Microsoft 365 Copilot is easy to demo and much harder to live with. The launch videos show a Word draft appearing from a single sentence and a Teams meeting recap that sounds like a colleague wrote it. Six months into a paid rollout, most IT teams are dealing with messier questions: why the Excel answers disagree with the finance model, why the SharePoint results include a document from a project nobody should see, and why the $30 seat has not reduced inbox time as promised.
This Microsoft 365 Copilot limitations review focuses on the gaps that show up after the pilot, not the polish of the keynote. The product is genuinely useful in narrow workflows. It is also bounded by licensing, permissions, app coverage, and the quality of your own data.
The $30 per-user price is only the visible cost
At $30 per user per month on an annual commitment, 100 seats cost $36,000 a year before you pay for the underlying Microsoft 365 licence. Copilot is an add-on. It does not replace Business Standard, Business Premium, E3, or E5. A 50-person firm on Business Premium can easily spend more on Copilot than on the base productivity suite, then discover that safe deployment needs extra work on labels, DLP, and SharePoint permissions.
That extra work is where budgets drift. A Copilot rollout is not a software install; it is a data governance project with a chat box on top. If you need Purview, advanced eDiscovery, or tighter DLP to make Copilot safe, the real project cost climbs well beyond the per-seat line. For a broader framework on comparing AI tool spend, this value-for-money review of AI tools for business is a useful sanity check.
Copilot is bounded by Microsoft Graph, including your permission mistakes
Copilot does not browse the open web to answer work questions. It queries Microsoft Graph, the same permission layer that controls SharePoint, OneDrive, Outlook, and Teams. That sounds safe until you remember Graph mirrors every inherited SharePoint permission, guest link, and stale Teams site. If the marketing folder is open to “Everyone except external users,” Copilot can summarize it for anyone who asks.
Data hygiene is not optional. Sensitivity labels and DLP help, but they are configuration projects, not toggles. A common failure mode is a pilot that looks clean because it uses a small test site, followed by a company-wide launch where Copilot surfaces an old salary spreadsheet or a draft reorg memo. The limitation is not the model; it is the state of your tenant.
It will not reliably find the one file you need
Copilot retrieves content; it does not know your naming conventions. If a contract sits on a local drive, an unindexed network share, or a third-party app with no Graph connector, Copilot won’t see it. Even inside Microsoft 365, a file called “Final_v7_approved_use_this_one.docx” may rank below an older document with a cleaner title. Search quality depends on metadata, freshness, and the words in the file.
The app-by-app reality check
There is no single Copilot skill level. Each Microsoft 365 app exposes a different slice of the model, and the gaps are practical:
- Outlook: Summarizes long threads and drafts replies, but it won’t cleanly triage an inbox, enforce your filing system, or replace rules. A 40-message thread with two decisions can read like one.
- Excel: Works best on clean tables. It can suggest formulas and spot trends, but it will not audit a messy 12-tab model or guarantee the formula points at the right column. Finance teams still need a human check.
- Word: Good for first drafts, rewrites, and summaries. Long policy documents can drift; Copilot may miss section-specific clauses or produce a citation that looks plausible but does not exist.
- PowerPoint: Generates a starter deck fast. Brand templates, charts, and layout polish remain manual, and sales teams often rebuild half the slides.
- Teams: Meeting recaps depend on transcription being enabled, consent handled, and the meeting being the right type. The recap is useful, not a legal record.
The pattern is consistent: Copilot accelerates the first 30% of a task and leaves the judgment-heavy last mile to you.
Hallucinations don’t stop at the firewall
“Grounded in your data” does not mean correct. Ask for a summary of a client account and Copilot can blend two similarly named projects, or quote an older statement of work because it ranked higher in search. In Excel, a generated formula can be syntactically perfect and semantically wrong. In Word, a confident paragraph can cite a policy clause that was deleted last quarter.
Build a review step into any workflow that touches money, contracts, HR, or customer commitments. Better prompts help, but they are not a control. The teams that get value from Copilot treat it like a fast junior assistant: useful for a first pass, never the final signature.
Long documents, short context, and the retrieval gap
Copilot does not read your entire 180-page employee handbook every time you ask a question. It retrieves relevant chunks and works from those. If the answer depends on a table in appendix C, ask for appendix C explicitly and check the output. If a project update pulls from five files, name the five files.
For recurring work, create a curated SharePoint page or a clean reference file instead of hoping the model finds the right needle. The limitation is practical: Copilot is a retrieval assistant, not a document database with perfect recall.
Admin controls and support are a second project
Deployment is not a license assignment. You need to decide which users get it, which data sources are in scope, how to label sensitive content, and how to measure adoption. Microsoft’s admin center gives logs and usage reports, but it won’t tell you whether the output is any good.
Support can feel like a do-it-yourself project. If you want a flavour, this six-month review of Copilot Pro support shows how much troubleshooting lands on the customer. For Microsoft 365 Copilot, the same pattern appears: the partner or IT team owns prompt training, champion programs, and the inevitable “why did it say that?” tickets.
Where the competition looks stronger, and where Copilot wins
Google Gemini for Workspace has its own limits, especially around plan requirements and app coverage. This Gemini for Workspace limitations review is a good reality check if you are weighing ecosystems. ChatGPT Enterprise offers flexible custom pricing and strong general reasoning, but it is not natively wired into Outlook, Teams, and SharePoint. The ChatGPT Enterprise review explains what the custom price actually gets you.
Copilot’s advantage is context: it can use the meeting you just had and the file your colleague shared in Teams. That integration is also the source of its biggest risks. A standalone chatbot cannot leak a SharePoint folder it never saw. Copilot can, if the permissions allow it.
A sensible pilot plan before you buy more seats
Pick 30 users in two functions, not the whole company. Set a baseline: minutes spent on meeting notes, time to first draft, and time to find a document. Run for six weeks and review weekly. Test five ugly tasks, not five polished demos:
- Summarize a 60-message thread with two decisions and one open question.
- Find the latest approved contract and list the three clauses that changed.
- Draft a project update from three files with conflicting dates.
- Build a formula on a messy table and explain the assumptions.
- Recap a meeting where one person disagreed with the group.
Track correction rate: how often a human has to fix Copilot output. If it is above 30%, the workflow is not ready. Then compare cost per accepted output, not cost per seat. For pricing benchmarks across vendors, this AI chatbot pricing comparison helps frame the negotiation.
Microsoft 365 Copilot is a useful assistant for some teams. It is not a plug-in that fixes bad data, unclear processes, or weak adoption. Treat this limitations review as part of the buying decision, not a footnote after the invoice.

