The most expensive sentence in many artificial intelligence plans is, "We need an AI strategy." It sounds serious, attracts a committee, and postpones the work of identifying the one process that should improve first.

Businesses do not earn a return from possessing AI. They earn it when a defined task becomes faster, cheaper, safer, or better without creating a larger problem elsewhere.

That narrows the starting point considerably. One repeated problem. One responsible owner. One measure taken before and after. Everything else is theater until those three things exist.

A suitable first use case is frequent, time-consuming, and bounded. It has recognizable inputs and an output that a competent person can judge. Examples include classifying support tickets, comparing supplier quotations, summarizing recorded meetings, drafting routine customer replies, or extracting fields from standard documents.

The task should also tolerate correction. A poor first draft of an internal summary can be fixed. An incorrect dosage recommendation, legal filing, safety instruction, or wire-transfer decision carries consequences that make it a poor place to learn.

This is why starting with a list of fashionable tools reverses the logic. A software demonstration will always find a problem that appears to suit the product. Start with the costly friction already visible in the business, then ask whether AI is the best intervention. Sometimes a clearer form, a checklist, or a database rule will be cheaper and more reliable.

Record the current process for two weeks. How many items are completed? How long does each take? How often is work returned for correction? What does the work cost? How do customers or colleagues rate the outcome? Without a baseline, every polished demonstration feels like progress.

Good evidence shows that carefully matched tools can help. An NBER study examined 5,179 customer-support agents after the introduction of an AI assistant. Productivity, measured by issues resolved per hour, rose 14% on average. Novice and lower-skilled workers improved by 34%, while the effect on the most experienced workers was small.

The mechanism mattered. The system had been trained on examples of successful and unsuccessful conversations and offered suggestions that agents could accept, adapt, or reject. It was attached to a specific workflow with a measurable result. It was not a general instruction to "use AI more."

A tool that succeeds at one task can fail at the next. In an experiment with 453 college-educated professionals, MIT researchers Shakked Noy and Whitney Zhang found that ChatGPT cut the time spent on mid-level writing tasks by 40% and raised independent quality ratings by 18%. The tasks included press releases, short reports, analysis plans, and sensitive emails.

The researchers also stressed that the assignments did not demand extensive factual checking or detailed knowledge of a company's customers and aims. The business lesson is not that AI works or that AI fails. It is that performance is local. Test the actual task with the actual people, data, review process, and standard of quality that your business uses.

This also means measuring the hidden work. If a draft arrives in five minutes but requires 40 minutes of fact-checking and rewriting, the tool has not saved 55 minutes. If support agents handle more tickets but repeat the same subtle error to hundreds of customers, throughput has risen while quality has fallen.

For a first test, choose one team and one workflow. Write down what data may be used, what must never be entered, who reviews the output, and what happens when the system is unavailable. Train the users on examples from their real work rather than generic prompt tricks.

Run the test for 30 days. Compare time per item, correction rate, output volume, customer effect, and total cost against the baseline. Include subscription fees, setup, review time, training, and any additional software. A cheap tool that consumes management attention can be expensive.

Then make one of three decisions: adopt, adjust, or stop. Adoption means the measured gain survived after quality and risk were counted. Adjustment means the task or controls need another test. Stopping is not failure; it is a small, contained experiment preventing a larger waste.

The transferable principle is deliberately unglamorous. Do not buy artificial intelligence and search for a use. Find one expensive piece of friction, test AI against it, and let the evidence decide.