AI at work usually means a chat window that knows nothing about the business it is being asked about. It answers well in general terms, and every piece of context has to be typed or pasted in first.

Microsoft 365 Copilot works the other way round. It sits inside Outlook, Teams, Word, Excel and PowerPoint, and it can reach into the emails, files, meetings and chats an employee is already permitted to open. Ask it about a client and it can find the thread, the meeting notes and the renewal document itself.

That is the whole proposition. Everything below is what falls out of it.

Copilot Chat is the part that finds things

The chat window is where Copilot behaves least like a chatbot and most like someone who knows where everything is filed.

• Retrieving documents without knowing where they live. Ask for a document by name or by topic and it pulls it from SharePoint, provided the person asking already has permission to open it. No folder archaeology.

• Pulling a client history out of scattered emails and files. The record of a relationship is usually spread across an inbox, a shared drive and three meetings nobody wrote up. Copilot can assemble it into one answer.

• Retrieving what was said in a meeting or client call held in Teams, weeks after the fact.

• Comparing two documents and reporting what materially changed between them. Two versions of a supplier proposal, two drafts of a contract.

• Drafting policy and compliance documents by pulling together existing material and meeting notes into a first version. The AI drafts, a person reviews and signs off.

In Outlook it works on the inbox

• Summarising long threads. A twenty-message chain about a delivery date compressed into what was agreed, what was asked, and what nobody has answered.

• Drafting replies for review. It writes, a person checks and sends.

• Tightening tone. There is also a personalisation feature that studies previously sent mail to learn how someone actually writes, so drafts come back sounding like them rather than like a machine. It needs switching on.

In Teams it handles the meeting

• Summarising the meeting once it ends, including what was decided.

• Identifying and listing action items, which is the part that otherwise depends on somebody volunteering to take notes.

• Catching up mid-meeting. Someone joining twenty minutes late can ask for a summary of the discussion up to the point they arrived, and rejoin properly rather than guessing.

In Word it handles the document

• Turning rough notes or a meeting transcript into a first draft. Scribbled discovery notes become a structured proposal.

• Reviewing an existing document and flagging unclear wording, repetition and missing information. This is the most dependable use of the lot, because the person supplies the substance and Copilot only inspects it.

• Adjusting structure, reordering and reshaping a document that says the right things in the wrong sequence.

In Excel it answers questions about the numbers

• Asking in plain English. "Which clients are approaching renewal and have not been contacted yet?" No formula, no pivot table.

• Turning data into charts without anyone hunting through menus.

• Spotting values that look wrong. A client premium sitting unusually high or low against the rest of the book is the sort of thing that hides in a spreadsheet for a year.

Why the context changes what the answer is worth

A generic AI asked to prepare someone for a client meeting produces a competent, generic meeting-preparation template, because a template is all it has.

Copilot can answer the same question from the actual email thread, the last call, and the file someone amended on Thursday. The first is advice about how to work. The second is the work.

This is also why the quality of the answer depends on the state of the business. Copilot is reading real company material, so well-organised files and clear records produce sharper answers than a shared drive nobody has tidied since 2023.

What a business gets beyond the time saved

It brings shadow AI usage in-house. Staff are already pasting client details into free chatbots. Providing a sanctioned tool moves that activity inside the tenant, under existing controls.

Administrators get an audit trail of what is being asked and what is being accessed.

It is permission-aware by design. Copilot surfaces only what the signed-in person can already open. An intern cannot ask what the owner emailed John, or what anyone is paid, because those files were never theirs to read.

Adoption becomes visible. Usage can be seen down to the individual, so it is possible to tell who has taken to it and who was licensed and never logged in.

It future-proofs the business. More capable AI is arriving, including automation and agents that carry out tasks rather than draft them. A business already running Copilot inherits that inside software it owns, without buying a new platform or retraining anyone.

What it will not do

It is not a source of truth. It is grounded in company data, so an answer is only as current and accurate as what is sitting inside the business. Anything going to a client gets checked first.

It can only see what the employee can already see. Where permissions are poorly organised, Copilot cannot reach the material. The flip side deserves saying: it also enforces permissions that already exist, so a folder shared too widely in 2023 becomes discoverable by asking a question in plain English. Tidy the permissions before turning it on.

Transcription has to be running before a meeting starts. Copilot cannot summarise a conversation it never attended, and finding that out after the meeting that mattered is a common first disappointment. It can be set on by default.

The same prompt can produce different answers. No single response can be treated as fixed truth, which makes human review a normal step rather than an occasional precaution.

It is not reliable for financial or legal matters. It can draft and summarise. It cannot make the call. Copilot is not a substitute for professional judgment, and it reduces effort without transferring accountability.

GCSE: the four parts of a prompt that works

The gap between a disappointing answer and a useful one is almost always the prompt. Four elements fix most of it, and they spell something easy to remember.

• Goal. A clear, specific instruction. "Draft a reply confirming Margaret Ellis's renewal date."

• Context. Why it is needed and who it is for. "She is a long-standing client and is a bit anxious about the price going up."

• Source. What it should work from. "Use our recent email thread with Margaret, and read the attached policy renewal document."

• Expectation. What the result should look like. "Keep it warm and reassuring, and under 150 words."

Run together, that is one instruction a person could speak aloud in fifteen seconds, and it produces a draft worth editing rather than a draft worth deleting.

Weak answers are usually missing two of the four, and the one most often left out is source. Naming the thread and the document is the difference between Copilot finding the right material and guessing at it.

The place to start is a task already done every week, chosen because the person doing it will know immediately whether the answer is right. A task where nobody can judge the output is the worst possible place to begin.

Sources: Microsoft Learn documentation on Microsoft 365 Copilot capabilities, data grounding and permissions, September 2026. Capabilities vary by licence and by how each organisation has configured its tenant.