GenAI at Work: A Tool Unlike Its Predecessors

The real change

Organizations have brought in plenty of new tools: email, spreadsheets, cloud services, phones, and collaboration platforms. Each required learning, process changes, and sometimes a fair amount of frustration.

Generative AI asks for a different kind of preparation. It produces language, images, summaries, and recommendations that can sound useful before anyone has checked whether they are useful. Its output depends on the task, the context supplied, the system configuration, and the data it can reach. That makes it a poor fit for the old rollout message: “Here is the tool. Please start using it.”

The question is not whether employees should trust an AI system. The question is what work it is approved to help with, what evidence a person must check, and what happens when the tool is wrong.

NIST’s Generative AI Profile is a useful reminder that generative AI can introduce or amplify risks across its lifecycle; it is not a substitute for a general-purpose technology rollout checklist. (NIST AI 600-1)

Give people an operating contract

Before a team is asked to experiment, leaders should be able to answer a few plain questions:

• Which tools and accounts are approved for this work?
• What information may be entered, and what information may never be entered?
• Which tasks are reasonable pilots, and which decisions must remain with a person?
• What must be checked before AI-generated material is sent, published, or used to make a decision?
• Where does an employee take a concern, an unexpected result, or a useful lesson?

That is an operating contract. It gives employees a way to use judgment instead of asking them to guess the boundaries. It also makes it possible to disagree with a use case without being labeled resistant to change.

Start with work, not with a tool tour

People learn new systems best when the practice is connected to work they actually need to do. A team might try a low-consequence task such as turning its own meeting notes into a draft agenda, preparing a first draft of a routine communication, or organizing questions for a research task. The team should know what a good result looks like and who will review it.

The pilot is not an audition for the employee. It is a test of the workflow. If the output takes longer to check than it saves, creates confusing handoffs, or cannot be safely used with the available information, that is useful evidence. The organization can narrow the use case, change the controls, or stop the experiment.

Managers have a different job now

Managers do not need to become prompt engineers before they can lead responsibly. They do need to make the operating contract real in the daily work:

• explain the purpose and limits of a pilot;
• make room for questions about accuracy, privacy, authorship, and workload;
• distinguish a bad output from a bad employee;
• record recurring failures and decide whether the process, the data, or the use case needs to change.

The NIST AI RMF’s discussion of human-AI interaction makes a related point: organizations need to define the different human roles and responsibilities around an AI system. (NIST AI RMF)

Measure the work, not the enthusiasm

An adoption count can show that people opened a tool. It cannot tell a leader whether the work became safer, clearer, faster, or harder to review.

For a small pilot, a team can look at a few concrete questions:

• Did the task take less time once review was included?
• Were sources, calculations, and important statements checked before use?
• Did the output create rework or reduce it?
• Did the pilot expose a policy or training gap?

The answers do not need to become an employee surveillance program. They should help the team decide what to keep, change, or discontinue.

The invitation

GenAI does not need to be treated as magic, a colleague, or an enemy. It is a system that can be useful in some work under clear conditions and a poor fit in others.

Leaders earn trust by naming those conditions, giving people a safe way to question them, and changing course when the evidence calls for it. That is slower than an adoption mandate. It is also a more credible way to learn what the organization can responsibly use.