An AI tool can finish a sentence long before you have decided whether it should be involved. But producing an answer is different from taking responsibility for it. Before handing over a workflow, ask three questions in order: Does the task repeat? Is there a reliable source of truth? Is a wrong output cheap enough to catch and correct?
This editorial test is not vendor guidance. It is deliberately narrower than “What can AI do?” Pew Research Center found in June 2026 that 42% of U.S. adults use chatbots to search for information, while 38% of employed adults use them for work tasks. Those figures describe use, not successful automation.
Run the test before choosing a tool
First, look for repetition. A chore that arrives in roughly the same form each week gives you a chance to define what goes in, what should come out and what counts as an exception. Deciding how to handle a new conflict between colleagues does not become routine because it happens at work. If each instance demands fresh context and judgment, stop here: AI may help prepare material, but the decision should not be automated.
Next, identify the authority the output can be checked against. Approved opening hours, an original document or notes from the meeting offer something concrete to compare with a draft. A chatbot’s confident answer is not itself that source. If you cannot point to a current, accessible record, you have no dependable way to tell whether the output is right. Create or find the source before building the shortcut.
Finally, price the mistake. Could someone notice and repair an error before it matters? A misgrouped item in a private planning list is usually easier to recover from than an incorrect promise sent to a customer. Where a wrong answer could change a health, legal or financial decision, use AI to organize questions, then put the decision and its verification with the appropriate person or primary document.
Passing all three questions does not mean “set and forget.” It means the task may support a bounded first pass, followed by a named human check. Failing one question is useful information too: narrow the task until it becomes checkable, or leave it with a person.
The same tool, two different calls
Consider a neighborhood club that sends a weekly email about upcoming events. Its coordinator has an approved calendar with dates, locations and registration links. Turning those entries into a draft email repeats; the calendar is the source of truth; and the coordinator can compare every detail before sending. An AI draft could be worth trying here, provided it uses the approved material and the coordinator checks the links and wording. The automation is the first version, not the act of publishing it.
Now consider a member who writes that an event cancellation caused a serious personal problem and asks for an exception to the refund policy. The coordinator may see refund requests often, and a policy document exists. Yet the case depends on circumstances the policy alone may not resolve. A mistaken refusal could damage trust; a mistaken promise could create an obligation. The tool might summarize the message privately or surface the relevant policy passage if the information can be handled appropriately. It should not decide the exception or send the response on its own.
The distinction is not writing versus customer service. The answer emerges only after the questions are applied in sequence. The weekly email would fail if the calendar were outdated or a draft went out unchecked.
Keep a human checkpoint that does real work
Microsoft’s 2026 Work Trend Index, a vendor study surveying 20,000 AI-using workers in 10 countries, describes a useful habit among its “Frontier Professionals.” In that surveyed group, 53% said they pause before starting to decide what AI should do and what a human should do, compared with 33% of other AI users. Some 43% said they deliberately work without AI sometimes to keep their skills sharp, compared with 30% of others. These are self-reported differences within Microsoft’s sample, not proof that the habit causes better performance.
The same survey found that 86% of AI users treat output as a starting point. Quality control of AI output, named by 50%, and critical thinking, named by 46%, led its list of important human skills. That is a better model for a small automation than silently accepting whatever arrives: decide the handoff, inspect the result against the source, and retain the ability to do the job when the tool is absent.
A check should have a specific job. For the club email, the coordinator verifies each event against the calendar and reads the message for confusing or misleading language. If checking every line takes longer than writing it, the scope is too broad. Perhaps the tool should only format calendar entries into a draft, leaving the introduction and any unusual announcement to the coordinator.
Watch what the shortcut requires
Over-automation often starts with a review step that is nominal rather than real: a person clicks approve without the source in front of them. Another failure is moving sensitive information into an unvetted service just to make a workflow possible. OECD research identifies skills, cost, time, privacy and security, and regulation as barriers to AI adoption. Its SME digital-security work also flags unauthorized “shadow AI” use when tools connect to company data.
A pilot can stay small. Try a few examples using information you are allowed to use, compare results with the source, and count the setup, correction and review as part of the work. Keep an exception route for anything ambiguous. If the tool needs access that has not been approved, or if nobody can own the final check, the process is not ready to run.
The practical filter remains short: Does it repeat? Can you check it against a reliable source? Can you absorb and correct a mistake before it causes harm? A task that passes can earn a supervised first pass. A task that fails should stay human-led, even if AI can produce a plausible answer.
