Almost every automation business case rests on one sum: hours saved, multiplied by an hourly rate. It is the formula in every vendor calculator and most internal proposals. It is also why so many automations are counted a success by the people who bought them and stay invisible in the accounts.
01 — What happened
Adoption became broad, the savings became measurable, and almost nobody checked where they went.
The European Central Bank published survey work on 26 August 2026 on what workers themselves report. The share using AI doubled in two years, from 26% in 2024 to 52% in 2026. Users save a median of three hours a week — about 7.7% of median working time — though only 48.8% of workers both used AI and saved time, pulling the economy-wide figure closer to 3.8%. The authors’ caveat matters more than the headline: those savings “only translate into higher productivity if workers turn the freed hours into extra output, rather than using them for less directly productive activities.”
The Federal Reserve Bank of New York reported on 1 September 2026 that 61% of service firms in its region now use AI, up from 40% a year earlier, and 51% of manufacturers, up from 26%. Yet usage is thin inside those firms: a median of 17% of workers at service firms, 7% at manufacturers. Three-quarters call their investment minimal to modest, and retraining — not restructuring — was the main workforce response.
Wide adoption, shallow depth, real but modest savings. The third finding should bother any owner. In the Thomson Reuters Institute’s 2026 AI in Professional Services Report, drawn from more than 1,500 respondents across 27 countries, only 18% said their organisation tracks the return on its AI tools at all — and those that do mostly track internal operational measures rather than revenue, client satisfaction or new business won.
Writing in the LSE Business Review in February 2026, Kristina Khutsishvili put the underlying argument plainly: minutes-saved metrics skew attention towards automating routine tasks and away from whether the technology produced anything new. The pattern is not an adoption problem. It is a measurement problem.
02 — Why it matters
Hours saved is an input. Money is an output. Between them sits an assumption almost no business tests.
The hours disappear in three common ways. The first is dilution: three hours a week is about thirty-five minutes a day, spread across interruptions, and time at that granularity gets absorbed by the working day rather than reallocated to anything you could name.
The second is shape. The New York Fed figures describe organisations where a minority of staff use the tools and the structure of work is unchanged. Freed capacity becomes output only if demand is waiting and someone points the capacity at it. Neither happens by default.
The third is that nobody looks. If 18% measure the return, the rest run on anecdote — and with no measured result you cannot tell a good automation from a bad one, so the next purchase goes to whichever demo was most impressive.
03 — The economics
An automation converts into money through exactly three routes. A business case that names none of them is not a business case.
Route one: a cost you stop paying. A licence cancelled, a vacancy not filled, contractor work brought in-house, overtime that stops. Cash, traceable to a line in your accounts.
Route two: revenue you can trace. Faster response on enquiries, better conversion, fewer renewals lost to being forgotten. The value shows up in work won, not time released.
Route three: a constraint removed. The process was capping throughput and now is not. You can take the extra volume, serve outside working hours, or accept the bigger contract without hiring first.
An automation that does none of the three bought convenience — legitimate, but budget it as convenience rather than writing it into a business case as a saving.
Take an illustrative service business with, say, 400 inbound enquiries a month, handled by two coordinators. Assume eight minutes each for triage, logging and first reply: roughly 53 hours a month. The standard formula multiplies 53 by a fully loaded hourly cost and declares a saving.
Now test it. Payroll is unchanged — nobody is leaving, and 53 hours is about a third of one person. Route one yields nothing, unless you were about to hire, in which case the deferred hire is the real number. Route two is where the value more likely sits: if first response falls from four hours to four minutes and that lifts quote-to-win by even two percentage points across 400 enquiries, the return is in work won, and your pipeline already measures it. Route three applies if those coordinators were why enquiries went unanswered in the evening — then what you removed is a ceiling, not an overhead.
Three costs are missing from that formula.
The running cost has changed character. Automation used to be a fixed build plus a fixed licence. Increasingly it is metered: Salesforce’s published Agentforce pricing lists $2 per conversation, or Flex Credits at $500 per 100,000, where a standard agent action draws 20 credits and a voice action 30. A successful automation then gets more expensive as it succeeds. Payback is no longer build cost divided by monthly saving; it is a margin per unit of work, and that margin has to survive growth — so model it at three times today’s volume. This is the shift we covered when vendors moved to charging per resolution rather than per seat.
Then the residual. Easy cases automate first, so what is left for people is the hard remainder: longer, more escalation, more judgement. Average handling time on the human queue usually rises.
Then supervision. Something upstream changes — a form field, a template, an API — and the automation needs an owner. Unowned automations fail quietly, which is the expensive way to fail.
The honest sum: value realised through route one, two or three, minus metered running cost and supervision, measured against the build. One discipline makes it work — write down the number you expect before you build, and the date you will check it.
04 — What businesses should do
Name the route before you build. Decide which of the three this automation is for, and what single number will move. “Fewer errors” is not a number. “First response under five minutes on all 400 enquiries, visible in the CRM” is.
Count the baseline for one month, by hand if you must. Volume, honest handling time including interruption and rework, fully loaded hourly cost. Without a baseline you will never prove the return.
If the hours are not leaving the payroll, say so out loud. Fifty-three hours that nobody leaves is capacity, not cash. Capacity is valuable when you have demand to fill it; when you do not, you have bought a quality-of-life improvement — a fair purchase, badly described.
Prefer automations whose return lands in a number someone already tracks — conversion rate, renewal rate, days to invoice. If proving the case needs a measurement system built first, it will not get proved.
Set the review date when you commit, not when it goes wrong. Ninety days, one named owner, three figures: unit volume, unit cost, and the metric you named at the start.
Apply a volume threshold honestly. As a rough test, a process running fewer than about fifty times a month at under ten minutes each is usually not worth automating: supervision will outlast the saving. Fix the process instead. The arithmetic of automating follow-up starts working only once volume and repetition are both high.
05 — Where technology fits
For many businesses reading this, the first build is not an automation at all. It is a way to count.
If you cannot say how many enquiries arrived last month, how long they took, or how many were answered within a day, the measurement problem comes first — and it is not an AI problem. It is a system-of-record problem: one place where the work is logged, so the baseline exists and the ninety-day review has something to read. Usually that is a CRM or a connected system, the same gap behind the cost of disconnected systems.
Where automation genuinely pays, it is often deterministic rather than intelligent: routing, reminders, handovers, data moving between two systems that should already be talking. Fixed cost, no residual, almost no supervision — unglamorous, and usually the best return in the building.
AI earns its place where work needs judgement at volume — reading, qualifying, drafting — and even then against a record that already exists, which is where AI actually creates economic leverage rather than where it merely demonstrates well.
And the legitimate no: if a process is run by one person a handful of times a week, and that person is not your bottleneck, you probably do not need an AI system for this yet — possibly not any system at all.
06 — RevenueStack’s take
Our view: the hours-saved formula survives because it is the only number that is easy to get. It flatters buyer and vendor at once, which is why it sits on both sides of the table and is challenged by neither.
The more useful questions are plainer. What did we stop paying for? What did we sell that we would not otherwise have sold? What can we now do that we could not before? Most businesses can answer all three in an afternoon — if they will accept “nothing yet” as one of the answers.
That willingness is the whole thing. The businesses getting a return from automation are rarely the ones with better tools. They are the ones that named the number before they built, then went back and looked. Almost nobody looks, which for now makes it a cheap advantage.
Sources
- AI adoption and the productivity promise: what workers report — European Central Bank, 2026-08-26
- Businesses Are Using AI to Transform Work, Not Cut Jobs — Federal Reserve Bank of New York, Liberty Street Economics, 2026-09-01
- 2026 AI in Professional Services Report: AI adoption has hit critical mass, but now comes the tough business questions — Thomson Reuters Institute, 2026-02-09
- AI productivity gains should be measured in more than minutes saved — LSE Business Review, 2026-02-18
- Agentforce Pricing — Salesforce, 2026-10-01
