What research actually says about AI at work
Five traceable sources on AI adoption, productivity and risks. Each figure includes what was measured and where its relevance ends.
The numbers in context
A global survey or forecast cannot promise the same result for your company. It can help you decide what to test in a pilot.
McKinsey, The State of AI 2025: this is the share of respondents reporting regular AI use. About one third reported scaling AI across their organisations. Using a tool is not the same as realising measurable value.
Brynjolfsson, Li and Raymond, Generative AI at Work: the study followed 5,179 customer support agents. After access to an AI assistant, productivity rose 14% on average and 34% for novice and lower-skilled workers. The result concerns a specific task at one organisation.
Gartner's 2025 forecast for the end of 2026, compared with fewer than 5% in 2025. This predicts a share of applications, not a measured cost saving or the proportion of companies running agents successfully.
IBM Cost of a Data Breach 2025: among 600 breached organisations, those with high levels of unsanctioned AI use had average breach costs $670,000 higher than those with low or no shadow AI. This is an association, not proven causation or a cost estimate for a small business.
Adobe Analytics, 2025 holiday season: year-over-year growth in clicks from AI tools to US retail websites. Adobe notes that the starting volume remained modest. Channel growth is not the same as a share of revenue.
How to calculate your own AI return
Start by measuring one repeatable task. Do not plug broad research averages into your budget.
How many cases do you handle each month, how many minutes does each take, and what does the work cost? Record errors and customer waiting time too.
Choose one bounded task. A person checks the output while you measure accuracy, time saved and running costs.
The value of work actually saved and other proven benefits minus monthly costs for the tool, review and corrections.
Divide the initial investment by a positive net monthly benefit. If that benefit is zero or negative, the solution will not pay back under those conditions.
See concrete work in the case studies. If you want to assess your own process, start with an AI audit.
Start with your own numbers
Describe one process that takes too much time. In a free 30-minute call, I'll tell you what is worth measuring and testing.