Getting better results from Microsoft 365 Copilot often has less to do with finding a magical prompt and more to do with how you work with AI. Treat Copilot like a new teammate by providing context, giving feedback, and matching each task to the right AI tool or agent.
“Copilot doesn’t get me.”
“Copilot doesn’t understand my work.”
“Copilot isn’t providing me the answers/output that I need.”
“Copilot isn’t giving me the value that I had hoped for.”
I hear these and similar types of questions on a regular basis, multiple times a week at times. And after the individual on the other end of the call has finished sharing their observations, and often frustrations, I ask a few questions of my own:
- Tell me more about the prompt that you sent to Copilot. How much detail did you provide?
- How often did you iterate on the prompt with Copilot before you gave up?
- Did you try different Copilot models/agents to obtain different/better answers?
Too often the answers are, in order, “Not enough”, “Maybe twice”, and “No”. I am not surprised by these common answers.
AI, and the surrounding ads, have done us all a great disservice. AI has been raised to the level of a magical technology that can (quoted directly from Microsoft’s Copilot website) “analyze your professional world, from your files to your preferred habits.” Get Teams meeting “context in a flash” and bring “your ideas to life with Copilot in PowerPoint.”
Just type your natural language prompt and you are off to the races. Little, if any, time is spent in these ads about the need for prompt engineering, setting real context for Copilot or an agent.
We expect the technology to be able to sift through our digital work worlds. To sort through thousands of emails and chats, many that are split into multiple threads and topics. To filter and understand all our pack-ratted files that include outdated info and duplicates scattered across multiple repositories.
True, the AI can process this massive amount of information very quickly, but, due to the messy nature of our environments, can struggle to make sense of it all.
To top it off, early on, the AI doesn’t know you. It doesn’t understand your way of thinking. Your tone. Your style. Your intentions.
Do we seriously expect these tools to infer all of that from a messy inbox and OneDrive repository?
The Human Analogy: Why AI Needs Coaching Too
Here’s the irony in all of this. If you had just hired an intern or new team member, would you expect the same of them? Would you sit them down and point them at your organization’s SharePoint site and tell them to get up to speed and then to start creating content?
Early on, would you accept their first attempt at generating analysis or net new documents? Not without review, feedback, and an iteration or two.
With humans, we still (mostly) practice empathy for what someone new is stepping into. We teach them our ways. We mentor them through their challenges. We invest time to make them better.
Why do we not take the same approach with AI?
We must accept that our work lives are much more complex than our personal lives. We are not just asking for a restaurant or vacation recommendation that can be put into context with a favorite cuisine or vacation lifestyle.
We are dealing with years, if not decades, of institutional knowledge and behavior. Yes, the AI can learn much faster and reference information loads that would initially overwhelm a human, but the AI still needs your nurturing to make it the most effective.
AI Onboarding Works Like Human Onboarding:
Treat Copilot like a new hire. It needs context, feedback, and iteration, and it improves with coaching.
Three Questions That Change Everything
Let’s revisit those three questions that I typically ask to make these recommendations real and actionable.
1. Prompt Quality: Are You Giving Enough Context?
What to Do: Be verbose in your prompt
Give the agent some background and context before asking for help.
Yes, you could just point it at a meeting recording or email chain but giving the AI your perspective and opinion matters. The AI can use that extra information to frame its output.
If you are optimistic about a challenge, it will do the same. If you are cautious and need convincing, the agent can provide additional context to its answers and provide more robust arguments.
If you are executing a complex financial analysis, explain what you want from the agent. Explain how our organization budgets and manages materiality and variance thresholds. Provide it with the lingo that you typically use in an analysis by either explaining it or pointing it to a prior analysis spreadsheet for reference.
Don’t just expect it to find that information on its own. Best case, it does. Worst case, it finds an outdated and incorrect analysis that you had forgotten about.
The good news is that the AI will learn. It will remember. I have been impressed with Copilot’s ability to make connections across my work patterns…now that I have taught it about our practice, offerings and approaches.
The upfront investment will pay off, and likely faster than if you were coaching a human.
Common Mistake: Assuming Copilot will “just know” your business context without being told.
2. Iteration: Are You Actually Coaching the AI?
What to Do: Don’t stop at the first answer
Sometimes your initial prompting efforts fall short, no matter how much context you give. Maybe your phrasing was off. Maybe, what you thought was enough context, still wasn’t enough. Maybe you accidentally introduced unintended bias or didn’t challenge the agent enough to validate its output.
Regardless of the cause, you are not satisfied with the output.
What would we do with our human intern? We (hopefully) would coach them, explaining where expectations were not met and how the output can be improved.
Do the same with your AI. Converse with it. Tell it what you expected and ask it why it didn’t meet the expectations, and you may need to do this multiple times.
The output will improve, and the speed of seeing the change will be impressive.
A Technical Note on Context Windows
Sometimes an agent’s context window (its ability to keep order across iterative thoughts in the same chat) can become too big and the agent can seem to lose its place.
If that happens, take the last output and start a new chat. Reference what has been done, upload the latest output and continue. This helps the agent reset and provide crisper answers.
A Technical Note on Context Windows
| Issue | What Happens | Fix |
|---|---|---|
| Context window overload | AI loses track of earlier instructions | Start a new chat |
| Long iterative threads | Responses become inconsistent | Re-anchor with latest output |
| Conflicting instructions | Output quality degrades | Reset and restate goal clearly |
3. Tool Selection: Are You Using the Right AI “Team Member”?
What to Do: Match the tool to the task
Skills and capabilities differ across individuals and teams, right? We learn, over time, to use the right team with the right skills for the right task at hand.
When something new comes up or a new person joins a team, time is spent understanding the need and capabilities of the new team member. Why do we not use the same approach with AI?
I think that, once again, the advertising around AI is failing us. Copilot was portrayed as a one-stop shop for all your needs. But not enough time was spent highlighting that Copilot is an interconnected set of tools.
Yes, you have Copilot chat for general questions and tasks, but then you have Cowork, albeit for an extra consumption fee, to execute complex, iterative, and multi-threaded tasks.
Also, if you want to not spend more than your license fee, you have agents such as Researcher and Analyst to dig deeper into topics and help you with numerical analysis respectively.
The PowerPoint, Word and Excel agents help you create content. Copilot embedded in those same applications can create and edit content, especially when using the Anthropic Claude models.
Then there are other agents, both out-of-box and third party that do specific bidding. Finally, you can introduce custom agents into the mix, grounded in specific knowledge and empowered with your custom actions and skills.
Each of these tools behaves a bit differently and has its own strengths and weaknesses. It is beyond the scope of this post to cover them all, but if your results are not what you need in one tool, try another.
Over time, you will find the right tool for the right task, and with expanded model availability and continued innovations, it may be worth revisiting tools over time to see how they have evolved. Just like you would a team and its growing team members.
Is this as seamless as we would like? No. It takes effort, but the time spent will likely still be less than you performing the work manually.
Even if it does not, trying out new ways of working is a worthwhile investment and learning experience.
Closing Thought: Treat AI Like a Junior Teammate
So, here we are. Surrounded by so much AI functionality, some of which we know and some that may be completely new.
Maybe the very human-based mindset outlined above will allow us to better connect with our AI “team” and understand “who” has what skills.
The next time Copilot disappoints you, ask yourself the same question you would ask about a new employee: Did I give it enough context, feedback, and coaching to succeed?
Want to Go Further?
Want to learn more about Copilot and agentic AI? Our Copilot deployment and adoption engagements make sure that you and others are using Copilot to its fullest potential.
Want to understand how your business needs can be solved with AI? Our AI use case and roadmap exercise will match your needs to the right tools.
If you just want to discuss your broader AI visions and goals and how to become a Frontier firm, reach out and we’ll discuss strategy and tactics. I look forward to hearing from you
