Microsoft Copilot or Claude? Use both. Just not for the same jobs.
I fired up a Microsoft 365 Copilot license the week before MAICON in Cleveland, then sat through Microsoft's own workshop and three days of CMOs showing what they built. Here is the honest split: where Copilot should lead, where Claude leads, and how to tell which is which for any job on your desk.
Your mail, meetings, files, and permissions are already there. Copilot's edge is where it lives, not which model it runs.
Your CRM, ad platforms, analytics, website, code, and browser. Anything that touches more than one system at once.
Under generic labels, at a flat rate, with Microsoft deciding the effort and context. Same model family, different experience.
Copilot's edge isn't the model. It's the address.
Copilot lives inside your Microsoft 365 tenant, with Work IQ grounding, permission-aware search, and admin governance built in. Claude's edge is reach across everything outside that tenant, and the depth of multi-step work it can do with all of it in one task.
The MAICON workshop made two things explicit that I had not seen written down anywhere. First, Claude is now available inside Copilot, but it shows up under generic labels ("Claude" or "Sonnet" in Chat, just "Auto" in the mini toolbars), through a Microsoft-managed experience at a flat per-user rate. Microsoft does not say what model, effort level, or context window you actually get on a given prompt. So Copilot can feel different from Claude directly, even on the same model family. The workshop's own advice was to test it side by side and to pick the model on purpose in the Word, Excel, and PowerPoint side pane.
Second, Copilot is two products with two bills. Everyday AI (Chat and the Office apps) is flat-rate under the license. Advanced AI (Cowork, Code, Autopilot, agentic SharePoint, and the newest frontier models) is metered in Copilot Credits on top, and in enterprise tenants it stays switched off until an admin creates a spending policy. If a feature you read about is missing, that is usually why.
Know your license. Then know what's metered.
What Copilot can do for you depends on the label under your name at microsoft365.com (scroll to the bottom of the home page), and on whether usage-based billing is turned on.
| Label | What you get | What's missing |
|---|---|---|
| Copilot Chat (Basic) | Chat only | No Copilot in Word, Excel, PowerPoint, or OneNote |
| M365 Copilot (Basic) | Standard Copilot in the apps | No add-on license, no priority access |
| M365 Copilot (Premium) | Add-on license, priority access, the full experience in the apps | Cowork, Code, and Autopilot still need credits turned on |
| Tier | How it's billed | What it covers |
|---|---|---|
| Everyday AI | Flat per-user subscription license | Copilot Chat; Copilot in Word, Excel, PowerPoint, Outlook, and Teams |
| Advanced AI | Usage-based Copilot Credits on top of the license | Cowork, Code, Autopilot, agentic SharePoint, the newest frontier models |
Model choice. Copilot Chat defaults to Auto. Effort runs Quick Answer, then Think Deeper, then Advanced Reasoning. More thinking means more tokens, which means more money. The Researcher agent can run Auto/Critique (GPT drafts, Claude does a second reasoning pass), Model Council (GPT and Claude research the same question in parallel, with a summary of where they agree and differ), or a single model directly. Claude and Model Council both need Frontier access and admin-enabled Anthropic models.
Claude's side of the same comparison: one subscription, the model is named, effort is set by the model you pick, and connectors are bring-your-own. Sandy Carter's reason at MAICON for choosing Claude over the agent frameworks she tried, token cost and the connectors actually available, is the same lens to apply here.
A lot of "Copilot things" are just things.
Both can summarize email and threads, draft documents and decks, reason over spreadsheets, search SharePoint and OneDrive, read Teams chats, run multi-source research, keep a persistent project context (Copilot Notebooks vs. Claude Projects), save memory and custom instructions, schedule recurring prompts, and build simple agents from instructions plus a few knowledge files.
Claude's Microsoft 365 connector already covers Outlook, Calendar, Teams, SharePoint, and people search. So many of the jobs people buy Copilot for are already askable in Claude, and Claude can chain them with HubSpot, Supermetrics, Ahrefs, or a client's UTM Grabber data in the same task.
The workshop's prompt discipline transfers unchanged, too. Goal, Context, Sources, Expectations is the four-line template. The six-part version adds Audience, Requirements, and Guardrails. Both work identically in either tool, and both are the real reason most people's early results disappoint: starved context produces starved work.
If the work stays inside the tenant, stay in Copilot.
The workshop's capability map, organized by where it actually pays off. The right column is the honest read on where Claude sits for the same job.
| Area | Why Copilot wins | Where Claude sits |
|---|---|---|
| Work IQ grounding | One toggle decides whether an answer draws on your organization's mail, meetings, chats, and files, with citations you can open. The "Find / Catch up / Prepare" prompts work out of the box. | Searches mail and SharePoint through the connector, scope by scope. No org-wide graph. |
| Editing inside Office | Real-time editing alongside Copilot in Word. Plan mode, Python, and Power BI grounding in Excel. Model picked per app in the side pane. | Exports files. Can't edit a .docx or .xlsx in place in your tenant. |
| Scheduled prompts | Run a proven prompt weekly without leaving Outlook. The workshop's Friday inbox review returns responses owed, decisions, deadlines, and cited emails. | Scheduled tasks exist, but run outside the tenant and reach mail through the connector. |
| Cowork with approval gates | Plans the work, shows you the plan, pauses before anything sends. The calendar-conflict demo and a 69-headshot directory cleanup were done end to end. | Same pattern is available in Claude, but Cowork's actions land natively in Outlook and OneDrive. |
| Recorded Skills | Record a PowerPoint formatting task once, reuse it by @mention. Slide styling, table cleanup, logo placement, chart styling. | Skills are written, not recorded, and don't act inside PowerPoint. |
| Brand Kits | Upload your brand guidelines PDF. Copilot extracts colors, voice, and usage rules and applies them to every Create output. | Brand lives in a Project's knowledge files. No app-wide kit. |
| Teams meetings | Recaps, action items, call delegation, and video recap require being in the call. | Reads Teams messages. Not a meeting participant. |
| Tenant-wide search | The graph index respects every SharePoint and Exchange permission automatically. | Connector search, permission-aware but query by query. |
| Governance | Purview integration, admin spending policies, audit trails, agent registration. This matters for our healthcare clients. | Enterprise controls exist but aren't tied to Microsoft's compliance stack. |
| Agent Builder → Copilot Studio | A declarative agent in minutes: instructions, three knowledge docs, starter prompts. Copilot Studio adds multi-channel, connectors, auth, and analytics. Power Automate for event-driven flows. | Projects and skills cover the declarative tier. The business-workflow tier is Make.com and HubSpot, not Microsoft. |
If the work touches more than one system, or a client will see it.
The MAICON main stage was, in practice, a catalogue of builds done in Claude. Veeam's creative studio, L2 Digital's 34-skill business operating system, Samsara's 100-plus production agents, Sandy Carter's podcast agent. This is where that work lives.
| Area | Why Claude wins | Where Copilot sits |
|---|---|---|
| Split reasoning through MCP | Conductor's pattern: your data (facts) → your LLM (reasoning) → your team (judgment). It runs today in Claude against HubSpot, Supermetrics, Ahrefs, UTM Grabber, and Google Drive in one task. | Federated MCP connectors (HubSpot, Canva, Notion, Linear) are read-mostly. No sandbox to run the numbers. |
| Code and a real shell | Scripts, data cleaning, WordPress and Divi, Make.com payloads, Python on raw CSVs. Keith Moehring's START-HERE.md business OS and Veeam's banner and collateral generators (zero tokens to run) were built in Claude Code. | Python is confined to Excel. Copilot Code is a metered Advanced AI feature. |
| A skills library with a manual | Moehring's prompt: "build an HTML operating manual showing a menu of all skills, each with purpose, prompt, and process steps, labeling which are human and which are AI." That is a Claude Code pattern. | Recorded Skills are app-bound. Agent Builder's skill-building skill is brand new. |
| Browser automation | Logs into client ad accounts, fills forms, posts approved page content and verifies the look (Mike Kaput's distribution step). | No browser agent in M365 Copilot. |
| Long-form reasoning and writing | Strategy documents, proposals, a book, with full project context, memory, and past conversations. And the voice that keeps your work from reading like everyone else's. | In-app sidebar with document context. Generic model labels. |
| Building for clients | Published dashboards, calculators, client-facing reports, and full agents: a Project plus a skill plus connectors plus a scheduled run plus approval gates. | Output lives as Office files, Pages, or Notebooks inside the tenant. |
| Attribution and AI-search measurement | CallRail's two-bucket split (paid LLM as its own channel; organic LLM pulled out of referral) and Conductor's chosen-vs-mentioned citation share are UTM Grabber and Ahrefs jobs. | Power BI grounding reads reports. It doesn't build the tracking. |
Five levels. One question per job.
Microsoft's five-level framework maps cleanly onto what we already run in Claude. Use it to tell yourself, or a client, which tool a given job belongs in.
| Level | Copilot tools | Claude equivalent | Who should own it |
|---|---|---|---|
| 1 · Individual productivity | Copilot Chat; Copilot in Word, Excel, Teams, Outlook | Claude chat with the Microsoft 365 connector | Either whichever is open |
| 2 · Repeatable workflow | Notebooks, Scheduled Prompts, Skills, Cowork Tasks | Projects, scheduled tasks, skills, memory | Copilot for tenant-bound jobs; Claude for anything touching outside data |
| 3 · Shared capability | Agents, Agent Builder, Brand Kit, Connectors | Shared Projects, MCP connectors | Claude clients and partners aren't in your tenant |
| 4 · Business workflow | Copilot Studio, Agent Flows, Power Platform | Make.com, HubSpot workflows, UTM Grabber reporting | Claude + Make.com Copilot Studio only for Microsoft-heavy shops |
| 5 · Emerging and frontier | Autopilot, Code, Home in Copilot, Frontier Program | Claude Code, computer use, browser automation | Claude already in production at Veeam, L2, Samsara |
Does this even need an agent?
Knowing when not to build one is part of agent literacy.
- A reusable prompt or Notebook is enough when one person uses it occasionally, the context changes every time, or there's no stable role or knowledge
- Build an agent for a defined group, a stable method, curated knowledge, and tools it needs to call
- Or, as Samsara put it: not 40 agents, six jobs
Prototype is not the finish line.
Identified → Prototype → Pilot → Production, with a gate at each step.
- Prototype means at least two rounds of subject-matter review before anyone else touches it
- Pilot means an opted-in group, a captured baseline, success criteria, an end date, and the decision meeting already booked
- For almost everyone, the honest destination this quarter is prototype or pilot prep, not pilot
Five safeguards. Either tool.
Sandy Carter built one marketing agent three times and named the five ways it went wrong. Her five safeguards are the governance pack for any agent, in either tool. Copilot ships some of them by default. In Claude they are design decisions you make on purpose.
| Safeguard | In Copilot | In Claude |
|---|---|---|
| Approved tools (no double agent) | Admin-governed connectors and agent registration | The connector list is yours to curate, per Project |
| Least access ("give it the list, not the login") | Graph permissions inherited per user | Scope each MCP connector; use read-only keys where offered |
| Access that ends | Licenses and connectors revoked centrally | Rotate and remove connector keys when an engagement ends |
| Human approval on sends, bookings, publishing | Cowork pauses before sensitive actions by design | Build the approval gate into the skill or the Make.com scenario |
| A record of every run | Audit trail, Copilot Analytics, Purview | Scheduled-task logs reviewed in a nightly digest |
That one cuts the other way from governance: work in public. Zapier's rule is that a DM is not in the brief, not seen by the team, and not findable by agents. Copilot's shared Notebooks and Claude's shared Projects are the same answer to the same problem.
Copilot for the inbox. Claude for the business.
We run on Microsoft 365 ourselves, so Copilot is the daily driver for email, calendar, meeting prep, and in-place editing of Office files. Claude is where every build that touches marketing data, code, the web, or a client lives: visitor intelligence, attribution reporting, the SDR agent, client dashboards, and the writing. Email, calendar, and SharePoint lookups go to whichever is already open.
If you just got a Copilot license, here is the first 30 days in the workshop's own crawl-walk-run order.
- Week 1 · Check the labelConfirm your license tier at microsoft365.com and whether usage-based billing is on. Without it, Cowork, Code, and Autopilot won't appear, and you'll think they don't exist.
- Week 1 · Pick the model on purposeOpen the Copilot side pane in Word, Excel, and PowerPoint and choose the model. Run one real task there and in Claude side by side. Keep the result as your baseline.
- Week 2 · Schedule the Friday reviewTurn on the inbox review as a scheduled prompt: responses owed, decisions, deadlines, cited emails. Check it twice, then leave it running.
- Week 2 · Record one skill, load one kitRecord a PowerPoint skill for your deck style. Upload your brand guidelines PDF as a Brand Kit so every Create output starts on-brand.
- Week 3 · Try Model CouncilRun Researcher with Model Council on one real business question. Compare it to the same question in Claude with your CRM and analytics connected.
- Week 4 · Write the governance packThe five safeguards, one page. Apply it to anything that sends, books, or publishes before you let it run on its own.
- MAICON 2026 Copilot for Business workshop deck (Microsoft; pricing and billing slide verified by the presenter October 11, 2026)
- MAICON 2026 sessions: Conductor, CallRail, Sandy Carter (Equs), Keith Moehring (L2 Digital), Jessica Hreha (Veeam), Holly Chen (Samsara), Dan Slagen (Zapier), Mike Kaput and Paul Roetzer (SmarterX)
- Microsoft Tech Community, What's New in Microsoft 365 Copilot, June 2026
- A Guide to Cloud, August 2026 Copilot updates
- A Guide to Cloud, May 2026 Copilot updates (federated MCP connectors)
- Microsoft Copilot blog. The workshop's advice stands: check the release notes before relying on a detail. Copilot changes fast.
Bring us your desk.
Fifteen focused minutes on your business. We'll sort your recurring work onto the five-level ladder and you'll leave knowing which tool gets which job, and what to build first.
Book the Fast 15