A company buys licences, launches a copilot and reports success one quarter later: every employee saves 40 minutes a week.
Across 100 people, that adds up to more than 3,000 hours a year. The spreadsheet looks excellent. By December, nobody can point to where those hours went, which cost disappeared or which business result improved.
Saved time creates value only when the organisation removes it from the workflow and assigns it to a measurable outcome: more sales, shorter lead times, fewer errors, higher throughput or lower cost.
Most dashboards measure tool adoption. The P&L requires a harder measure.
Poland: rapid adoption, little movement in employment
According to the National Bank of Poland's Quick Monitoring Survey, 42% of Polish companies were using AI in the second quarter of 2026. The figure stood at 14% at the end of 2024. Adoption tripled in 18 months.
The effect on employment remained marginal. Fewer than 1% of surveyed companies had reduced headcount because of AI. The rate was 0.1% in manufacturing, 1.8% in trade and 1.7% in transport. Meanwhile, 35% of companies had not begun analysing potential AI applications; among micro and small businesses, the figure reached 46%.
The survey primarily captures reported use and employment. The distance between those measures is revealing: installing a tool rarely changes the structure of work on its own.
The spread across Polish adoption studies is equally useful. Eurostat reported that 5.9% of Polish enterprises with at least ten employees used AI in 2024. A 2026 study by the Polish Agency for Enterprise Development and Jagiellonian University found active use among 23% of 1,822 employers across 12 industries with high AI potential. The National Bank of Poland reported 42%.
The range is sevenfold. Each study uses a different date, sample and definition. This makes "our company uses AI" a poor KPI. Does it mean one licence, regular use by part of the team, an automated workflow or a result visible in EBIT? Adoption says little without a threshold, frequency and business outcome.
Value leaks at three levels.
1. The individual: the verification tax
In August 2026, while prioritising AI implementations for a 120-person agency, I added a verification multiplier to the model. Popular frameworks such as ICE and RICE ignore the cost of checking an output. With generative AI, that cost can decide whether a use case is worth pursuing.
I use this formula to rank candidates:
Priority score = (annual volume × unit time saving in minutes) / (implementation effort in person-days × verification multiplier × data-class multiplier)
This formula sets the order of implementation. Financial ROI requires a separate calculation, which comes later in the article.
The verification multiplier ranges from 1.0 to 3.0. A low value fits tasks whose outputs can be assessed quickly or tested automatically. A high value applies when the work requires a full read, source reconstruction or expert approval.
If AI cuts the preparation of an analysis from 30 minutes to ten, while checking takes another 15 minutes, the saving is five minutes. A dashboard tracking the first draft will report 20.
The scale is already measurable. In Sage and IDC research, finance professionals spent almost 13 hours a week reconstructing, validating and defending AI outputs. Forty-eight per cent reported 15 hours or more and 19% reported 30 hours or more. In Adaptavist research, 42% of knowledge workers spent more time checking AI output than they saved by using it, while 52% regularly corrected AI-generated work from colleagues. Workday estimated that correction, rewriting and verification consume about 37% of the time initially saved.
There is also a cost transferred to the recipient. BetterUp and Stanford used the term "workslop" for low-quality AI-generated material. Each incident took close to two hours to resolve and carried an estimated monthly cost of $186 per employee.
Time to first draft works well in a demo. ROI depends on total cycle time: preparation, checking, correction, approval and exception handling.
2. The organisation: no baseline
A return needs a reference point. When a company has no pre-implementation measure of process time, hand-offs, error rates and rework, it can only report activity later.
Building a baseline does not require a six-month analytics project. I use a short task map: six questions and roughly ten minutes per person. It captures the task, frequency, unit time, waiting time, cost of errors and required level of control. Several hundred responses are grouped by function and process. The result is a map of working hours that reveals high-volume opportunities with a manageable verification burden.
After launch, we return to the same map and compare the complete cycle. We then check whether the released capacity reduced a queue, increased completed cases or lowered external spend.
McKinsey classifies only 6% of surveyed organisations as "AI high performers": companies attributing at least 5% of EBIT and significant value to AI. In its 2026 State of AI survey, nearly three-quarters of that group had fundamentally redesigned workflows. Among other organisations, the figure was roughly one-quarter.
Polish data shows the same measurement gap beyond AI. According to Bank Gospodarstwa Krajowego, digitalisation carries high or very high strategic priority for 64% of SMEs, yet 91.5% do not measure results through KPIs. The market still lacks a method connecting investment with financial outcomes.
The formula: from minutes saved to financial ROI
I calculate AI implementation ROI in five steps.
1. Net hours
Start by deducting the time spent operating the AI system, checking its work, correcting it and handling exceptions:
Net hours = annual task volume × (time before AI − time with AI − verification time − rework time) / 60
Annual task volume is the number of completed tasks. All time inputs are minutes per task.
2. Captured hours
Some of the net saving will disperse across the working day. Apply a capture factor between zero and one:
Captured hours = net hours × capture factor
The factor reaches 1.0 when the company removes work from schedules, cuts overtime or outsourcing, avoids planned hiring, or increases completed volume. A factor of zero means that cost, throughput and revenue remain unchanged.
3. Annual business benefit
Annual benefit = avoided labour and external spend + incremental contribution margin + avoided error cost + retired-tool savings
Recovered hours can be valued at an hourly rate when they reduce cash expenditure such as overtime, temporary labour, outsourcing or planned additional headcount. When the capacity moves into sales or service, use the incremental contribution margin generated by higher throughput. Counting the same hours at an hourly rate and adding the margin they generated would overstate the result.
4. Total cost of ownership
Year-one TCO = one-off implementation cost + annual running cost
The one-off cost includes discovery, process mapping, integration, testing, migration, training and change management. Running cost covers licences, API usage, monitoring, support, governance, security and continued development.
When verification time has already been deducted from net hours, it should not appear again in TCO.
5. ROI and payback
Year-one ROI = (annual benefit − year-one TCO) / year-one TCO × 100%
Payback in months = one-off implementation cost / ((annual benefit − annual running cost) / 12)
What remains of the 40 minutes
Assume 100 employees report saving 40 minutes a week across 46 working weeks. The gross result is 3,067 hours.
Verification and rework consume 37%, leaving 1,932 hours.
The capture factor is 0.6, giving the company 1,159 usable hours.
Three hundred hours reduce outsourced work and overtime by EUR 36,000.
The remaining capacity generates EUR 60,000 in additional contribution margin.
Fewer errors and retired tools add EUR 30,000.
The annual benefit is EUR 126,000. One-off implementation costs EUR 50,000 and annual running cost is EUR 60,000.
Year-one ROI = (126,000 − 110,000) / 110,000 = 14.5%
Payback = 50,000 / ((126,000 − 60,000) / 12) = 9.1 months
The first spreadsheet showed 3,067 saved hours. The financial calculation produces EUR 16,000 of year-one return. That figure can survive a conversation with the CFO.
3. Leadership: conflict with the current revenue model
The hardest implementations often sit inside the part of the business that currently makes the most money.
One question helps expose the problem: what is this company's "search ads"? Which revenue line exists because the customer cannot get an easy answer, cannot serve themselves or pays for time rather than output?
This is where resistance becomes strongest. Even with capable technology and a competent team, the organisation may limit scale because full deployment would erode current revenue.
Chegg built its business around paid access to study support. Generative AI and answers displayed directly in search weakened both pillars. The company cut 22% of its workforce in May 2025 and another 45% in October. Its shares lost about 99% of their value from the 2021 peak.
BigLaw shows a less dramatic version of the same tension. AI reduces the time needed for research, document review and first drafts. Hourly billing still rewards billable time. In 2025, according to The American Lawyer's Am Law 100 ranking, aggregate revenue across the Am Law 100 grew 13% to $178.95 billion, while the average blended hourly rate reported by Brightflag increased 8.3% to $1,145. The incentive structure allows law firms to retain part of the productivity gain. Clients do not always receive a shorter bill.
SaaS faces a related pressure. An agent performing the work of several users may reduce the number of licences a customer needs. A vendor charging per seat has reason to add AI features while protecting a commercial model tied to user count. That slows agent-first workflow design.
AI's largest potential often sits where full deployment requires a change in how the company earns money. The decision quickly becomes political.
How to calculate an AI ROI that can reach the P&L
Measure the baseline in hours. Do it before buying licences or selecting a vendor. Include queues, errors, corrections and hand-offs.
Put verification cost into use-case selection. It should affect implementation priority at the business-case stage.
Measure the complete cycle. Start when the task is assigned and stop when the output is accepted and ready to use. The first draft is one step.
Examine the process that threatens your own revenue line. It will reveal the available value and the barriers that a technology workshop will miss.
Name an owner for the recovered hours. Before implementation, decide how much time should be released, who will capture it and which result it will support.
Unassigned time quickly disappears into meetings, inboxes and additional tasks. Its value emerges when it changes throughput, cost or revenue. That value belongs in the ROI calculation.
At ROI and Shine, we use this method across AI diagnosis, process mapping, implementation and post-launch measurement. It gives management a business case grounded in operating data and gives the implementation team a clear target to deliver.
Sources
- National Bank of Poland (NBP), Quick Monitoring Survey, second quarter of 2026, as reported by the Polish Press Agency (PAP)
- Eurostat, use of artificial intelligence in enterprises, 2024 data
- Polish Agency for Enterprise Development (PARP) and Jagiellonian University, 2026 study of 1,822 employers in 12 industries
- Sage and IDC, research on finance professionals and AI output verification
- Adaptavist, research on knowledge workers and time spent checking AI output
- Workday, research on time saved by AI and time lost to correction and rework
- BetterUp Labs and Stanford Social Media Lab, "workslop" research
- McKinsey, The State of AI 2026
- Bank Gospodarstwa Krajowego (BGK), study on digitalisation priorities and KPI measurement among Polish SMEs
- Reuters, Forbes and Chegg company announcements on 2025 workforce reductions and share price
- The American Lawyer, Am Law 100 ranking for 2025 results
- Brightflag, Hourly Rates in Am Law 100 Firms: 2025 Edition