The two named reasoning agents every M365 Copilot Premium user has and most never use. One does deep multi-step research across work + web. One runs Python to analyse data files. Both take longer than standard Copilot — and deliver output standard Copilot can't. They share a monthly query allowance; the discipline of spending it well is the whole skill.
Before the features, the reframe. Everything else in this playbook has been about prompting — you ask, Copilot answers. Researcher and Analyst are different. You commission; the agent plans, retrieves, reasons, reviews, and delivers. The difference isn't speed; it's shape of output.
Think about how you'd ask a senior colleague for help with two different tasks. Task 1: "Can you tell me what the last ORR guidance said about rail accessibility?" They tell you in a sentence. Task 2: "Can you prepare a briefing on the UK transport sector's current approach to digital asset management, covering main operators, regulatory context, industry research, and Arup-relevant angles?" Different request. They'd go away for an afternoon, come back with something substantive. Researcher and Analyst are the AI equivalents of the second kind of request.
"Researcher is designed for deeper reasoning of your tasks. By taking longer processing time, Researcher delivers a more comprehensive output, making it ideal for detailed research projects."
Three specific shifts this produces:
Standard Copilot Chat returns answers in 2-30 seconds. Researcher and Analyst take 5-15 minutes to complete a query — sometimes longer for complex requests. That isn't a bug; it's the mechanism. During those minutes, the agent is planning the approach, retrieving sources iteratively, reasoning over what it finds, cross-checking, drafting, and revising. The output shows the work that standard Copilot can't produce because it's optimised for speed.
Standard Copilot produces a paragraph or short bulleted answer. Researcher produces a multi-page structured report with sections, citations, visual elements, sometimes charts. Analyst produces a reasoned analysis with the Python code visible, intermediate results, and business-language interpretation. These are decision-grade artefacts — the kind you'd attach to a client briefing, paste into a Word document, or present to a team.
The shared monthly quota changes the relationship. You don't fire off a Researcher query for something you could get from standard Copilot in 10 seconds — that's a wasted query. You save them for tasks where a reasoning agent's depth and rigour genuinely justify the cost. Each query becomes a small commissioning decision. The discipline forces better use.
For Arup advisory specifically, the two agents map directly to the two most common billable activities: researching a topic to produce a deliverable (Researcher) and analysing data to produce an insight (Analyst). These aren't marginal productivity boosts — they're agents that do work consultants would otherwise spend half-days on.
Two agents, six distinct capabilities worth knowing. Core Researcher and Analyst are the starting points; the additional capabilities (Critique, Computer Use, output conversion, the shared quota) shape how you actually use them.
Two named agents. Similar framing (deep reasoning, minutes-to-answer, pre-pinned in the app). Very different capabilities. Knowing which to commission for which task is the primary skill.
Model: OpenAI's deep research model, with Critique multi-model review.
Works on: web + work content (emails, files, meetings, chats) + third-party connectors.
Output: multi-section structured report with citations. Convertible to PowerPoint / PDF / infographic / audio.
Takes: 5-15 minutes typically. Longer for complex topics.
Languages: multiple supported — check current coverage in M365 Copilot.
Model: an OpenAI reasoning model, optimised for analysis.
Works on: data files you attach (Excel, CSV, PDF, XML, PowerPoint). Also M365 Graph data.
Output: reasoned analysis with Python code visible, insights in prose, charts, tables.
Takes: 3-10 minutes typically.
Languages: 8 supported (expanding).
A simple rule that captures 95% of cases:
The edge cases where both might apply:
The strongest advisory workflows often use both agents sequentially. Typical pattern: Analyst crunches the client's data to identify trends; Researcher investigates what's driving those trends in the broader market; you synthesise the two into a client briefing. Two quota queries for a deliverable that used to take two days; one day's turnaround is still faster than most competitors.
Fewer technical prerequisites than most chapters — the named agents work out of the box for any M365 Copilot (Premium) user. But the quota constraint and the "this is not standard Copilot" expectation need explicit setting.
1. Confirm M365 Copilot (Premium) licence. Researcher and Analyst are only available with a full M365 Copilot licence. Copilot Chat (Basic) users cannot access them. Check at copilot.cloud.microsoft — label under your name should read M365 Copilot (Premium). See Part XII for full tier discussion.
2. Verify admin enablement. Copilot administrators can enable or disable Researcher and Analyst at tenant level. If you don't see them in the Copilot app's Agents section, check with IT — it may be a policy setting rather than a licence gap.
3. For Critique and Computer Use, confirm Frontier program status. Multi-model Critique and Computer Use require tenant enrolment in the Frontier program. If enrolled, these capabilities are available; if not, you get standard Researcher. Frontier enrolment is an admin decision.
4. Understand the quota. Researcher and Analyst share a monthly query allowance — confirm the current cap in the M365 admin centre, as Microsoft adjusts it. Resets at the start of each calendar month. Count is not currently surfaced in-product — you're informed when close to or at the limit. Plan accordingly.
5. Set expectations on timing. Researcher and Analyst take 5-15 minutes per query. That is not slow — it's the mechanism. Submit the query, do other work, come back. Don't abort mid-query; don't re-submit while waiting. Consultants new to these agents often don't realise they're supposed to leave the query and return. Setting expectations up front saves frustration.
The highest-value default use of Researcher for advisory consultants. Day 1 of a new engagement. Brief the agent with client, scope, and specific questions. Come back to a multi-page report. An afternoon's research becomes 15 minutes.
The single biggest difference between a good Researcher output and a poor one is the specificity of the brief. Vague prompt, vague report. Structured brief, structured report.
Even with Critique multi-model review, verify critical claims before relying on them commercially. Three checks:
Budget 15-20 minutes for review. Total time: 30-45 minutes for a report that used to take half a day.
Once reviewed, use the March 2026 format conversion: PowerPoint for internal pre-kickoff briefing; PDF for the engagement's Loop workspace archive; infographic for a one-page client-facing summary; audio overview if you'll listen during your commute to the kickoff.
The second natural Researcher use case. Who are the other firms in this space? What are they offering? Where are Arup's distinctive angles? What's moving in the market? Historically a 1-2 day desk research effort; now a 20-minute commission.
Competitive analysis has a particular AI failure mode: sometimes conflating capabilities between adjacent firms. "Mott MacDonald has a strong X practice" when it's actually Jacobs — because the two firms have similar phrasing on their websites. Always verify any competitor-specific claim before using it in a pitch or proposal. Cost of miss-attribution in a proposal is real.
For substantial competitive research needing gated industry databases (Gartner, Forrester, industry analyst reports), Researcher with Computer Use (Frontier only) extends reach. The agent can log into subscription databases with your approval, pull the research, integrate with public-web analysis. Same one query, richer output.
The March 2026 capability that closes the loop. A Researcher report becomes a slide deck, PDF, one-page infographic, or audio briefing — without re-writing. The workflow that makes Researcher's output immediately useful rather than "a source document for my actual deliverable."
One query, three deliverables. Run Researcher query. Once report is ready, convert to PowerPoint for Tuesday's internal briefing. Convert to PDF for the engagement's Loop workspace archive. Convert to infographic for the one-page summary to share with the client partner. Three deliverables, one query, maybe 10-30 minutes of conversion + editing time across all three. This is the workflow pattern that turns Researcher from "a research tool" into "an output-production engine."
Converted outputs are good starting points, not final deliverables:
The primary Analyst use case for advisory work. Client sends spreadsheets. You need to understand what they say, what patterns are in them, what insights emerge. Historical pattern: export to Excel, build pivot tables, eyeball for patterns, hope you didn't miss anything. New pattern: attach the files, ask the question, read the analysis.
Expect 2-4 pages of structured output:
The Python code being visible is the feature. Read it:
Don't trust headline numbers; trust the code that produced them.
The second Analyst use case. Not client data — Arup's own data. Team portfolio performance, engagement metrics, practice-area trends, internal benchmarks. The analyst-on-call you never had, available for the 15-minute "I'm curious about this" questions that normally go unanswered.
The single pattern worth establishing: when you notice yourself wondering about internal data, run an Analyst query rather than noting it down "for when I have time." The friction cost is 10 minutes; the decision value is often meaningful. Some specific Arup patterns:
1. Quarterly engagement mix review. Export the team's engagement register. Ask Analyst: "analyse this register by client sector, engagement size, and outcome — what's the overall mix, what's changed year-on-year, any concentration risks?"
2. Pipeline velocity analysis. Export pipeline data. Ask: "for engagements won vs lost over the last 12 months, what's the average time from first contact to contract? Are there patterns in what converts faster?"
3. Practice-area revenue patterns. Revenue data by practice area and quarter. Ask: "identify any seasonal patterns, growth or decline trends, and practice areas where revenue variability is unusually high."
4. Internal capacity analysis. Timesheet data across the team. Ask: "who's been most utilised over the last three months? Any signals of over-utilisation? How does utilisation map to seniority?"
5. Client portfolio concentration. Client data. Ask: "what proportion of our total revenue comes from our top 5 clients? Has concentration risk increased or decreased over 18 months?"
Two rules for internal-data Analyst use:
The discipline most consultants struggle with. A monthly allowance feels like a lot at first. By day 20, most consultants find they've used most of it on marginal tasks and don't have capacity for the ones that actually mattered. The skill is knowing which queries deserve the slot.
For a typical advisory consultant, a defensible allocation across the monthly allowance might look like:
Not a prescription; a starting template. Adjust based on actual work. Senior consultants in research-heavy roles may lean more toward Researcher; consultants in data-heavy workstreams may lean toward Analyst.
The discipline works better with a weekly cadence:
Things consultants sometimes run named-agent queries for that don't deserve the slot:
Six use cases, two agents, one table.
| Your scenario | Use this | Agent | Time |
|---|---|---|---|
| New engagement kickoff research | Engagement brief template | Researcher | ~30 min |
| Competitive landscape for a proposal | Competitive landscape prompt | Researcher | ~25 min |
| Monthly market intelligence check | Sector change prompt | Researcher | ~20 min |
| Client sends multi-file data for analysis | Multi-file Analyst prompt | Analyst | ~15 min |
| Quick internal data curiosity question | Attach + ask pattern | Analyst | ~10 min |
| Research report needs client-ready slide deck | Convert to PowerPoint | Researcher | ~15 min |
| Listen to brief during commute | Convert to audio overview | Researcher | ~3 min |
| Gated database research (Gartner, etc) | Researcher with Computer Use | Researcher (Frontier) | ~20 min |
| Trend detection across time-series data | Time-series Analyst prompt | Analyst | ~15 min |
| Simple factual question with current web info | Standard Copilot, NOT Researcher | Copilot Chat | ~30 sec |
| Quick email draft or summary | Standard Copilot, NOT named agents | Copilot Chat | ~30 sec |
| Small data query (single spreadsheet, basic aggregation) | Excel Copilot, NOT Analyst | Excel | ~2 min |
Patterns from training advisory teams on the named agents. Almost every bad outcome maps to one of these.