What it is
Choose a single process where the cost is easy to see: answering routine enquiries, preparing quotes, processing invoices. Measure it, apply an AI tool with a person reviewing the output, and measure again. Move to the next process only when the first one shows a result.
What the evidence shows
Four studies explain why this advice is common.
- Gains show up in specific tasks. In a study of 5,172 customer-support agents at work, an AI assistant raised issues resolved per hour by 15%. The task was frequent, well defined and easy to measure.
- Broad availability changes little. A study covering the Danish labour market found that AI chatbots were widely adopted, yet earnings and hours did not change. An earlier version of the study put average time savings at 2.8% of working hours. In a field experiment with 7,137 workers at 66 firms, workers randomly given an office AI assistant spent about two hours less on email a week, and the amount and mix of their other work did not change. Three of the four authors worked for Microsoft, which sells the assistant.
- Work nobody was doing may gain most. In randomised experiments on a large online retail platform, the largest gain, 16.3% more sales, came from a chatbot answering buyers who had previously had no live help at all. AI-written product descriptions raised sales only for products that had fewer than 50 words of description before.
Many projects are also dropped. In an S&P Global survey reported in 2025, 42% of companies said they had abandoned most of their AI initiatives. One process with a clear measure is cheap to stop if it does not work.
How to do it
- List three to five processes that take the most staff time or cause the most delay.
- Pick the one that is most frequent and has the clearest standard for a good result. A process you are not covering today, such as unanswered enquiries or missing product descriptions, may gain most.
- Record today’s figures for two to four weeks: time per case, error rate, cost.
- Apply an off-the-shelf AI tool. Keep a person reviewing every output at first.
- Run it for a set period, then compare with the starting figures.
- Decide: keep, change or stop. Write down why.
- Only then pick the second process.
What to measure
- Time per case, before and after.
- Error or rework rate, before and after.
- The time the reviewer spends checking. It is part of the cost.
To see what these figures mean in money, put them into Will AI pay back?. The tool compares one chosen process with a rollout to everyone.
Where it fails
- The process is rare or vague. If it happens a few times a month, there is too little to measure.
- The process is broken. AI on top of a bad process gives a faster bad process. Fix or redesign it first.
- The pilot never ends. Set the decision date before you start.
- Narrow or wide is not settled. In an OECD survey of 2,018 small and medium-sized firms, only 2% of firms using AI for isolated tasks reported a transformational effect, against 23% of those using it across the business. The effects are what firms said, the OECD says the survey is not representative, and about three-quarters of the answers on AI came from Japan. Two other sources point elsewhere. A Federal Reserve analysis found that US small firms that had built AI into their processes had no better hiring expectations than firms still experimenting. In payroll data from Gusto, which sells payroll software, firms using AI for everyday tasks showed the same hiring gain as firms that had built it deep into their operations. Whether starting narrow or going wide gives more is not settled.
By company size
- Micro and small: one process is the whole programme for the first months. That is fine.
- Medium and mid-market: run two or three in different departments, each with its own baseline and owner.
- Enterprise: large firms usually have pilots already. Their problem is turning those pilots into redesigned workflows, so this practice matters less.
See it applied
Company cases
Companies that did this
- Duolingo The first target was course writing, where a single new course could take a small team years.
- eBay eBay aimed machine translation at one barrier: buyers in Latin America could not read listings from US sellers.
- EIST The owner picked one task: pricing custom parts, which could take hours per request and was costing the firm orders.
- Herges Stahl- und Blechbau The project was limited to one process, rolling sheet metal, with a design that can be extended to welding and assembly.
- Hitachi Energy The work was aimed at one product and one cost: transistor chips that fail the final test are scrapped.
- META-Regalbau The analysis covered one process: picking the many separate parts of an ordered shelving system in the warehouse.
- Milkforbubs The shop aimed the tool at one task: answering the same customer questions again and again, many of them at night.
- Ministry of Clean The software was aimed at one problem: a growing customer list that was being managed in spreadsheets.
- Rue La La The work was aimed at one decision: the first price of products the retailer had never sold before, which bring in most of its revenue.
- Temple & Webster The first figures the company reported were for one process: answering customers before and after a sale.
- Walmart Walmart pointed AI at one job: filling in and correcting product data, which its chief executive said “affects nearly everything we do”.
Sources
- Brynjolfsson, Li and Raymond (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. What this study did and found
- Humlum and Vestergaard (2025). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777 (revised March 2026). What this study did and found
- Dillon, Jaffe, Immorlica and Stanton (2025). Shifting Work Patterns with Generative AI. NBER Working Paper 33795; arXiv 2504.11436, version 4 (November 2025). What this study did and found
- Fang, Yuan, Zhang, Donati and Sarvary (2025). Generative AI and Sales Productivity: Field Experiments in Online Retail. arXiv 2510.12049, version 6 (29 June 2026). What this study did and found
- Gartner (2024). Prediction that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. gartner.com
- CIO Dive (2025), reporting an S&P Global Market Intelligence survey: 42% of companies abandoned most of their AI initiatives. ciodive.com
- OECD (April 2026). Empowering SMEs in the age of AI. Includes the D4SME survey of 2,018 small and medium-sized firms; the OECD says it is not representative. oecd.org
- Aarons and Sarkar (October 2026). Federal Reserve Bank of New York, Liberty Street Economics. Analysis of AI use and hiring expectations in the Small Business Credit Survey. libertystreeteconomics.newyorkfed.org
Related practices
- Measure time saved per task against a baseline Record how long a task takes and how good the result is before you add AI. Without that starting figure you cannot tell a real gain from a hopeful one. Do first · Effort: low (days, one person)
- Map tasks before deciding what to automate Break jobs into tasks and look at each one. Published scores rate how much of a job AI could do, its “exposure”. Treat a high score as a place to look, not as a decision to automate. Do first · Effort: low (days, one person)
- Redesign the workflow instead of adding a tool to it Giving people an AI tool inside an unchanged process changes little. Decide again who does each step, in what order, and where a person checks. Do next · Effort: high (months, several teams)