AI is the hammer, not the blacksmith

Treating AI as a tool does not mean pretending it cannot automate anything. It can already help people draft, classify, summarize, compare, translate, and surface patterns. That can remove repetitive steps and give people more time for the work that requires context, relationships, and judgment.

The distinction is between replacing a task and replacing responsibility for the outcome. The International Labour Organization's 2025 exposure index found that one in four workers globally is in an occupation with some generative AI exposure. But because most occupations contain tasks that still require human input, the authors describe job transformation as the more likely broad effect. Exposure is not the same as observed job loss, and an occupation is more than the easiest step to automate.

A better workplace question is not, “Can AI do this job?” It is, “Which part can AI accelerate, what must be checked, and who owns the final decision?”

One of the easiest AI failures to miss is a polished answer that sounds complete enough to skip the check. Clear language, confident tone, and a tidy citation list can create the feeling that the work has already been verified.

NIST's Generative AI Profile calls confidently stated false content confabulation. It also notes that fabricated logic or citations can make an answer look supported when it is not. NIST recommends fact-checking generated information and reviewing sources and citations, especially when information comes from multiple or unknown sources.

This matters in ordinary work. A meeting recap can assign an action to the wrong person. A research summary can flatten an important exception. A customer-service draft can invent a policy. A spreadsheet explanation can sound mathematically sound while reading the wrong column. The language may be smooth in every case; the underlying work can still be wrong.

When AI produces the first draft, human judgment moves downstream. Instead of spending all of the effort creating, the person must decide what to ask, whether the response fits the situation, what evidence supports it, what is missing, and whether it should be used at all.

In 2025, Microsoft Research and Carnegie Mellon researchers surveyed 319 knowledge workers about 936 self-reported examples of using generative AI at work. Greater confidence in AI's ability to complete a task was associated with less self-reported critical thinking. The same study found that critical thinking shifted toward verifying information, integrating the response, and overseeing the task. This was a cross-sectional, self-reported survey, so it does not prove that AI caused people to think less. It does show why a workflow cannot assume that judgment will happen automatically.

That is the adoption risk hiding behind convenience: when the first draft feels effortless, the review can start to feel optional. In a responsible workflow, that last step is where the professional earns trust.

Not every AI-assisted task needs a committee. The review should match the consequence of a mistake. That keeps lightweight uses useful without pretending that every output deserves the same level of trust.

The consequences depend on the workflow. In a legal filing, the failure can be immediate and public. In Mata v. Avianca, attorneys cited nonexistent judicial decisions generated by ChatGPT. The federal court did not treat the mere use of AI as the problem; it sanctioned the lawyers after they abandoned their gatekeeping duty, failed to verify the cases, and continued making misleading statements after the problem surfaced.

A customer-facing workflow creates a different version of the same accountability problem. In Moffatt v. Air Canada, an airline chatbot gave a passenger incorrect instructions about receiving a bereavement fare after travel. The British Columbia Civil Resolution Tribunal found that Air Canada had failed to take reasonable care to keep the chatbot accurate and held the company responsible for the information on its website. The decision does not identify that chatbot as generative AI, so it is not an example of an LLM hallucination. It is a useful reminder that an organization cannot outsource accountability to automated output simply because the interface delivered it.

The lesson is not to avoid AI. It is to place a check where an error becomes a consequence: before the filing, before the customer receives the answer, before the recommendation becomes a decision, and before the automated action becomes difficult to reverse.

A good workflow makes responsibility visible. People should know what the tool is allowed to do, what evidence it used, when a review is required, and who can stop or correct the result. That structure matters more than adding a vague “human in the loop” label after the system is already built.

For most professional work, the following loop is a strong place to start: