.png?width=207&height=108&name=Mega%20menu%20%E2%80%93%20Featured%20Image%20Template%20(1).png)
UK
France
Germany
Italy
Spain
Poland
.svg)
Australia
Japan
USA
Canada
Over the past six months most of the enterprises we work with moved from a single Copilot deployment to running Claude alongside it. The question employees ask is no longer whether to use AI but which tool, for this task, with this data. Here is the framework we teach.
Short answer: start in Copilot when the work already sits inside an open Microsoft 365 file and you need it changed in place. Start in Claude when the work requires drafting, synthesis, or analysis across several permitted sources, or when you need a new output such as a deck storyline, a report, or an interactive artifact. Many tasks begin in Claude and finish in Copilot. Before either, confirm the data classification and the purpose are both approved.
Why are enterprises running Claude and Copilot at the same time?
Because the tools have converged enough to overlap but not enough to be interchangeable, and because governance was settled for Copilot first.
In our enterprise work this year, the pattern repeats across sectors: a global airline group, a payments company that had just completed an acquisition, an investment firm with five offices, and a multilateral development institution all described the same situation. Microsoft 365 Copilot was already licensed and governed. Claude arrived later, usually because leadership wanted stronger performance on complex, multi-document work, and in several cases because of Claude Code, Cowork, or connectors into systems like Jira and GitHub. One client described Claude as "the enterprise standard" while keeping Copilot in place for Microsoft-native tasks.
The most common construct we now see is a general-population tool (Copilot, or Gemini in Google shops) for everyone, with Claude licensed for subsets of the population whose work is heavier on synthesis and building: product managers, designers, engineers, business analysts. Six to nine months ago the construct was one tool for all.
The trade-offs are real. Two tools are harder to govern than one, license costs stack, and the rollout logic (who gets which tool, and why) is often unclear to the people expected to use them. The enablement job is to make that logic explicit.
What three questions should an employee ask before using either tool?
Permission, location, and reasoning. Together they decide whether to use AI at all, which tool to start in, and how much model capability the task needs.
- Permission. Can this information be used in an AI tool for this purpose? Both the data classification and the intended use must be approved. Permitted data can still be used in a prohibited way.
- Location. Where should the work happen? In the open file, or across several sources?
- Reasoning. How much capability does the task require? A spelling pass and a multi-report synthesis should not run on the same settings.
We teach these as a pre-send checklist rather than a policy document. The questions are the same for every task, which is what makes them stick.
When is Copilot the right starting point?
When the work already lives inside an enabled Microsoft application and the goal is to change that file in place.
Copilot's strength is proximity. If the task is "fix this formula in the workbook I have open" or "reformat these six slides," the file is the context, and leaving it to paste content elsewhere creates version risk. Representative Copilot-first tasks:
- Updating formulas or cleaning a range in an open Excel workbook
- Reformatting or tightening slides in an open PowerPoint deck
- Summarising a Teams thread or an Outlook chain inside the application
- Focused edits to a Word document that is already structured
Some organisations are Copilot-only by design and have built a real estate of Copilot agents. One consulting client we are training this autumn runs seven agents for proposal work alone (content generation, assembly, compliance review) and its question is orchestration in Copilot Studio, not whether to add a second tool. For them, the decision framework still applies; the "location" answer is simply always Microsoft.
When is Claude the right starting point?
When the task requires drafting, synthesis, or analysis across longer or multiple permitted sources, or when the output does not yet exist.
Claude's strength is turning context into something you can keep working on. Upload several approved reports and meeting notes, and Claude can analyse and combine them, draft a defined output, and refine it through conversation. The output arrives as a downloadable file (Word, Excel, PowerPoint, PDF) or as an artifact you can inspect and revise beside the chat. Representative Claude-first tasks:
- Analysing several long reports and developing a presentation storyline
- Producing a first-draft leadership update from supplier reports and meeting notes
- Building a reusable Project with a knowledge base for recurring work
- Turning a method one expert follows into a Skill the whole team can run
Can a single task use both tools?
Yes, and the best workflows often do. A task may start in Claude and finish in Microsoft 365.
The table below maps four common workplace tasks to the tool to start in, with the reason for each.
| Task | Start in | Why |
|---|---|---|
| Update formulas in the workbook currently open in Excel | Copilot | The work is inside the open file |
| Analyse several long reports and create a presentation storyline | Claude | Synthesis across sources |
| Reformat slides in the presentation currently open in PowerPoint | Copilot | In-app editing of a live file |
| Develop a storyline from several documents, then refine the slides | Both | Different stages, different tools |
Capabilities are converging and personal preference plays a role. Treat this as a starting heuristic, not a boundary.
How should companies set a data boundary that works across both tools?
Set one classification ceiling that applies regardless of tool, and pair it with a purpose test.
The airline group's approach is a useful model. Four classifications: Public, Internal, Confidential, and Strictly Confidential. The first three are approved for Claude; Strictly Confidential (personal data, safety and security data, board-restricted material) is prohibited in every format, including screenshots and pasted extracts. Confidential is the ceiling.
Then the purpose test. Summarising approved business documents, drafting market analysis, and creating first-draft reports are approved purposes. Personal projects, working around blocked systems, HR and people processes, and monitoring or comparing individual colleagues are out of scope, even when the underlying data is permitted. The rule we teach: Claude drafts; you review, verify, edit, and own the action.
How does memory change what employees should share?
It doesn't. Memory never expands what the organisation permits you to enter.
When memory is enabled, Claude may save useful context (role, preferred language, output format) and apply it in later conversations. That reduces repeated setup. It does not change the data boundary. We teach three habits: save stable, permitted context deliberately; review saved memories periodically rather than only when something looks wrong; and remove anything outdated.
What does this look like in an enterprise training programme?
A short, sequenced programme that opens with tool choice and governance before touching prompting technique.
In the airline group's eight-session Claude enablement programme, the first session is entirely about the governed environment: choosing between Claude and Copilot, handling data appropriately, and managing memory. Prompting comes second. Participants then complete a real workplace task twice, once in each tool or at two different model settings, and record which they preferred and why. That comparison, not a slide, is what builds judgement.
Key takeaways
- Most enterprises now run Claude alongside Copilot; the enablement job is to make the rollout logic explicit.
- Copilot: the work is inside an open Microsoft file and should change in place. Claude: synthesis across sources or a new output.
- Ask three questions before every task: permission, location, reasoning.
- One classification ceiling and one purpose test apply across both tools. Memory never changes what is permitted.
- Teach governance and tool choice first; prompting technique second.
Frequently asked questions
It depends on where the work lives. Use Copilot when the task is inside an open Microsoft 365 file and needs in-place edits. Use Claude when the task requires drafting, synthesis, or analysis across several permitted sources, or produces a new document, deck, or artifact. Many tasks start in Claude and finish in Copilot.
An enterprise setup where employees have access to more than one generative AI assistant, typically Microsoft 365 Copilot plus Claude or ChatGPT. It offers better fit per task but adds governance, cost, and rollout complexity, which is why clear tool-selection rules matter.
Classify data (for example Public, Internal, Confidential, Strictly Confidential), set the highest permitted tier as a ceiling that applies in every format, and add a purpose test so approved data cannot be used for prohibited purposes such as HR decisions or comparing colleagues.
No. Memory stores permitted context such as role and format preferences to reduce repeated setup. It does not expand the data classification an organisation allows employees to enter.
In the deployments we see, no. Claude is typically added for complex, multi-source work and for newer capabilities such as Skills, Claude Code, and connectors, while Copilot remains the tool for Microsoft-native, in-file tasks.
Build a Claude and Copilot programme for your teams
Correlation One designs enterprise enablement programmes that start with governance and tool choice, then build to Projects, Skills, and agentic workflows.