Microsoft Copilot

We asked Copilot in Excel to build a monthly revenue forecast from 18 months of actuals in a SharePoint-linked workbook, seasonal adjustment, 6-month forward projection. Copilot generated a FORECAST.ETS formula structure in 12 seconds and explained why it chose that approach over a linear trend. Then we asked it to replace all 2024 date references with 2025. It declined, it can only add new columns, not modify existing data. Two prompts, two meaningfully different capability tiers. Neither boundary is clearly communicated in Microsoft's marketing.

That asymmetry is the honest shape of Microsoft Copilot in April 2026: real strengths in Teams recaps and cross-M365 search, hard walls elsewhere, a $30/user/month enterprise price most organizations are struggling to justify, and a privacy controversy that delayed a flagship feature by six months.

Microsoft Copilot at a glance. September 2026

Microsoft Copilot is a family of AI assistants built on OpenAI's GPT-series models and distributed across four surfaces: a free consumer app at copilot.microsoft.com and on Windows 11, iOS, and Android; Copilot Pro at $20 per month; a Microsoft 365 Copilot enterprise add-on at $30 per user per month embedded in the full Office suite; and a free Copilot Chat tier included with eligible M365 licenses. The product launched in November 2023 as a rebrand of Bing Chat, itself the third identity the product had carried in 12 months, a pattern that has become a running frustration for IT administrators. Paid Microsoft 365 Copilot advertises Latest OpenAI models: GPT-5X with priority access and Model choice; Microsoft Learn’s Copilot Cowork picker documents Auto plus GPT 5.5 (Frontier), GPT 5.6 Sol, GPT 5.6 Terra, Opus 5, Claude Sonnet 5, and Claude Fable 5 (Preview-admin/data-retention).

A mandatory disambiguation: this page covers consumer Microsoft Copilot and the M365 Copilot enterprise add-on. GitHub Copilot, the code autocomplete tool for developers integrated into VS Code and JetBrains, is a separate product reviewed on this site as tool 5. The two products share a brand name and a parent company; they share nothing else operationally.

What Microsoft Copilot is actually good at

Teams meeting recaps are the most consistently praised capability across all user research. When transcription is enabled, Copilot generates chapter-level summaries, attributes action items to named participants, and flags open decisions, typically within seconds of a call ending. Two of 14 action items in our test were attributed to the wrong participant (both were speakers with similar-sounding names), and Microsoft's documentation notes that some chapter markers may fail to generate depending on transcription quality. For most meetings, accuracy runs at 90–95%, good enough to use as a working draft, not reliable enough to skip reviewing.

"Teams meeting recaps are Copilot's most reliable capability, generating detailed summaries, action items, and follow-up emails with approximately 90–95% accuracy.". IT consultant review via industry analysis, 2025

Microsoft Graph grounding is the structural advantage no standalone AI tool can match. M365 Copilot grounds responses in the user's actual organizational data, emails, Teams messages, SharePoint documents, OneDrive files, within existing permission boundaries. Querying "What did we agree with this client last quarter?" returns a synthesized answer from relevant emails and Teams threads the logged-in user already has access to. Copilot cannot escalate permissions or surface files the user is locked out of. This is the single most important mechanical distinction from ChatGPT or Claude, which have no live access to your work data.

Python in Excel (Wave 2, September 2024) lets Copilot Pro and M365 Copilot users run Python code natively inside a spreadsheet for advanced data analysis, no round-tripping to a separate environment required. No competitor offers this natively inside a productivity suite. Outlook email drafting is also consistently rated as reliable: Copilot drafts replies with full thread context and adapts tone on request, benefiting from the same Graph grounding as Teams recaps.

"I use ChatGPT for anything creative or research-heavy, it just thinks better. But the moment I need to draft a slide deck or summarise a Teams call, Copilot is unbeatable because it already has the context.". Product Manager, quoted in enterprise AI comparison research, 2026

Where Microsoft Copilot breaks, the failure modes users keep hitting

Excel accuracy is the most documented and underreported problem. Testing by Office Watch in 2026 documented systematic failures in the COPILOT() spreadsheet function: it omitted Gold Coast Airport (population 700,000+) when asked to list airports in cities over 500,000 people, and sorted US Presidents by inauguration date when asked to sort by age at inauguration. Identical prompts returned different results on repeat runs. Microsoft's own support documentation advises the function is "best suited for scenarios where deterministic accuracy is not required" and warns against using AI-generated outputs for "financial reporting, legal documents, or other high-stakes scenarios." That disclaimer lives in help docs, not in the function insertion interface.

The ROI evidence is difficult to read around. A Q1 2026 Gartner Enterprise AI Survey found 74% of companies cannot demonstrate tangible business value from M365 Copilot. Recon Analytics found only 35.8% of licensed users actively use it. At $30/user/month across 1,000 employees, that is $360,000 per year for a tool that two-thirds of seats are not engaging with. The pattern is consistent: pilot programs succeed with early adopters, broad rollout stalls when tenant data quality is poor. PowerPoint generation has its own recurring complaint, generic Microsoft-standard templates requiring substantial manual formatting before any presentation is client-ready.

The Recall privacy controversy is the most cited example of Microsoft's AI ambition outpacing its security readiness. Announced in May 2024, Recall takes a screenshot of everything on screen every few seconds and stores it for AI-powered local search on Copilot+ PCs. Security researcher Kevin Beaumont's DoublePulsar analysis found that early builds stored this data in unencrypted plain-text files accessible to any user-level process. Microsoft pulled the June 2024 launch, delayed to October 2024, delayed again, and finally shipped a redesigned opt-in version in late 2024 with encryption at rest and Windows Hello just-in-time decryption. The feature is now off by default, the correct outcome after a six-month delay. The episode is a fair data point about how Microsoft ships AI features at speed.

Rebranding fatigue is a separate documented complaint: from Bing Chat to Copilot to Microsoft Copilot to Copilot Chat in under 24 months, with features added and removed without clear communication. Copilot was removed from Notepad, Photos, and Snipping Tool in March 2026, leaving users uncertain which integrations are stable.

Microsoft Copilot vs. ChatGPT Team vs. Google Gemini for Workspace

ChatGPT Team ($25/user/month) is $5 cheaper per seat and platform-agnostic, it works regardless of which productivity suite you use. Both products route through GPT-series models. The mechanical difference: ChatGPT Team has no live access to your organizational data; you paste context manually. M365 Copilot reads from Microsoft Graph in real time. ChatGPT Team adds DALL-E image generation (M365 Copilot generates no images) and a broader third-party integration catalog (Dropbox, Slack, Notion). The decision is one question: is live Graph grounding worth $5/month more plus the constraint of M365-only operation?

Google Gemini for Workspace is the structural parallel. Gemini sits inside Gmail, Docs, Sheets, Slides, and Meet the same way Copilot sits inside Outlook, Word, Excel, PowerPoint, and Teams. The critical pricing difference: since January 2025, Gemini AI features are bundled into all Google Workspace Business and Enterprise plans at no extra charge. Copilot costs $30 per user per month on top of an M365 base license. Gemini also offers a 2-million-token context window versus Copilot's 512K. The practical answer: use whichever AI is embedded in the productivity suite you already pay for. There is no cross-compatibility. Copilot does not touch Gmail or Google Drive; Gemini does not touch Outlook or SharePoint.

Is the paid tier worth it?

The free consumer Copilot includes voice, vision, and reasoning (Think Deeper) since February 2025. It caps at 30 chat turns per hour and has no Office app integration. For casual research and drafting, it is competitive with the free tiers of standalone LLMs, just not connected to your work data.

Copilot Pro (Microsoft 365 Premium, $19.99/month) adds Office app integration in Word, Excel, PowerPoint, Outlook, and OneNote; a 256K token context window; DALL-E 3 image generation; and 120 chat turns per hour. For individual M365 subscribers who need draft and summarize inside Office, this is the logical entry point.

The M365 Copilot enterprise add-on is where the math gets hard. A promotional rate of $18 per user per month (annual) runs (promotional period ended), then reverts to $30. Add the M365 Business Standard base license ($12.50/user/month annual) and you reach $42.50 per user per month after July 2026. Against a documented 35.8% active usage rate, the per-active-user cost roughly triples. M365 Copilot earns its price for organizations with high Teams meeting volume, long Outlook threads to manage, and clean tenant data. It does not earn its price for speculative deployments where the expectation is that employees will discover uses post-purchase.

One billing trap: agents in the free Copilot Chat tier are metered at Azure rates, 30 messages per tenant Graph query. Organizations that assumed Copilot Chat was unconditionally free have found unexpected Azure bills when agents run at scale.

Best use cases (and when to skip it)

Use M365 Copilot when: your team spends significant time in Teams meetings (recap value alone can justify the per-seat cost at 10+ meeting hours per person per week); you need AI-assisted Outlook management with full thread context rather than manual pasting; or you are already committed to M365 and the June 2026 promotional rate expiry makes timing the evaluation relevant.

Skip it or wait when: your SharePoint is disorganized, your meetings are not transcribed, or your OneDrive is cluttered. Copilot's output quality scales directly with tenant data governance and weak inputs produce weak outputs; you need auditable, deterministic Excel outputs, the COPILOT() function is explicitly not recommended for financial reporting or legal documents; or PowerPoint generation is the primary use case, since the editing overhead of generic template output is substantial.

Getting started with Microsoft Copilot

The fastest honest test: enable Copilot in Teams and use it on a real meeting. If the recap quality, chapter summaries, attributed action items, matches your actual meeting workload, you have a defensible signal for the subscription. For Excel, treat the COPILOT() function as a structural drafting aid: generate formula frameworks, verify before sharing. The accuracy warnings in Microsoft's own documentation are real.

Enterprise buyers with the June 30, 2026 promotional rate in view: the $18/user/month annual rate versus the $30 post-June rate saves $14,400 per year at 100 users. That is worth factoring into a Q2 2026 evaluation timeline. Last tested: April 2026.