We ran Claude and Microsoft Copilot against the same three months of real mailbox and calendar data — five judgment-heavy prompts, scored blind by a third AI with no stake in either. Claude took Round 1 four to one. The decisive factor was not raw capability but a willingness to name uncomfortable patterns rather than tidily catalogue activity.
Most "Claude vs Copilot" comparisons test the wrong thing. They run clean, synthetic prompts in controlled sandboxes and report which model writes a tidier paragraph. That tells you almost nothing about how either tool behaves on your actual work — the messy, three-months-deep reality of a real inbox and a real calendar.
So we ran the test we actually wanted to see. The M365 Bake-Off puts Claude and Microsoft Copilot against identical, genuine Microsoft 365 data and asks them the kind of questions a busy operator actually needs answered. Round 1 is Outlook: mail and calendar. This is independent testing by Acuity AI Advisory, with no involvement from Microsoft or Anthropic.
How the test was run
Both tools accessed the same data through the Microsoft Graph API: roughly three months of real commercial emails, calendar entries, and professional relationships. No synthetic data, no curated test environment, no prompts engineered to flatter one tool over the other.
- Tool A — Microsoft Copilot (GPT-5.5 Thinking)
- Tool B — Claude (Opus 4.8 Extended)
- Arbiter — Grok 4.3 (xAI), scoring outputs blind, without knowing which tool produced which answer.
The arbiter matters. Grok was chosen precisely because it has no commercial stake in the outcome — it sells neither Copilot nor Claude. Each answer was scored on its merits, with the labels hidden.
The five prompts
The prompts were deliberately judgment-intensive. Anything can summarise an email thread; the interesting question is whether a tool can reason about what the data means.
- Strategic Drift — where has my time and attention actually shifted, and is that aligned with strategy?
- Pipeline Momentum — which commercial conversations are moving, which have stalled, and why?
- Relationship Audit — which professional relationships are two-way, and which have gone dormant?
- Commitment Tracking — what have I promised and not delivered, ranked by how late it is?
- What Actually Matters — the three most important things to handle before Friday. Not the most urgent — the most important.
The result: Claude 4, Copilot 1
Scored blind, prompt by prompt:
- Strategic Drift → Copilot (9.0 vs 8.5)
- Pipeline Momentum → Claude (9.0 vs 8.7)
- Relationship Audit → Claude (9.0 vs 8.7)
- Commitment Tracking → Claude (8.9 vs 8.4)
- What Actually Matters → Claude (9.2 vs 8.5)
Two things stand out before the verdict. First, the scores are close — both tools are genuinely capable, and Copilot's single win was a clean one. This is not a story about one tool being bad. Second, the margins widened on the prompts that required judgement rather than comprehensiveness.
What separated them
The clearest way to describe the difference: Copilot behaved like a thorough consulting analyst; Claude behaved like a sharp strategic advisor.
Copilot's strength was structure. It produced comprehensive, well-organised output with ready-to-use frameworks — the kind of thing you could paste into a deck. On Strategic Drift, that thoroughness won the prompt outright.
Claude's strength was signal. It carried a higher signal-to-noise ratio and showed more willingness to name the uncomfortable thing — the relationship that has quietly gone one-way, the commitment that is now embarrassingly late, the priority everyone is avoiding because it is not on fire. On the final prompt, the arbiter specifically noted that Claude "better embodied the 'not the most urgent — the most important' distinction" by tying its recommendations back to strategic priorities rather than to the loudest items in the inbox.
That is the through-line across all five prompts. The gap was not a capability gap — both tools could see the same data and reason competently about it. It was closer to a character gap: a willingness to surface inconvenient truths instead of producing a complete-but-comfortable catalogue of activity.
Why this matters for how you buy and deploy AI
For Irish organisations standardising on Microsoft 365, the instinct is to take Copilot because it is already in the tenant. That is a reasonable default, and on structured, framework-style work Copilot earns it. But the Bake-Off is a reminder that "which AI is best" is the wrong question. The right question is best at what — and for the judgement-heavy work that senior people actually need help with, the answer is not automatic.
This is exactly the kind of decision we help organisations make on the evidence rather than the brochure. Our independent Claude vs Copilot comparison for Irish organisations sets out the trade-offs for a buying decision, and the diagnostic behind this test — analysing real M365 signal for priority drift and cognitive load — is the basis of our Cognitive Mirror executive diagnostic. If the goal is getting genuine return from a Copilot rollout rather than just licences, the Copilot ROI work is where that starts.
Round 2: Word
Round 1 was about reasoning over data. Round 2 moves to production — both tools as native Word add-ins on the same board-level document, same blind-scored methodology. The result complicates the story told above: Claude won on output quality again, and still lost the round. Worth reading the two together before drawing conclusions from either.
If you want this run against your own environment rather than ours — your mailbox, your calendar, your real priorities — that is what the Cognitive Mirror diagnostic does. For a straight, vendor-neutral read on whether Claude or Copilot fits your organisation, start with the Claude vs Copilot comparison or just talk to us.