Delegating Work to AI: Stop Prompting, Start Asking

For the last few years, much of the conversation around generative AI has focused on prompting. Organizations built prompt libraries, trained employees on prompt frameworks, and encouraged people to improve the quality of their instructions to tools like Microsoft Copilot.

That was a necessary step. Better instructions generally produce better results.

But as AI capabilities mature, prompting is becoming only one part of the equation. The more important skill is becoming delegation.

Instead of asking, “How do I write a better prompt?” organizations should increasingly ask:

What work can we assign to AI?

That shift may seem subtle, but it represents a significant change in how AI fits into the workplace.

From Individual Tasks to Complete Outcomes

Most early Copilot usage has been task-based.

Summarize a document. Draft an email. Analyze a spreadsheet. Prepare meeting notes. Create a list of ideas.

These are valuable productivity improvements, but they typically involve one request followed by one response. The user remains responsible for moving the work from one step to the next.

The emerging model is different.

Consider a customer feedback process. Instead of asking AI to summarize a collection of comments, the assignment could be to review feedback from the previous quarter, identify recurring themes, highlight issues that appear to be increasing, create an executive summary, and recommend several actions for leadership.

The first request asks AI to perform a task.

The second defines an outcome.

Capabilities such as Copilot Cowork are accelerating this transition by allowing AI to work across multiple steps, information sources, and tools. The role of the user begins moving from directing every individual action to defining what needs to be accomplished.

That is much closer to delegation than prompting.

Better AI Requires Better Delegation

Delegation has always been a management skill, and many of the same principles apply when assigning work to AI.

A strong AI assignment starts with a clearly defined outcome.

“Analyze these project files” leaves significant room for interpretation.

“Determine why the last five projects exceeded planned hours and identify the three most likely causes” creates a much clearer objective.

Context matters as well. AI needs to understand the relevant time period, customers, audience, business objective, and information sources.

Boundaries are equally important. An assignment may specify that AI should not contact customers, alter source data, make assumptions when information is missing, or take an external action without approval.

Finally, the expected deliverable should be clear.

The result may be an executive summary, presentation, recommendation, customer communication, spreadsheet analysis, or action plan.

The more clearly the assignment describes success, the more effectively AI can contribute to producing it.

AI Training Needs to Evolve

This change also has implications for how organizations approach AI adoption.

Prompt training still has value, but teaching employees how to communicate with AI is no longer enough. Organizations should also help people identify work that can be delegated.

A useful place to start is with recurring business activity.

Which processes require employees to gather information from several sources? Which reports are recreated every week or month? Which tasks involve multiple applications, documents, or handoffs? Which analyses are delayed because assembling the information takes too much time?

Those questions move the conversation away from individual AI features and toward work design.

That is where larger productivity gains are likely to emerge.

Saving several minutes drafting an email is useful. Delegating a multi-step process that previously consumed several hours can fundamentally change how capacity is used.

Accountability Still Belongs to People

Greater delegation to AI does not eliminate human responsibility.

If AI prepares an analysis, recommendation, or customer communication, someone still needs to own the outcome.

Humans determine what good looks like. They decide which information can be trusted, where judgment is required, what risks are acceptable, and when approval is necessary.

This distinction will become increasingly important as AI moves beyond generating information and begins taking more actions on behalf of users.

AI can perform more of the work without becoming accountable for the business decision.

That responsibility remains with people.

My Take

I think we are spending too much time teaching people how to ask AI better questions and not enough time teaching them how to redesign work around it.

Prompting matters, but I do not believe the biggest gains from AI will come from perfecting the wording of individual requests. They will come from identifying complete pieces of work that can be responsibly delegated.

That means leaders need to start thinking differently about AI adoption.

Instead of measuring success by how many people are using Copilot, look at what changed because they are using it. Did a process get faster? Did a handoff disappear? Did someone get meaningful capacity back? Did the quality or consistency of an outcome improve?

The goal should not be to create better prompters. It should be to create an organization that understands how to divide work effectively between people and AI.

Moving Beyond the Prompt

Prompting is not disappearing. It will remain part of how people interact with AI.

But it is unlikely to remain the defining skill of enterprise AI.

Copilot introduced millions of users to AI assistance. Agents expanded the conversation toward AI taking actions. Experiences such as Cowork are pushing the model further toward AI completing broader pieces of work.

The next stage of AI adoption will not be measured by how many employees can write an impressive prompt.

It will be measured by how effectively organizations learn to define outcomes, establish boundaries, provide the right context, and decide where human judgment belongs.

The question is no longer simply: What should we ask AI? It is becoming: What work should we assign to it?