The hidden inefficiency in everyday AI use

I see the same pattern regularly. Someone needs a quick answer, so they enter a short prompt into Copilot. The first response is not quite right. They add more detail, ask another question, change the format and try again. They eventually reach the answer they wanted, but only after several attempts.

The issue is not that Copilot failed. The issue is that the user had not clearly explained the goal, context, source information or expected output at the start.

As adoption grows, small inefficiencies become operational habits. A few extra prompts may not seem important for one person. Across hundreds or thousands of users, the organisation may be paying for the technology while still accepting avoidable rework, inconsistent outputs and limited reuse of what people have learned.

High prompt volumes can look like strong adoption but can also indicate that users are struggling. A usage dashboard may show that people are active, but activity alone does not tell us whether Copilot is helping them complete work.

This changes the questions technology leaders should ask. Instead of asking the obvious questions such as: How many prompts were submitted or how many users opened Copilot or how many licenses were assigned? Technology leaders need to ask I more critical questions that reveal whether Copilot is genuinely improving how work gets done. These questions matter because activity does not always equal value. Critical questioning helps separate busy usage from meaningful adoption, showing whether Copilot is reducing effort, improving consistency and helping people produce better work faster.

How this becomes more relevant when developing Copilot Agents

Copilot Studio agent instructions influence which tools and knowledge sources are selected, how tool inputs are completed and how responses are generated. This is why I see an agent’s instructions as part of the application design, not an optional description box.

Consider an IT service desk agent. A loose instruction such as “Assist service desk users in solving technical IT issues” leaves too much undefined.

The agent needs to know:

• Which knowledge sources are authoritative
• What questions it must ask
• Which diagnostic sequence to follow
• When it should stop
• When it must escalate
• What tools it can call
• What the final response should contain
• What it must never do

An IT support agent should not rely on every engineer inventing the right question while a user waits. It should provide approved starting prompts for common scenarios such as account access, device enrolment, connectivity or application faults. It should collect the required facts in a consistent order and return a response that clearly separates diagnosis, recommended action, evidence and escalation.

The impact of inefficient prompts

Repeated prompting does more than consume user time, it could potentially also increase cost. Microsoft’s current Copilot Studio documentation states that a classic answer uses one Copilot Credit, a generative answer uses two, an agent action uses five, and tenant Graph grounding uses ten. Each prompt from a pay-as-you-go user may therefore create an additional charge. Helping users write better prompts can reduce both wasted time and avoidable cost.

What’s the fix?

1. Treat prompting as a core AI skill

Provide short, role-based training using real tasks. Teach goal, context, source and expectations. Include source checking and response verification.

2. Build a governed prompt library

When someone develops a prompt that consistently works, that knowledge should not remain in their chat history. Microsoft 365 Copilot supports saving prompts for later use, and the Copilot Prompt Gallery allows users to explore, save and share prompts with colleagues.

A prompt library can prevent different teams from repeatedly solving the same problem. It can also help new users start with proven examples rather than a blank prompt box. The organisation begins to capture not only what people know, but how they ask AI to apply that knowledge.

Start with a small number of high-value scenarios. Give each prompt an owner, purpose, audience, approved sources and review date. Retire prompts that are outdated or unused.

3. Use suggested prompts in Agents

When building agents, give users clear, commonly used prompts as starting points. This helps them provide the right details upfront, improves the agent’s response, and saves time because users do not need to work out what to ask.

4. Measure task outcomes

Track completion time, retries, manual rework, quality, user confidence and business KPIs. Prompt volumes and active-user counts are supporting measures, not the final outcome.

‍