Microsoft Copilot vs ChatGPT Enterprise vs Claude (2026)

Copilot vs ChatGPT Enterprise vs Claude is a decision about the shape of your work more than about which model is smartest. In our judgment, Microsoft 365 Copilot is the sensible default when work happens inside Word, Excel, PowerPoint, Outlook and Teams, while ChatGPT (Business or Enterprise) or Claude (Team or Enterprise) usually fits research, drafting and analysis across many sources. Often the honest answer is two tools, so the real decision is which work goes where, settled by a 30-day bake-off on your own tasks.

Key figures (updated 23 October 2026)

• Microsoft 365 Copilot: $30.00 per user per month paid yearly, $31.50 paid monthly, on an annual commitment and a qualifying Microsoft 365 plan.
• Microsoft 365 Copilot Business (up to 300 users): $21.00 list, $18.00 in year one for purchases from July 1 to December 31, 2026.
• ChatGPT Business: $20 per user per month billed annually, $25 monthly; ChatGPT Enterprise has no published price.
• Claude Team: $20 per seat per month billed annually, $25 monthly; Claude Enterprise is $20 a seat plus all usage at API rates.
• Training: off by default on all three business offerings, though Anthropic says explicitly submitted feedback may be used.
• Evidence: a UK government pilot found no evidence that Copilot's time savings improved productivity, and we found no independent test of all three.

Cite as: AIDOLS Research Team, Microsoft Copilot vs ChatGPT Enterprise vs Claude (2026), aidolsgroup.com, 23 October 2026.

This guide covers prices, a 100-user cost example, data terms, controls, integrations, models, agents, a 30-day bake-off and 15 vendor questions. AIDOLS does not resell platform vendor licenses. "Copilot" means Microsoft 365 Copilot, which Microsoft's documentation now calls Microsoft Copilot, not GitHub Copilot.

Choose by where the work lives: how many sources the answer needs, and where the finished work lands, which the AI Fluency course calls the one-sentence rule: one document, or across the work. These starting points are AIDOLS judgments.

By the shape of the work

By organization type

As of October 2026, Microsoft 365 Copilot lists at $30 per user per month and ChatGPT Business and Claude Team at $20 a seat billed annually; ChatGPT Enterprise is quote-only. These are US list prices; local prices differ, and Microsoft says its web prices are for marketing purposes only.

What the per-seat price leaves out

Seat prices exclude metered work. Microsoft bills Copilot Cowork, Code, Autopilot and the Work IQ APIs in Copilot Credits on top of the license, and Copilot Studio capacity costs $200 a month per 25,000 credits or runs on a pay-as-you-go meter. OpenAI sells extra Business usage as workspace credits and meters Enterprise usage at contract rates. The Claude Enterprise seat covers access only, with every token billed at API rates, so it does not compare with a $20 Team or ChatGPT Business seat that includes usage.

Prerequisites and purchasing limits

• The add-on's qualifying plans go well beyond E3 and E5: Business Basic, Standard and Premium, F1 and F3, Microsoft 365 Apps and Office 365 E1, E3, E5 and F3 qualify, among others.
• Self-serve ChatGPT Business, renamed from ChatGPT Team on August 29, 2025, has no invoicing, purchase orders, Zero Data Retention or BAAs; those need a contracted plan such as Enterprise.

At October 2026 US list prices, 100 users cost $24,000 to $36,000 a year on one tool (Standard seats), and $31,200 to $42,000 on Copilot plus 25 cross-work seats. The arithmetic assumes annual billing unless noted and excludes tax, the Microsoft 365 base licenses you already pay for, and usage-based charges.

The 25-seat cohort and the 80/20 split are AIDOLS planning estimates; ChatGPT Enterprise is quote-only. ChatGPT Business and Claude Team cost the same at list, so quality, controls and data terms will separate them.

On business plans, none of the three trains on your content by default. The differences are feedback buttons, retention defaults and the contract that governs.

Feedback buttons are not handled the same way

Unlike OpenAI and Microsoft, Anthropic says material users explicitly report, such as thumbs-up or thumbs-down feedback, may be used to train its models. Claude Team and Enterprise owners can switch feedback off with the "Rate chats" setting (Anthropic), so decide whether to, and tell staff what the buttons do.

Personal accounts are the real leak

OpenAI's Free, Plus and Pro plans may use conversations for training unless the user opts out, and Anthropic's pricing page lists training on Free, Pro and Max as opt-out. Staff on personal accounts sit outside every term above, so decide what staff may paste into any AI tool and which accounts count as work accounts.

This is general information, not legal advice: have counsel review vendor terms against your obligations, and check your regulator's current guidance.

Single sign-on comes with every business tier; SCIM, audit logs and compliance APIs are Enterprise features on ChatGPT and Claude, while Copilot inherits your Microsoft Entra identity and logs interactions in Microsoft Purview.

ChatGPT Enterprise and Claude Enterprise also add IP allowlisting and customer-managed keys (ChatGPT where available). Sources: Microsoft on enterprise data protection and Copilot audit logs; OpenAI on the Business plan, SCIM, Enterprise admin controls and residency; Anthropic's plan comparison and Enterprise plan.

Data residency for EU, UK and Canadian organizations

Data residency has two halves, where data is stored and where the model runs, and vendors often promise only the first.

• EU and UK. Copilot is an EU Data Boundary service for EU customers, and the boundary covers the EU and EFTA, not the UK. Anthropic models in Copilot sit outside it and are off by default there and in the UK. OpenAI-operated models in Copilot, on by default for eligible commercial customers since July 24, 2026, are inside it but outside in-country commitments. Microsoft expects in-country Copilot processing in the UK and the US, among others, by the end of 2026, a roadmap rather than a contract term.
• Canada. Microsoft now expects in-country Copilot processing in Canada in 2027; until then, it says customers outside the EU may have queries processed in the US, the EU or other regions. OpenAI can store new Enterprise customers' data in Canada but offers inference residency only in Europe, the US and the UAE, so on our reading Canadian model processing happens elsewhere. Anthropic documents a US-only inference setting for usage-based Claude Enterprise at 1.1 times API rates, and Canadian regions on AWS Bedrock and Google Cloud Vertex AI for API access; ask Anthropic about app storage. See also Canadian AI compliance under PIPEDA and Quebec's Law 25.
• Connectors. OpenAI says data sent to connected apps, MCP servers or web search may leave the residency region, and Anthropic's US-only setting does not cover connectors.

What the vendors say about audits and standards

Microsoft states that Copilot's compliance offerings include GDPR, ISO 27001, HIPAA and ISO/IEC 42001. OpenAI reports completed SOC 2 Type 2 audits for ChatGPT Business and Enterprise. Anthropic lists SOC 2 Type 2, ISO/IEC 27001, 27017 and 27018 and CSA STAR. These are vendor statements: ask for the reports under NDA and check their scope.

All three now work inside Office files and reach Microsoft 365 content, each through the user's own permissions.

Inside the document: Copilot natively, ChatGPT and Claude through add-ins

A Copilot license covers Copilot in Word, Excel, PowerPoint, Outlook and Teams (Teams needs its own license) and the Researcher, Analyst and Facilitator agents; without it, Copilot chat inside Word, Excel, PowerPoint and OneNote is limited to standard access or unavailable, depending on licenses and tenant settings. ChatGPT for Excel, Word and PowerPoint is generally available to Business and Enterprise users, though add-in chats do not sync with ChatGPT history. Claude for Excel, PowerPoint and Word is generally available on paid plans, with Word edits as tracked changes and Claude for Outlook in beta.

Across the work: Graph, connectors and MCP

Copilot grounds answers through Microsoft Graph and Work IQ: Microsoft 365 content, data from Copilot connectors, the semantic index and Copilot Memory. ChatGPT uses connected apps such as SharePoint, Teams, Outlook, Google Drive, Slack and GitHub, plus Projects, GPTs, Company Knowledge and Deep Research. Claude's Microsoft 365 connector searches SharePoint, OneDrive, Outlook and Teams after a one-time Entra Global Administrator consent, with write tools only if an admin enables them. Connectors increasingly use the Model Context Protocol (MCP), which Anthropic gave to the Linux Foundation's Agentic AI Foundation in December 2025 and says ChatGPT, Gemini and Microsoft Copilot have adopted.

The permission problem all three share

Copilot surfaces only data a user can at least view. ChatGPT's SharePoint app retrieves only what the connected account can open and needs Sites.Read.All and Files.Read.All with admin consent; where OpenAI's site and label controls are enabled, admins can allow or block site collections and exclude sensitivity labels. Claude's connector uses delegated permissions and tenant-wide Sites.Read.All, because site-level scoping is not supported. The UK government's Copilot experiment warned that searching with a user's access "may highlight issues when users have access to files they shouldn't." In our judgment, that makes oversharing an answer problem in all three tools: fix SharePoint oversharing before any assistant reads it.

As of October 2026, Copilot runs on OpenAI models (Microsoft's plan tables list "GPT-5.X") with model choice for licensed users. Anthropic is a Microsoft subprocessor, its models on by default for most commercial tenants outside the EU, EFTA and UK, and OpenAI-operated models have been available as a subprocessor since July 9, 2026. ChatGPT's newest are the GPT-6 family, including GPT-6 Astra, which was off by default for Enterprise workspaces at launch; Claude's are Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, each with a 1M-token context window, in chats on paid plans too.

Do not buy a model name. OpenAI retired GPT-5.5 from ChatGPT on October 14, 2026, and Copilot now offers models from two providers. Buy for the work shape, the controls and the contract, re-run your bake-off tasks when a model changes, and ignore vendor-run benchmark tables. The same applies to people: train them on the work, not on one vendor's buttons.

An AI agent carries out multi-step work instead of answering one prompt; each vendor bills agents at least partly outside the seat.

• Microsoft. Licensed users can create and use internal agents in Copilot Studio within the license; external channels or unlicensed users need a standalone Copilot Studio license. Copilot Cowork runs multi-step tasks across Microsoft 365 apps on Anthropic models, pausing for approval at key checkpoints and billed on usage, on top of a required Copilot license. Agent 365, which Microsoft calls a control plane for agents, lists at $15.00 per user per month.
• OpenAI. ChatGPT Work is the agent for research, analysis and finished deliverables. Business seats include it with ChatGPT Agent, Deep Research and Codex, and Enterprise admins choose which roles get Work, workspace agents and Codex.
• Anthropic. Claude Cowork takes a goal and works across your files and tools on paid plans, and Claude Team also includes Claude Code. On Claude Enterprise both bill at API rates. It is not Microsoft's Copilot Cowork.

Microsoft's pay-as-you-go agent billing is disabled by default; keep it that way until a budget and an owner exist. Put write actions such as sending mail behind approval and logging, because agents that read email, web pages or shared files can be steered by planted instructions (prompt injection); Copilot's audit records flag detected cross-prompt injection. And agents pay off only when the process is redesigned around them, the case made in agentic process automation without a multi-year program.

We found no independent test of all three on the same tasks, so judge task by task; the verdicts below are AIDOLS judgment.

What the public trials actually found

The public evidence is Copilot-only and largely self-reported. In the UK's cross-government experiment, 20,000 staff self-reported saving 26 minutes a day on average, with limits on complex, nuanced or data-heavy work. The UK Department for Business and Trade's 1,000-license pilot found small time savings but no evidence that they improved productivity; scheduling and image generation took longer, and respondents reported hallucinations. In Australia's six-month government trial, only a third of participants used Copilot daily, mostly to summarize and rewrite.

A preregistered Harvard Business School and BCG experiment with 758 consultants shows why your own tasks matter more than any benchmark: AI users completed 12.2% more tasks, 25.1% faster inside the AI's capability frontier, but were 19% less likely to be correct on a task outside it. It used GPT-4 in 2023, so take the method, not the numbers.

Run the same 10 real tasks through each candidate for 30 days, score blind, verify every number and review security in parallel.

1. Pick 10 real recurring tasks and record a baseline. Take them from the people who own the work and record time, rework and errors first, so you can tell a result from a demo.
2. Write one brief per task and freeze the inputs. State the job, inputs, output and standard, and give every tool the same source files.
3. Set each tool up fairly. Business-grade plans, the same connectors, permissions and SSO, training and feedback settings checked, monthly billing or trials.
4. Run every task in every tool in weeks 1 to 3. Use the same testers and three runs per task per tool, and keep every output.
5. Score blind and verify. Strip tool names, have the task owner score each output against the rubric, and run a fixed verification routine on every output that logs each factual or numeric error.
6. Review security and contracts in parallel. Send each vendor the checklist below and check the data terms, residency, retention, subprocessors and feedback handling.
7. Decide in week 4. Use the weighted matrix below and the cost per accepted output, then write the team tool map.

The 30-day timeline, 5 to 10 testers per tool and three runs per task are AIDOLS planning estimates.

Scoring rubric template

What the pilot costs

Ten testers on monthly ChatGPT Business or Claude Team seats cost $250 per tool (10 × $25). Microsoft's business plans with Copilot have a one-month free trial that converts to paid automatically, without promotional pricing. Claude Enterprise is annual, so test on Team seats.

Score each tool on six weighted criteria and treat a data or residency failure as a veto.

The weights are AIDOLS planning estimates; agree them before anyone sees a score.

The two-tool pattern

In our judgment, the pattern that fits many Microsoft 365 organizations is Copilot for people whose work lives in Microsoft 365 files, mail and meetings, Copilot Chat for the rest, and ChatGPT or Claude for a named group doing cross-source research and analysis. Write the split down as a one-page team tool map that routes work shapes, not vendors, with an owner and a review date.

Two tools mean two sets of controls and contracts. Keep two from becoming five with one data rule, one owner, one review cadence and standing instructions kept as plain text, which limits AI vendor lock-in.

• Permissions. All three inherit your Microsoft 365 permissions, so fix oversharing first, starting with HR, finance and client sites.
• Subprocessor settings. Decide on Anthropic models and OpenAI-operated models (Microsoft-operated OpenAI models stay on either way) in the Microsoft 365 admin center under AI providers operating as Microsoft subprocessors; the change needs an AI Administrator or Global Administrator, and access can be limited to Entra groups.
• Defaults. Turn off "Auto-apply new services" on Copilot Credit spending policies (on by default) so new metered services get reviewed first. Enable GPT-6 Astra by decision, leave Claude's connector write tools off, and set Claude Enterprise retention for projects as well as chats, because project retention overrides chat retention.
• Data rule. Say which accounts count as work accounts, what may be pasted, and whether staff may use feedback buttons.
• Training and measurement. Train people on their own recurring tasks and count change in output; see a Copilot adoption plan measured by output, not logins.

• High-volume processes. A task run hundreds of times a month needs a built process with fixed inputs and an exception rule, not a chat assistant.
• In-country processing in Canada today. None of the three apps documents it as of October 2026, so decide what may go in, or build on Claude through Canadian cloud regions.
• Google Workspace. Evaluate Gemini first; Google includes it on Business Standard.
• No baseline. Measure today's time and error rates before you buy.
• Coding. Compare coding agents separately.
• Licenses only. If you just need licenses bought and assigned, your reseller does that; you do not need an advisor, AIDOLS included.

Send the same list to each vendor and compare the written answers. For a fuller set, the AI vendor security questionnaire has 40 questions as a free CSV.

1. Which tier are we buying, and is "no training by default" in the contract, not only an FAQ?
2. How is thumbs-up or thumbs-down feedback used, and can admins disable it?
3. What are the default and minimum retention periods, and deletion timing, including after termination?
4. Where is data stored at rest, where does inference run, and when do in-country options arrive?
5. Which model providers are subprocessors, and which are on by default?
6. Are any models offered under the model provider's own terms?
7. Do SSO, SCIM and role-based access come with our tier?
8. What do audit logs or the compliance API capture (prompts, files accessed, agent actions), and can they feed our SIEM?
9. Do Microsoft 365 connectors need tenant-wide Sites.Read.All, and do they support site allowlists and sensitivity-label exclusions?
10. Which agent write actions exist, which need approval, and are they logged?
11. What is metered beyond the seat, and which spend caps and alerts exist?
12. How are model changes announced, and can admins restrict models?
13. What does the IP indemnity cover and exclude?
14. Which independent audit reports, such as SOC 2 Type 2 or ISO/IEC 27001, can we read under NDA?
15. How do we export chats, projects and instructions if we leave?

AIDOLS builds training around each team's own tools and tasks, whichever assistant you choose.

• Teach the routing. The AI Fluency course is vendor-neutral by design, and every written lesson is free to read with no sign-up; its which-tool-when unit covers routing by work shape and the team tool map. Paid enrollment adds guided builds, cohort office hours and a graded capstone; individual enrollment is not open yet, and organizations can ask about a cohort.
• Check one workflow first. With the free workflow map, AI agents interview the 2–3 people who run one workflow, and within three business days you get a one-page map and a written verdict on whether each bottleneck needs software, a process change or AI. There is no obligation.
• Bring your shortlist. Book a 15-minute call with your shortlist and bake-off scores.