The Copilot
app itself.

The three-mode mental model

Work mode — grounds on your Microsoft Graph only. Use when the answer lives entirely in your own M365 tenant: your files, emails, meetings, chats, SharePoint, Teams. Fastest, most relevant for "what have I already got?" questions. Examples: "What's on my calendar this week?", "Summarise what Sarah and I have discussed about the Harrow contract", "Find the last draft of the stakeholder mapping document".

Web mode — grounds on public web only (with the Bing search service). Use when the answer lives entirely outside your tenant — external frameworks, public guidance, industry reports, general knowledge. Examples: "What does PAS 2080 say about scope 3 emissions?", "Summarise the latest ORR guidance on rail accessibility", "What are the key differences between NEC3 and NEC4?".

All (both) mode — grounds on Microsoft Graph AND the web. Use when the answer requires combining internal context with external reference material. Examples: "How does the approach in our last Harrow report compare with best practice in recent ORR guidance?", "Synthesise our internal lessons-learned with external industry reports on the same topic", "What would a reasonable next step be given our scope and the latest regulatory changes?".

The common failure — wrong mode for the question

Two consistent mistakes advisory consultants make:

• Asking Work questions in Web mode — "what's my calendar?" with Web grounding returns generic advice about calendar management, not your calendar. Frustrating, obvious in hindsight, happens constantly until you build the habit.
• Asking Web questions in Work mode — "what does PAS 2080 say?" with Work grounding searches your own files for references to PAS 2080. If you've never downloaded a copy, Copilot has nothing to work with and invents plausibly-sounding but wrong content.

Semantic search — what it actually does differently

Traditional SharePoint / Outlook search is keyword-based — it finds files containing the exact words you typed. Copilot Search uses a semantic index (vectorised representations of Microsoft Graph content) that understands relationships, synonyms, and meaning. Queries like "our Q3 revenue performance" find relevant financial reports, meeting notes discussing revenue, and email threads about Q3 targets — even if those exact words don't appear in the documents. This is the shift from "find containing" to "find about."

Six natural-language query patterns that work

Data-source filtering — use the right panel

Copilot Search results show an data source filter in the right-hand panel letting you narrow to specific systems: SharePoint, Teams, Outlook Emails, OneDrive, Viva Learning, Power BI, or any third-party system connected via Copilot connectors (Salesforce, ServiceNow, Confluence, etc.). Use it when a broad search returns too many cross-source results — filter to the one system you think the document lives in, refine from there.

Copilot Answer at the top

For certain natural-language queries that look more like questions than searches, Copilot Search shows a Copilot Answer at the top — a concise AI-generated response that may cite files from your tenant, web sources, or admin-curated answers your organisation has set up. Treat Copilot Answers as a quick orientation; the ranked files below are still the authoritative source material for anything that matters.

Search → Chat handoff

The natural workflow: find with Search, synthesise with Chat. Once Search returns the 3-5 files that matter, open them (or copy their references) and switch to Chat to ask "summarise across these four documents and tell me where they agree and disagree". Search does retrieval; Chat does reasoning. Use each for what it's designed for.

1. Brainstorming. The classic voice-natural task. Walking to a meeting, you have a half-formed idea. Start voice chat. "I'm thinking about how to structure the next phase of the Harrow engagement. We have three workstreams — digital, organisational, and commercial — and I'm trying to figure out the sequencing. Walk me through some options and help me pressure-test them." Copilot asks clarifying questions; you respond; it suggests structures; you interrupt to redirect. By the time you arrive, you have a shape. Transcript saves for the desktop write-up later.

2. Rehearsing a pitch or difficult conversation. "I need to have a difficult conversation with my client about a scope change. Let me explain the situation; then you play the role of the client and challenge me; I'll practice the response." Works surprisingly well — voice chat's conversational mode makes Copilot a competent sparring partner. Particularly valuable for pre-meeting prep on high-stakes calls.

3. Thinking through a problem. "I'm trying to work out why our stakeholder mapping isn't landing with the client. Let me talk through what I've done so far; then help me identify what might be missing or where I've misread the room." Externalising thought via voice is structurally different from typing — it's faster, looser, more associative. Copilot as thinking partner rather than assistant.

4. Voice catch-up in Outlook mobile. A specific feature that deserves its own shout. Open Outlook mobile → Copilot → Voice catch-up. Copilot summarises unread emails aloud and guides you through actions — draft reply, archive, pin, flag — all by voice. Perfect during a walk when you don't want to read but want to clear the inbox. "Yes, archive that one. Draft a reply to Sarah's saying I'll review by Friday. Skip the next one."

5. Memory-assisted follow-up. Because voice chats reference your Personalisation settings memories (January 2026 onwards), you can ask things like "what was the last action item I had on the Harrow engagement?" and Copilot pulls from stored memory rather than starting fresh. The conversation feels continuous across sessions. For the pattern to work, populate memories deliberately — see Section 4, prerequisite 4.

Voice etiquette — practical notes

• Reduce background noise when you want accuracy. Wind and traffic degrade transcription quality significantly.
• Interrupt freely. Copilot handles interruption well — it stops, listens, responds to what you just said. Don't wait politely for it to finish.
• Adjust pace by asking. "Speak faster," "slower," "more energetic," "more concise" — all work. Copilot adapts for the rest of the session.
• End the chat explicitly. Tap End voice chat when done. Otherwise the session stays open and counts against your daily voice capacity.
• Review the transcript on desktop. Voice chats appear in your Copilot Chat history — pick up the next day on desktop to refine the thinking into actual output.

Five desk-bound use patterns that genuinely work

1. The coding or formula lookup. You're in Excel writing a complex DAX measure or Python script in VS Code. You need to remember the syntax for something. Instead of alt-tabbing to a browser: "Hey Copilot — what's the DAX syntax for a rolling 12-month sum?" Copilot answers while your hands stay on the keyboard. You copy the syntax verbally or from the on-screen text.

2. The fact-check. You're writing a briefing and want to verify a statistic before including it. "Hey Copilot — is the UK's current BIM Level 2 mandate still in force as of 2026? Anything changed recently?" You get a verbal answer; if it's uncertain, you dig deeper in Chat. Quick fact-check without losing your writing flow.

3. The concept refresh. You're reviewing a document and hit a term or concept you only partly understand. "Hey Copilot — quickly remind me what PAS 2080 is and how it differs from ISO 19650." Thirty-second verbal briefing; back to the document.

4. The timer / reminder. "Hey Copilot — remind me in 20 minutes to check the oven," or "— in 45 minutes I have a client call; alert me at 10 minutes before." Useful for pomodoro-style work, calls, meetings you don't want to miss.

5. The parallel prompt. You're in a focused writing session and an unrelated idea occurs to you. Rather than breaking flow to capture it: "Hey Copilot — remind me later I need to think about the stakeholder mapping approach for the Leeds engagement." The memory gets captured; you continue writing.

Where it breaks down

• Open-plan offices. Voice carries. Colleagues don't need to hear your Copilot queries. Either use headphones with a microphone or disable the feature when working in shared spaces.
• Client calls and meetings. Mute your microphone or disable Hey Copilot before joining. You don't want "Hey Copilot" to trigger during a Teams call.
• False wakes. Saying "hey copper" or similar phrases occasionally triggers activation. Manageable but worth knowing — check whether Copilot is listening before assuming your speech was private.
• Screen-locked state. Doesn't work when the PC is locked, which is correct behaviour but sometimes unexpected.

Three dictation patterns that work well

1. The post-meeting brain-dump. Meeting just ended; you have twenty things in your head. Rather than typing them in order, dictate a messy prompt: "I just finished a meeting with the Harrow client. Three things came up that I need to think through. First, they want to accelerate the timeline by four weeks, which I don't think is realistic. Second, their IT director raised concerns about data sovereignty that we hadn't heard before. Third, the commercial structure we'd assumed needs to change. Structure these into a short note I can share with the team, with recommended next steps for each." Copilot produces a clean structured output from your verbal mess.

2. The long-form briefing request. You need a substantial briefing — two or three pages on a topic. Typing the prompt takes five minutes. Dictating it takes ninety seconds. "I need a team briefing on how UK transport authorities are implementing digital asset management. Cover: the regulatory context as of 2026, three specific authorities doing this well, common implementation patterns, typical pitfalls, and what advisory support they tend to need. Cite specific sources. Length around two pages. Professional tone, for an Arup audience." Faster to dictate the brief than to type it; the output quality is identical.

3. The accessibility case. For consultants with RSI, dyslexia affecting typing speed, or any other reason typing is slower or more painful than speaking, dictation is the primary input method rather than a convenience. The speech-to-text is genuinely good and improves with clearer voice work (see below).

Practical dictation notes

• Speak clearly but naturally — you don't need to enunciate like a presenter. Normal conversational speech works.
• Pauses help with punctuation. Natural phrase breaks usually get punctuated correctly.
• Say "new paragraph" or "new line" if you want explicit breaks.
• Proper nouns trip it up — project names, technical acronyms, personal names. Review the transcript before submitting for these.
• Background noise matters less than you'd think. Reasonable environments work.
• Nothing is stored server-side beyond the completed prompt. Audio is discarded after transcription.

Dictation vs voice chat — the distinction

Two different modes, easy to confuse:

• Dictation — speak, Copilot transcribes into the prompt box, you review and submit, Copilot responds in text. One-way input.
• Voice chat — speak, Copilot responds in speech, back-and-forth conversation. Two-way.

Use dictation when you want a careful prose response you'll read. Use voice chat when you want a conversation you'll listen to.

The pattern behind why these work

Notice what all three scheduled prompts have in common:

• They pull from work data (calendar, emails, chats), which means they require the Premium tier to work well
• They produce a single consolidated briefing rather than a list of answers — format matters for scannability
• They run at moments of natural transition — start of day, end of week, start of week. This is when consultants naturally take stock, and Copilot doing the preparation saves the cognitive work
• They're triggers, not endpoints — reading the briefing doesn't close anything; it sets up the work you'll do next

Other scheduled prompts worth considering

• Weekly project status summary — every Monday, summarise activity on a specific client engagement Teams channel
• Monthly industry scan — once a month, a web-grounded prompt on a sector you advise in, returning the month's most significant developments
• Quarterly lessons-learned trigger — every quarter, a prompt reviewing which engagements closed and asking you to reflect on what worked
• Daily 4pm wind-down — before logging off, what didn't get done today that matters?

Pruning the garden

Scheduled prompts accumulate. If you build ten over the year and only read three of them, the other seven are noise. Schedule a quarterly review of scheduled prompts in your own calendar — check which you're actually using, which you skim past, which you've stopped finding useful, and retire the rest. Good scheduled-prompt discipline is half building them and half removing the ones that no longer serve.