. In the first month, review takes 60 minutes per client because nobody trusts the drafts yet: 10 hours returned. By the third month, with clearer instructions and fewer corrections, review takes 30 minutes: 20 hours returned.

That row tells a board more than any adoption chart. It shows a real task, a real before and after, and a trend that depends on the quality of the documentation and the review, not on how many people logged in. It also shows why per task numbers matter more than per seat ones, a point we worked through in the real cost of a bank reconciliation.

Three ways the number gets inflated

• Counting gross, not net. Leaving out review time makes every handover look like a full saving. It also hides the drafts that are not yet good enough.
• No baseline. If nobody measured the task before, the saving is a guess, and guesses drift upwards. Measure first, even roughly.
• Counting hours nobody can find. If the ledger says 40 hours came back last month and nobody can say where they went, they probably went into overtime that stopped, slack that absorbed the next deadline, or nowhere. That is fine to admit. It is not fine to report it as a gain.

What to put in front of the board

A one-page quarterly view is enough:

• Tasks and steps handed over this quarter, and at what autonomy level.
• Net hours returned, per task and in total.
• The corrections trend, so the board sees whether trust is being earned.
• Where the hours went: more clients absorbed, faster close, more review, advisory work, training.
• The next candidates, ranked by how ready they are: how often they run, how rule bound they are, and how well they are documented.

That last line connects the metric to the work still to do. A task that scores well on everything except documentation is not an AI project yet. It is a documentation job for this month. Keeping this kind of ledger as a side effect of running the work is part of what we are designing Beaks to do, but the columns above work in any spreadsheet.

Keep the ledger visible to the team

A handover ledger that only the board sees will be read by the team as a list of whose work is next. Do the opposite. Share it. Credit the people who documented a task well enough for an agent to draft it, because that documentation is the reason the hours came back. Show plainly that the returned hours went into growth, review and better work, and let the people doing the task say whether the numbers match their week. They are the only ones who can.

One thing to do this week

Pick one recurring task where someone already uses AI for part of it. For the next cycle, have the person doing it note the minutes spent, including checking and fixing the output. Then estimate, honestly, how long the same cycle took before. That is your first ledger row, and the first number your board should see.