When to Use What

The decision nobody trained you for

Your licensing decision happens once. Your governance decision happens once a quarter. The mode decision happens every time someone opens a laptop and thinks “I need to get this done.”

That decision used to be trivial, because there was one answer: do it. Now there are six, they all look reasonable, and nobody has been given a rule. So people default — to the mode they discovered first, usually chat, and they stay there. That is why usage dashboards look busy while nothing in the business moves.

Most people don’t have a Copilot problem. They have a mode problem.

Why tool-first thinking fails

Enablement usually teaches tools: here is Copilot in Word, here is Copilot Chat, here is Copilot Studio, here is Cowork. Feature tourism. It produces people who can name five things and use one.

The failure is structural, not motivational. A tool-first frame asks “what can this do?” — an open question with no stopping point. A task-first frame asks “what am I trying to produce, and who has to be in the room while it happens?” — a closed question with an answer. Only the second one ends in a decision.

The gym version: nobody walks into a gym and asks what the leg press can do. You know what you came to train. The machine follows.

The Mode Switch: four questions

Question 1 — Do I need knowledge, or an artefact? Knowledge means you need to understand, compare, recall, decide. Artefact means something has to exist at the end: a document, a deck, a model, a message, a decision paper. Knowledge goes to chat or a reasoning agent. Artefact continues to question 2.

Question 2 — Do I already know what “good” looks like? If you cannot describe the finished result in two sentences, you are not producing yet — you are still exploring. Go and explore properly, then come back. Delegating an unclear outcome is the most expensive mistake in this whole guide.

Question 3 — Does this task repeat? Once is a task. Weekly, or done by four other people the same way, is a process. Processes get built once and run many times. That is an agent, and only that is an agent.

Question 4 — Do I have to be in it? Ask this about judgement, accountability and relationship, not about control. If your judgement shapes every paragraph, work alongside the AI inside the artefact. If your judgement only shapes the brief and the review, hand over the goal and check the result.

And underneath all four: does AI contribute anything at all? If the thinking is the work — the hard call, the difficult conversation, the four-line email that decides a relationship — do it yourself. If the thinking is the work, do it yourself.

The five modes, and the honest exit

Ask — Copilot Chat. For knowledge you don’t have or can’t remember: what did we decide, what does this policy say, summarise this thread, explain this contract clause. Cheap, instant, forgettable. The tell-tale sign you’re in the wrong mode: you have pasted the same context three times.

Explore — reasoning agents (Researcher, Analyst) and Notebooks. For questions that deserve more than one pass: market pictures, option comparisons, “what would this cost us”. Slower by design, sourced, and worth the wait exactly when the answer will be quoted in a decision. The tell-tale sign: you asked chat and got something plausible you wouldn’t dare put in front of your board.

Build with — Copilot inside the app. Word, Excel, PowerPoint, Outlook, Teams: the artefact exists, you are in it, the AI works next to you on the thing itself. This is the mode most enablement under-teaches and most professionals under-use, because it’s the least spectacular and the most productive. The tell-tale sign you should be here: you keep copying chat output into a document by hand.

Delegate — Copilot Cowork. For a one-off job that is genuinely long: gather this, compare that, produce the pack, prepare the follow-ups. You describe the outcome, it plans the steps, and it stops at checkpoints for your approval. Since general availability in June 2026, this runs on usage-based billing via Copilot Credits — so this mode has a visible price, which is a feature, not a flaw: it forces the question of whether the job was worth doing.

Automate — agents. Built once, used many times, ideally by more than one person. An agent is a small piece of software with a job description, an owner and a lifecycle. Building one for a task you’ll do twice is not automation, it’s a hobby. Which agents you’re allowed to build, and who owns them, is the governance question — that guide is linked below.

Do it yourself. Always on the board, never a failure. Some work is short enough, personal enough or consequential enough that the fastest route is your own hands.

Chat is where you ask. The app is where you build. The agent is where you repeat.

Three days, three roles

The project lead, Tuesday morning. Status report due Friday, same format every month. Chat would answer questions about it; the app would help write this one; but it repeats — so this is the agent. She builds it once with her template and her sources, and Friday stops being a task.

The account manager, Wednesday afternoon. Client review next week, twelve months of scattered history across mail, Teams and files. One-off, long, outcome describable (“a risk-ranked overview with the open items per topic”), and her judgement isn’t needed while it’s assembled — only afterwards. Delegate it. She reviews the result with the time she saved.

The CIO, Thursday. Board paper on the AI budget. The artefact matters, the reasoning is his, every sentence carries accountability. He explores first with a reasoning agent to get the numbers straight, then writes in the app with Copilot next to him — and the three-sentence paragraph that says what he actually recommends, he writes alone.

Four anti-patterns

Everything into chat. The most common one. Chat becomes a universal interface, artefacts get produced by copy-paste, and nothing is ever reusable. Symptom: long threads, short outputs.

Agents for one-off tasks. The most expensive one. Building takes an afternoon, maintenance takes forever, and the task happened twice. Symptom: an agent inventory nobody can name an owner for.

Co-working where delegation would do. The quietest one. People sit through work the AI could have finished while they did something else, because sitting there feels like control. Symptom: your calendar is full and your judgement was needed for nine minutes of it.

AI exactly where the thinking is the work. The most damaging one. The performance conversation, the strategic call, the apology. Symptom: it reads fine and means nothing.

What’s coming — and why the model doesn’t change

Microsoft has begun to add a further category: always-on agents with their own identity, acting continuously within the policies an admin sets. Microsoft Scout is the first of them, currently in preview through Microsoft’s Frontier programme, with broad availability not yet fixed.

Note what that changes in the four questions: nothing. It’s a stronger answer to question 4 — work that runs without you — with an admission price in governance. The modes are stable because they follow how work is shaped, not how products are named. That is the point of deciding by the task.

What this means for enablement

Teach the four questions before you teach a single feature. In practice, three things move the needle:

Put the decision on one page and hand it out — the tactic board and the cheatsheet below exist for exactly this. Make managers answer the mode question out loud in team meetings for two weeks (“which mode did you use for that?”), because the decision becomes a habit through repetition, not through a slide. And measure modes, not usage: a team that moved half its recurring work into agents has changed something; a team with a high chat count has not.

Self-check for your organisation: Can your people name the five modes? Can they say which one they used this morning, and why? Does anyone in your enablement material tell them when not to use AI? Do your agents have owners? Is anyone tracking what delegated work costs you — and what it produced?

If more than two answers are no, your problem isn’t adoption. It’s that nobody made the decision teachable.