The Prompt Coach Agent for Microsoft 365 Copilot - Teaching People to Ask Better

The dirty secret of most Copilot rollouts is that the tool works fine and the people do not know how to use it. An organisation buys the licences, runs a launch email, maybe does a lunchtime demo, and then wonders three months later why usage is flat and the feedback is a shrug. The technology is not the problem. The problem is that most people type a vague half-sentence into Copilot, get a mediocre answer, conclude the tool is overhyped, and go back to doing things the old way. They never learned to ask well, and nobody taught them.

The Prompt Coach agent is Microsoft's attempt to address exactly this, and it is a more interesting piece of the puzzle than it first appears. It is a template agent for Microsoft 365 Copilot whose whole job is to help people write better prompts. Microsoft's documentation on the Prompt Coach agent template walks through what it is and how to deploy it. What I want to talk about is why prompting skill is the actual bottleneck for AI adoption, and how a tool like this fits into fixing it.

The real problem with Copilot adoption

Let me be blunt about something we see on nearly every engagement. The gap between people who get enormous value from Copilot and people who get almost none is rarely about intelligence or seniority or technical ability. It is about how they ask.

The person who gets value has, often without realising it, learned to give Copilot context, to be specific about what they want, to say what format the answer should be in, to give an example, to iterate when the first answer is not right. The person who gets no value types "write an email" and is disappointed the email is generic. Same tool, wildly different results, and the difference is entirely in the asking.

This is a skills gap, and it is the single biggest thing holding back return on the AI investment most organisations have already made. They have paid for the licences. The capability is sitting there. And it is being wasted because the people using it were handed a genuinely new way of working with zero training on how to work that way. We would never hand someone a complex new piece of software and expect them to be productive with no guidance, yet that is exactly what happens with Copilot constantly. The licence gets provisioned, the skill never does.

What the Prompt Coach agent does

The Prompt Coach agent tackles this at the point of use, which is the clever part. Rather than a training course people attend once and forget, it is an agent that sits inside the Copilot experience and helps people improve their prompts in the flow of actual work.

The idea is that someone can go to the Prompt Coach and get help constructing a better prompt for whatever they are trying to do. It coaches them on the elements of a good prompt: being clear about the goal, providing enough context, specifying the format and tone they want, giving examples where useful. Instead of just handing over a prompt to copy and paste, it teaches the underlying pattern, so people get better at asking over time rather than staying dependent on being given answers.

As a template, it is meant to be a starting point. You deploy it, and you can customise it to your organisation, adjust the guidance, tune it to the kinds of tasks your people actually do. It is built using the standard Copilot agent extensibility approach, so it fits into the same framework as other custom agents you might build. That matters, because it means Prompt Coach is not a walled-off gimmick, it is an example of the same agent-building capability you can use for other things once you understand it.

That in-the-flow, teach-the-pattern approach is genuinely the right instinct. The best time to learn how to ask Copilot something well is the moment you are trying to ask it something, not in a training room three weeks earlier when the example felt abstract and you were thinking about lunch.

How we think about this on client work

When we help an organisation roll out Copilot or broader AI capability, the prompting skills gap is one of the first things we plan for, because it is one of the first things that sinks these rollouts. And a tool like Prompt Coach is one piece of a larger answer, not the whole thing.

Something like Prompt Coach is useful because it provides continuous, in-context help rather than a one-off training hit. People forget most of what they learn in a course within a fortnight. A coaching agent that is there every time they use Copilot has a much better shot at actually changing behaviour, because it meets people at the moment they need it. This is a real strength and worth taking seriously.

But, and this is the honest bit, a coaching agent on its own does not fix adoption. It helps people who are already engaged and reaching for the tool. It does very little for the larger group who have quietly decided Copilot is not for them and have stopped opening it entirely. For that group you need the human side: showing them concrete examples relevant to their actual job, building the habit, giving them a reason to come back. A coaching tool amplifies engagement that already exists. It does not manufacture it from nothing. That distinction matters a lot when you are planning a rollout.

This is why our Copilot training is built around real workflows and real examples for the specific teams involved, not generic prompting theory. A finance person and a marketing person need to see Copilot doing their actual job, with their actual kind of task, before it clicks. A generic coaching agent helps, but the initial spark usually has to come from a human showing them something that makes them think "oh, that would genuinely save me an hour a week." Prompt Coach then keeps that momentum going once they are in.

Where it is still rough

Fair assessment, because it is early and template agents have their limits.

The quality of coaching depends heavily on how well the agent is set up and tuned to your context. Out of the box as a template, the guidance is general. To get real value you will want to customise it towards the tasks and language of your organisation, and that customisation is work. It is not a switch you flip and walk away from. Treat it as a starting point that needs shaping, not a finished product.

There is also a slightly awkward chicken and egg problem. The people who most need prompting help are often the ones least likely to go and seek out a prompt coaching agent of their own accord. The people who will happily use a Prompt Coach are frequently the ones already engaged enough that they were going to figure it out anyway. Getting the tool in front of the disengaged majority, at a moment they are actually receptive, is a real challenge that the tool itself does not solve. That is a change management problem, not a technology one, and no agent template makes it disappear.

And as with anything in the Copilot agent space right now, the capability is moving fast. Template agents like this are useful signposts for what is possible, but the specifics of how you build and deploy custom agents keep evolving, so anything you set up today should be treated as a living thing you will revisit, not a set-and-forget deployment. Build it expecting to maintain it.

The bigger picture

The Prompt Coach agent is a small thing that points at a big truth: the bottleneck for AI value in most organisations is human skill, not model capability. The models are already far more capable than the way most people use them. Closing that gap, teaching people to actually ask well and think of AI as a genuine collaborator rather than a slightly disappointing search box, is where the real return lives. And it is mostly a people-and-habits problem wearing a technology costume.

Tools like Prompt Coach help. So do custom agents built for specific workflows, which take the "you have to know how to ask" burden off people entirely by wrapping the good prompting into a purpose-built tool. That is often the better answer for a common, repeated task: rather than coaching everyone to prompt well for it, you build an agent that just does it properly, and the prompting skill is baked in where the user never has to see it. Deciding when to coach and when to build the agent instead is exactly the kind of judgement call our AI agent builders make on client work, because sometimes the right move is teaching people to fish and sometimes it is handing them the fish already cooked.

If you are further back and still working out where AI can genuinely help your organisation, before you even get to prompting, that is a strategy conversation, and it is what our business AI strategy work is for.

Where I would start

If you have Copilot deployed and adoption is flat, do not assume the answer is more licences or a sterner email. Look honestly at whether your people actually know how to use it. A tool like Prompt Coach, combined with proper role-specific training and a few genuinely useful custom agents for common tasks, will do far more for your return than anything else you can buy. The capability is already paid for. The skill is the missing piece.

If you are earlier and still rolling out, build the human side in from the start. Plan for the skills gap, because it is coming whether you plan for it or not, and Prompt Coach is a good addition to that plan rather than a substitute for it.

The Prompt Coach agent is a smart, honest acknowledgement from Microsoft that the tool is only as good as the person asking. If you want help turning a flat Copilot rollout into something people actually use and get value from, that is very much what we do. Have a look at our Copilot training or get in touch and we will work out what your teams actually need.