Foundations
& Tools
Understand the stack, know your licence, learn the rules, see the patterns — Then use the right tool.
Understand the stack, know your licence, learn the rules, see the patterns — Then use the right tool.
Five things worth knowing before you choose a tool.
The five tabs below provide everything required to begin using Copilot at Arup. Confirm your access and locate the relevant Arup policies in First Steps; identify the appropriate tool for any task in the Decision Tree; review all 12 AI tools with usage guidance in the Toolkit; view the tools advisory teams use most frequently in Most Used; and explore governance, the AI stack and agent patterns in Understand Better.
Confirm which Microsoft 365 Copilot tier is enabled for your account before you start using Copilot workflows.
Wording and visibility vary depending on rollout, organisational policy, and which apps are enabled for your account.
Microsoft offers three licence tiers:
Confirm your tier in m365.cloud.microsoft before relying on a Premium-only feature in client work — assignment can vary per app.
Arup-internal policies, training and guidance referenced throughout this playbook. Sign-in required.
SharePoint PDF — Arup's cyber-security policy and practical guidance for handling client data.
Open in SharePointArup Group's official AI Policy — ethical, safe and compliant AI use across the firm. Read before deploying AI on client work.
Open policyMoodle course covering Microsoft 365 Copilot fundamentals and licence tiers available at Arup.
Open Moodle courseSharePoint PDF — step-by-step guide for building your own Copilot Studio agents at Arup.
Open guideThe fastest, safest way to use Copilot for common advisory work.
| Need | Recommended tool | Why |
|---|---|---|
| Writing | Word Copilot | Best default starting point |
| Meetings | Teams + Copilot | Most mature meeting workflow |
| Slides | PowerPoint Copilot | Fastest delivery acceleration |
| Research | Researcher | Strong synthesis capability |
| Data | Excel Copilot | Best for structured analysis where enabled |
Concrete first tasks. Low risk, high value.
A quick path to the right Microsoft AI tool. Follow YES → when the question fits, or follow the black arrow down to the next question. All tools assume Arup's M365 Copilot (Premium) licence unless flagged.
Twelve AI surfaces grouped by use case. Click any tool to expand details.
Copilot drafts directly into Word documents using native Styles, grounded in up to 20 reference files via the /filename syntax. Vibe writing treats drafting as a dialogue — you describe intent and the document scaffolds. Work IQ gives Copilot access to your SharePoint, OneDrive, emails, and meetings.
/filename sources./filename references religiously — ground every draft in real sources./filename instead.Five distinct AI surfaces, each handling data differently. Edit with Copilot reshapes data and builds dashboards live. =COPILOT() persists AI inside cells. Analyst is a Python-validated agent for complex analysis. Vibe working means handing off the modelling problem and steering rather than writing every formula.
=COPILOT() functionNative formula for classification, extraction, normalisation at scale.Ctrl+T) with clean headers — no merged cells, no blank rows.Ctrl+T) before inviting Copilot — Tables are first-class.=COPILOT() for numerical calculations — use SUM, AVG, IF (determinism required).=COPILOT() for lookups — XLOOKUP is deterministic; =COPILOT() can fabricate values.=COPILOT() = 100 calls / 10 min; Analyst counts against the shared Researcher + Analyst monthly allowance (check current limit in the M365 admin centre).=COPILOT() cannot run in Confidential-labelled workbooks (Purview protection). Non-determinism is the risk: never client-face a =COPILOT() result without Paste Values freeze or Analyst validation.
Six distinct workflows, separated by source (multi-file reference, single doc, from-scratch, data-driven, rebrand, single-slide). PowerPoint Copilot handles layout; reasoning models handle content depth. The signature pattern separates prompt engineering, content generation, and layout into three steps.
Copilot in Outlook restructures how you touch email — not faster, but in the right order. It ranks your inbox, drafts emails from one-sentence intent, coaches your tone before send, schedules meetings from threads, and prepares you before meetings. Five distinct surfaces — most consultants know one and miss four.
Copilot in Teams turns live meetings into indexed, searchable assets — and turns shared channels into ongoing AI workspaces. Six distinct surfaces: Intelligent Recap structures recordings; Custom Templates apply formats; Facilitator co-authors visible notes; Channel Agent becomes the project's dedicated teammate.
OneNote is your notes container; Copilot in OneNote summarises and rewrites your own notes. Copilot Notebooks is structurally different — an AI workspace grounded only on curated references you add. Six surfaces total — most teams confuse the two and miss four of the six.
Copilot in Microsoft Edge restructures research from sequential reading to parallel triage. Every web page, PDF, and video becomes a queryable surface grounded in the content you're viewing. The sidebar stays live across your session — collapsing afternoon research into minutes.
Loop is a set of portable, synchronised content blocks (components) that live in the cloud but appear natively in Teams, Outlook, Word, and Loop pages. Copilot Pages are AI-generated canvases that start from a Copilot Chat response. Together they replace document versioning with persistent live state.
.loop files — Loop's compliance applies (sensitivity labels, DLP, eDiscovery, audit). Component visibility respects source page access — document who sees what before pasting.
The Copilot app (m365.cloud.microsoft or taskbar icon) sits above all in-app Copilot surfaces. It centralises conversation history, invokes agents, runs Search across work data, enables voice chat, and schedules recurring prompts. The orchestration layer around in-app Copilot — not a replacement for it.
m365.cloud.microsoft. Check your licence tier below your profile name.m365.cloud.microsoft — anything below Premium means features won't work.Two pre-pinned Microsoft agents for M365 Copilot Premium users. Researcher produces multi-page structured reports with citations. Analyst executes Python on data files for quantitative analysis. Both take 5–15 minutes per query and produce artefacts consultants would otherwise create manually. They share a monthly query allowance — check current limits in the M365 admin centre.
Six content-generation surfaces in Copilot's Create tab — image generation, infographics, posters/banners, video, forms, and prompt-to-presentation. All are brand-kit-aware when configured. Infographics are the highest-value advisory use case: a weekly client brief becomes a one-page visual summary in 2–5 minutes.
Four advanced tools for building and shipping custom agents beyond Agent Builder's boundaries: Copilot Studio (low-code, 1,500+ connectors), M365 Agents Toolkit (pro-code, declarative agents), Microsoft Agent Framework v1.0 (open-source SDK, multi-agent patterns), and Microsoft Foundry Agent Service (managed production runtime).
Confidence over coverage. Start where the team already does most of its work.
Best for: meeting recap, action capture, channel context.
Most mature meeting workflow. Replace note-takers with Recap and Custom Summary templates.
Open Teams workflow →Best for: drafting, rewriting, and structured reports.
Best default starting point for advisory writing. Ground drafts in 2–5 source files using /filename.
Best for: turning briefs and notes into client-ready decks.
Fastest delivery acceleration. Use the two-stage reasoning + layout pattern for grounded decks.
Open PowerPoint workflow →Best for: structured analysis, reshaping data, building dashboards.
Strongest for data work where enabled. Use Analyst for client-facing accuracy.
Open Excel workflow →Best for: commissioning multi-source research with citations.
Treat each query as a commissioning decision — the shared monthly query allowance means commission deliberately.
Open Researcher workflow →Four checkpoints that apply to every AI workflow. Click any card for the full position.
A named person remains answerable for any AI-influenced decision.
Generative AI is probabilistic and confidently wrong with non-trivial frequency. Every workflow that incorporates AI must include an explicit review point — a person who reads, challenges, and accepts or rejects the output. If the reviewer cannot explain how the output was produced, the review has not happened.
Accountability also covers the gap between review moments — particularly in agentic systems that operate continuously. The owner must have the authority to stop, change, or take responsibility for what the system does.
Sensitive material stays inside approved tools and approved data locations.
Choose where data lives, where models run, and which jurisdictions govern them before deployment, not after. EU AI Act compliance, data residency, and vendor concentration shape which providers are appropriate for which projects.
Use Enterprise Data Protection (the green shield in Edge / Copilot) for any Arup or client material. Personal accounts and consumer Copilot modes are not safe for confidential work.
Important AI outputs are verified against sources before they inform a decision.
Hallucination is a structural feature of generative models, not a bug to be patched. Treat plausible, fluent outputs as drafts that need cross-checking — particularly for facts, names, dates, regulatory references, and quantitative claims.
Use citations where the tool provides them (Researcher, Edge sidebar, Copilot Chat). Click through; don't quote what you haven't read.
Use approved tools at the access level enabled for your account.
M365 Copilot tiers determine which features behave correctly. Confirm your tier before relying on a Premium-only feature in client work — what works for a colleague may not be enabled for you. Check rollout, organisational policy, and per-app assignment.
Researcher and Analyst share a 25-query monthly quota where enabled. Treat each call as a commissioning decision, not a casual prompt.
Deeper context, pillars, and patterns. Expand only what you need.
LLMs are the reasoning engine — a general-purpose reasoner trained on text, code, and structured data. Copilot surfaces that engine inside Microsoft 365, grounded in your work data via the M365 Graph. Agents are configured Copilots with rules, instructions, and curated source sets, ranging from prompt-based helpers to multi-agent production systems.
LLM is the engine. Copilot is the steering wheel inside M365. Agents are the cruise control for repeatable tasks.
The Premise — the AI we choose. AI is a complementary factor in our work, not a substitute. Anything not actively chosen sits outside our governance frame.
Pillar 01 — Human-in-the-loop. Every consequential AI output is reviewed by a person who can challenge it. Review by someone who cannot meaningfully second-guess the output is theatre, not governance.
Pillar 02 — Sovereign. We choose where data lives, where models run, and which jurisdictions govern them — before deployment, not after.
Pillar 03 — Accountable. Every consequential decision that AI contributes to has a named human owner with authority to stop, change, or take responsibility.
Pillar 04 — Considered. We use AI deliberately, aware of cognitive, social, and environmental costs. The question "should I use AI for this?" is asked, not assumed.
The Foundation — Ready. Before any pillar applies: data ready, people ready, process ready, outcome ready. AI scales weak systems rather than fixing them.
Microsoft offers three tiers: Free Chat (web Copilot, no Graph grounding); M365 Copilot (in-app Copilot with Graph grounding, Researcher + Analyst with a monthly shared usage allowance, Enterprise Data Protection); Frontier & Premium (Edit with Copilot in Excel & Word, Computer Use, frontier model access, custom enterprise agents — request via team lead).
Confirm your tier before relying on a Premium-only feature in client work. Wording and visibility may vary depending on rollout, organisational policy, and per-app assignment.
Bookmark a tool: open foundations.html#tool-<name> to land directly on its detail panel.