An operations team that bills nobody cannot price an hour, which is usually why the first business case came back from finance. Hours saved times a loaded hourly cost is an honest measure of capacity, and in a cost center it is not a saving. Salary is a fixed line. An hour freed and not redirected leaves the P&L where it was.
So measuring AI in operations means keeping three things apart: how much capacity the team got back, what it did with that capacity, and which cash line moved as a result. Only the third is a saving. The first two are what you manage, and they are where this falls apart.
The useful part is that operations already holds the records that answer the question: you count volume, your systems stamp every item in the queue, and somebody keeps a list of what had to be done twice.
Why the hours-saved model collapses in a cost center
In a firm that sells time, a freed hour has a price and a place to go: another matter, another engagement, another invoice. Operations has neither, so the model most measurement pages hand you is missing its last step.
Show that a six-person finance operations team now closes the month in four days instead of six, and finance asks the only question that matters: did we spend less money. If those six are still on payroll at the same salaries, the answer is no. Not yet.
That is not a reason to drop the number. It is a reason to stop presenting capacity as cash, because the moment a CFO spots a notional dollar figure, every other figure in the deck loses credibility too. We have watched a decent program lose its renewal over one line of arithmetic nobody believed, which is why our walkthrough of a finance-ready model keeps the assumptions separate.
The surveys show the same gap from above. McKinsey’s August 2026 survey of 1,719 respondents across 97 countries found 80% reporting improved personal productivity but only 37% attributing any EBIT impact to AI, with around 6% qualifying as high performers. That is the hours-to-dollars conversion failing at scale, in firms far better resourced than yours.
Capacity has a price, and the price is not a saving
Price the capacity anyway. Just label it correctly.
The Bureau of Labor Statistics puts employer costs for civilian workers at $49.46 per hour worked in June 2026, split as $33.85 in wages and $15.61 in benefits. For most operations roles at a mid-sized company that is defensible, and it is the right figure because it is what the employer pays, not what the employee takes home.
Four hours a week across a team of twelve is about 2,500 hours a year, or about $123,000 of capacity. Put it in the deck as capacity created, with a second row underneath called capacity redirected, and leave that row empty until a manager fills it in. The format turns a claim finance distrusts into a decision finance has to help make.
Be conservative on the input. Self-reports run high: BCG’s 2026 survey of close to 12,000 employees and leaders found 42% of frontline AI users claiming eight hours a week back, rising to 60% in marketing. The randomized evidence is calmer. A field study of 7,137 knowledge workers across 66 firms found adopters spent about two fewer hours a week on email and no measurable change in the volume or mix of their work. Two hours is a real result. Eight is what people feel.
Four numbers your operations team already keeps
The ladder runs from the number you have this afternoon to the one that takes a quarter.
| Layer | What to count | Where the data already lives | Time to move |
|---|---|---|---|
| Volume per person | Items handled per full-time person per month | Ticket system, ERP, close checklist | 1 to 2 months |
| Cycle time | Calendar hours from intake to done, one named deliverable | Queue timestamps you already store | 2 to 3 months |
| Rework rate | Share of items touched twice, corrected, or escalated | Credit memos, reopened tickets, audit adjustments | 1 to 2 quarters |
| Cash lines | Overtime, temp labor, outsourced processing, deferred hire | Payroll and the GL | 2 to 4 quarters |
Volume per person is the one nearly every team can produce today and nearly nobody uses, because it feels crude. It is also the metric the strongest study in this field chose. Brynjolfsson, Li and Raymond tracked 5,179 customer support agents and measured issues resolved per hour, finding a 14% gain on average, concentrated almost entirely among newer and lower-performing staff. Throughput per person, not hours saved, because support operations has no billing rate either.
That concentration matters. A gain showing up mostly in people who joined this year is the program working as the evidence predicts; a gain showing up only in your two strongest people means you trained the already-capable. Our bands for what good looks like at six and twelve months cover the adoption side.
Count full-time equivalents on the process, not headcount in the department: a team that quietly lost a person and held volume flat produced the result, and one that added a contractor did not.
Pick one recurring deliverable and time it
One. Not a dashboard of nine.
The deliverable has to recur, belong to a named owner, and already carry timestamps: the monthly close, the vendor onboarding packet, the weekly exception report, the renewal quote, the audit response. Record median calendar hours from intake to delivered for the three months before the sessions and the three after, so one disastrous quarter-end does not swamp the signal.
Cycle time is the strongest number a non-billable team has. It holds up when volume spikes, it holds up when someone leaves, and it is the number the rest of the business feels, so a sales lead or a controller confirms it unprompted. That does more in a board meeting than any chart you build.
Pair it with a cost-per-item figure when your process has a public benchmark. Accounts payable does: APQC’s cross-industry data puts the median total cost of processing one invoice at about $6, including people, systems and overhead. Moving toward the top quartile of that distribution gives you an external yardstick, which beats an internal percentage nobody can calibrate.
Rework is where a cost center actually leaks
Operations does not mainly lose money on slow work. It loses money on work done twice.
A corrected invoice, a reissued quote, a reopened ticket, an audit adjustment, a packet returned for a missing field: each costs the original effort, plus the rework, plus some of the relationship. Most teams have never put a number on it, which makes it the first thing to start counting, before anyone is trained.
Count it as a share, because volume moves. Reopened tickets over tickets closed. Corrected invoices over invoices posted. Packets returned over packets submitted. Then hold the definition still for a year, even when a better one occurs to you in month four: a redefined metric is an uncountable metric.
Rework also guards against the failure an assistant introduces. A team that gets faster and sloppier is not improving, and throughput alone will not show it: a first-draft-everything habit raises output and the correction rate together, which nets out worse. Running rework beside volume catches that inside a quarter, and if the numbers move the wrong way, the diagnosis of an enablement push that did not land separates a skills problem from a process one.
The three forms finance will book
Capacity becomes a saving in three ways. Everything else is a story.
A cash line goes down. Overtime hours, temp agency invoices, outsourced processing fees, seasonal contractors, expedited shipping caused by late paperwork. These sit in the GL, they are already budgeted, and a reduction needs no argument. If your team runs overtime at close, start there.
Volume rises and headcount does not. If the business grew 18% and operations absorbed it at flat headcount while cycle time held, you avoided the hire the old ratio demanded. State the ratio over time, name the role you did not post, and let the CFO price it. A number finance builds is a number finance defends.
A planned hire gets deferred with a date. Weaker but concrete, provided the requisition existed in the plan. Deferring a hire nobody approved is not a saving, and claiming it is the fastest way to lose the room.
| Claim | Will finance book it | Why |
|---|---|---|
| “We saved 2,500 hours, worth $123,000” | No | Salary line unchanged: capacity, not cash |
| “Overtime at close fell from 180 hours to 40” | Yes | Payroll line, already budgeted |
| “Volume up 18%, headcount flat, cycle time held” | Yes | Avoided hire, priced by finance |
| “The approved Q3 coordinator hire moves to next year” | Usually | Real requisition, dated, verifiable |
| “The team feels less overwhelmed” | No, and say it anyway | Retention argument, not a financial one |
The last row earns its place. Attrition in a processing team is expensive, and the support-agent study found the assistant cut turnover alongside the throughput gain. Report it separately from the money.
The redirect decision does more than the tool
The finding that should reshape how you run this: BCG found 66% of respondents receive limited or no guidance on what to do with the time they save, and more than half say they are not reinvesting it in more strategic work.
In a billable team that leak is partly self-correcting: an unused hour shows on a timesheet as unbilled. Operations has no such mechanism. Freed capacity dissipates into a slightly calmer week, nothing gets recorded, and the team ends up with real gains and no evidence.
So the artifact that decides your result is not a dashboard. It is a one-page note, written by the operations manager in the week the sessions end, naming the backed-up work the freed hours will go to: the reconciliation nobody has had time for, the vendor cleanup, the analyst work currently bought outside. Date it. Review it monthly. That note turns a capability into a number.
McKinsey’s data points the same way from the other end. Nearly three-quarters of its high performers report fundamentally redesigning workflows because of AI, against about a quarter of everyone else, and cost reductions are reported most often in service operations and supply chain, which is to say in the teams that bill nothing. The money is in operations, and it appears only when somebody changes how work flows rather than handing the same work to a faster typist. That is the case made in why buying licenses is not the same as building capability, and the method behind these numbers is set out in our guide to proving capability moved.
A cheap test of whether the redirect happened: ask four people to name, in one sentence and without opening anything, the work that is now different. “The exception report goes out Tuesday instead of Thursday” is a program with a second year. “It helps with emails” is a tool with good reviews.
Two numbers that mean stop
Write both down before the first session, because a measurement plan with no failure condition is a hope with a percentage attached.
Month six, no named deliverable has moved. Not a small gain, no gain, on any recurring work you timed. Cycle time flat, volume per person flat, rework flat. More training will not fix it. Either the work is not the kind an assistant helps with, or the sessions taught features instead of the team’s own documents.
Month twelve, healthy usage and no cash line or absorbed volume. Harder to accept, because the usage report looks fine and people say good things. If capacity was created for a year and nothing downstream moved, the constraint was never skill. Nobody made the redirect decision, and the fix is a management conversation, not another purchase order.
Neither verdict means the team is bad at its job. They tell you which of two problems you have, and only one responds to enablement spend.
Frequently asked questions
How do you measure AI in operations when nobody tracks billable hours?
Measure throughput, cycle time and rework instead of hours, then convert only where a cash line or an absorbed volume genuinely moved. Volume per full-time person per month and median calendar hours on one recurring deliverable both come from records you already keep, and neither needs a billing rate.
Can I use hours saved times salary cost in the business case?
Use it, but label it capacity created rather than money saved, and leave an empty row underneath for capacity redirected. The arithmetic is sound and the conclusion is not: salary is fixed, so an hour freed and not reassigned changes nothing on the P&L. Presenting capacity as cash costs you the CFO’s trust in every other figure.
What is a realistic time saving per person?
Plan on one to three hours a week for a knowledge worker using an assistant on real work, not the eight hours self-reports suggest. A randomized study of 7,137 workers across 66 firms measured about two fewer hours a week on email among adopters.
Which single metric should I pick if I can only track one?
Median cycle time on one recurring deliverable. It survives volume spikes and staff changes, it is already timestamped in your systems, and another department will confirm it unasked.
How long before any of this shows a result?
Volume per person moves in one to two months, cycle time in two to three, rework over one to two quarters, cash lines over two to four. A financial result reported at month three is noise, and demanding one kills working programs early.
How do I prove it was the training and not something else?
At this size you cannot prove it cleanly, so control what you can. Hold the definitions fixed, record three months of baseline, note every other change that hit the process in the same window, and use an untrained team as a comparison if one exists. Present the result as evidence rather than proof, and name the confounds first.
What if the team gets faster but quality drops?
That is a known failure mode, and rework rate catches it. Count reopened tickets, corrections and returned items as a share of volume and run it next to throughput monthly. A team producing more output and more corrections has not improved, and the fix is usually a review step in the workflow, not another training day.
What should I start recording before the sessions begin?
Three things: volume per full-time person on the process, median cycle time on the deliverable you chose, and rework as a share of volume. Baselines cannot be reconstructed afterward, and a program without one produces an argument, not a measurement.
Key takeaways
- In a team that bills nothing, hours saved is capacity, not cash. Price it at the $49.46 per hour employer cost, label it capacity, and keep a row for what was redirected.
- Use what you already hold: volume per full-time person, median cycle time on one named deliverable, and rework as a share of volume.
- If you can track one number, track cycle time: it survives busy seasons and staff changes.
- Finance books three things: a cash line that fell, volume absorbed at flat headcount, or a dated deferral of an approved hire.
- Rework catches what throughput hides: a team getting faster and sloppier at once.
- Two-thirds of respondents get limited or no guidance on what to do with freed time, so the manager’s redirect note decides your result more than the tool does.
If you want this applied to one real process in your own operation, with the baseline set before anyone is trained, book a free AI readiness call. We will pick the deliverable, name the three numbers worth recording, and say whether your work is the kind an assistant moves.
Arthur Wandzel
