Microsoft Copilot

What Microsoft Copilot is

Microsoft Copilot is Microsoft’s family of AI assistants for conversation, information search, content creation and work within its applications. It is a proprietary service: using it does not grant rights to its underlying technology. The brand covers personal and professional experiences; identifying the account, application and licence is part of understanding the product.

Its practical distinction is proximity to document-based work. A request may end in a chat answer, a document change or a task spanning several applications. These are different outcomes: reading a proposal does not have the same effect as saving changes to a shared file or performing actions with organizational permissions.

Reference documentation: Copilot Terms of Use for individuals.

Background and relationship to other assistants

Development built on Microsoft and OpenAI’s collaboration, with investments in 2019 and 2021 and a new phase announced in January 2023. That partnership connected research, Azure infrastructure and the integration of models into products.

Microsoft retired Cortana in 2023 and listed Copilot among alternatives. This does not mean Copilot is a renamed version of the same programme. Nor should the general assistant be confused with GitHub Copilot, which focuses on software development, or identical models and permissions be attributed to every product sharing the brand. This profile connects each capability to the experience documenting it.

Reference documentation: Microsoft and OpenAI extend partnership · End of support for Cortana.

From Bing Chat to Copilot: 2023

Microsoft introduced the new conversational Bing on 7 February 2023. The initial combination of search and OpenAI models was called Prometheus. On 16 March it announced Microsoft 365 Copilot, bringing together language models, Microsoft Graph and productivity applications.

Bing’s waiting list disappeared in May. November brought two milestones: general enterprise availability of Copilot for Microsoft 365 on the first, and the announcement that Bing Chat would become Copilot, alongside Copilot Studio, on the fifteenth. This sequence explains why an older guide may describe a screen or name that no longer matches the current experience.

Reference documentation: Reinventing search with a new AI-powered Microsoft Bing and Edge · Introducing Microsoft 365 Copilot · Announcing the next wave of AI innovation with Microsoft Bing and Edge · Introducing Microsoft Copilot Studio and new features in Copilot for Microsoft 365.

Expansion and the personal experience in 2024

January 2024 brought Copilot Pro for individuals and wider commercial availability for small businesses, removing the previous minimum licence purchase. In October Microsoft redesigned the personal experience and introduced Voice and the experimental Labs programme, featuring Vision and Think Deeper.

This broadened the idea of an assistant: users could speak rather than type a question, or explore visual assistance in supported experiences. These historical announcements are not a current feature comparison. When reconstructing an older process, retain the date and function name used, especially when the procedure depends on images, speech or document access.

Reference documentation: Expanding Copilot for Microsoft 365 to businesses of all sizes · An AI companion for everyone.

Memory, actions and personalization in 2025

The April 2025 announcement introduced Memory, Actions and Pages, and extended Vision to mobile and Windows. In October Microsoft introduced Mico, an optional avatar accompanying conversations, and new group experiences.

These milestones matter because they mark a move from answering questions towards remembering context and carrying out requests. Memory aims to provide continuity; an action changes something outside the conversation. Within an organization, this difference requires clarity about whether the desired output is an explanation, a draft or an operation. A friendly tone and expressive interface do not amount to knowledge of a workplace or a human relationship with the person asking.

Reference documentation: Your AI Companion · Microsoft stellt 12 neue Funktionen für Copilot vor.

Unification and changes during 2026

Since August 2026 Microsoft has been unifying its Copilot applications while keeping personal and professional accounts separate. Group Chat, Podcasts and Deep Research in the personal application retired on 18 August. They should not appear in a current feature list without explaining that retirement.

On 25 September it announced Home, Code and Autopilot: Home brings together Chat and Cowork; Code shares underlying technology with GitHub Copilot; Autopilot was conceived as a persistent agent. The announcement distinguished Frontier rollout from private preview. It does not establish that all three experiences are enabled for every account.

Reference documentation: Copilot App Updates for Personal, Work, and School Accounts · Introducing the new Copilot with Home, Code and Autopilot.

How it retrieves context and prepares an answer

Copilot combines generative models with retrieved information and the tools of the experience being used. In Copilot Chat, the Work IQ control lets an eligible licence use Microsoft 365 work context; with it switched off, grounding is limited to the web and attached files. Two people can therefore ask the same question and receive different answers because different context is available.

When reviewing an output, distinguish four layers: the request, the material actually accessible, the generated interpretation and the operation performed. If a document is missing, requesting a more detailed answer does not resolve its absence. If two procedures contradict one another, fluent writing may conceal the contradiction rather than resolve it.

Reference documentation: Frequently asked questions about Microsoft Copilot Chat.

Accounts, licences and access options

Professional documentation distinguishes basic Copilot Chat, standard access to Microsoft 365 features, priority access with an additional licence and usage-billed capabilities. Availability spans web, desktop and mobile. These categories should not be confused with personal subscription names.

Before buying or designing a workflow, describe the need in observable terms: consulting an attached file, searching work information, editing a spreadsheet or delegating a long task. Then check which licence and settings support each step. Costs should include consumption where applicable, data preparation and review time; counting only messages sent provides an incomplete picture.

Reference documentation: What is Microsoft Copilot?.

Inputs, files and information quality

File upload supports documents such as PDF and DOCX, XLSX workbooks, PPTX presentations, images, and text or data formats such as CSV and JSON. Accepting a format does not mean correctly interpreting every element it contains.

A scanned document, a table split across pages and a screenshot with small text pose different problems. Useful extraction requires specifying the expected fields, how to flag illegible content and where to retain the reference to the original. In a spreadsheet, an empty cell is not necessarily zero; in a presentation, a speaker note may change a headline’s meaning. Verification should follow those differences rather than stop at confirming successful upload.

Reference documentation: File upload in Microsoft Copilot.

Writing and reviewing in Word

Copilot in Word can prepare drafts, expand and restructure content, rewrite selections, and work from templates or reference documents. Source citations are also supported. Microsoft notes that some experiences roll out gradually depending on licence, version and settings.

For a technical report, the request can separate three jobs: preserving facts, improving order and adapting language for the reader. Asking for all three without priorities may result in a simplification removing an important condition. Useful review compares the draft with the original at the level of claims, figures and exceptions. Keeping corporate formatting supports visual consistency but does not establish that the content has been approved.

Reference documentation: Draft and add content with Copilot in Word.

Analysis and editing in Excel

Excel distinguishes edit, plan and chat modes. Copilot can generate formulas, charts and PivotTables, filter data and modify a workbook. Saved changes are visible to people sharing the file, making the choice of mode particularly important.

A good request asks for an explanation of the formula and preservation of original columns. Calculating delays, for example, requires a cutoff date and a decision about closed records or missing due dates. Accept the result when it can be recalculated and traced back to source rows. A convincing account of a trend does not compensate for an incorrect denominator or an inadvertently applied filter.

Reference documentation: Frequently asked questions about Copilot in Excel.

Presentations and visual summaries

Copilot in PowerPoint can add slides from a reference file. This allows a presentation to start with existing material instead of manually rebuilding it each time.

Visual summarization requires deciding what must remain visible. When turning an occupational safety report into a presentation for managers, a slide about outstanding actions should preserve the date, scope and decision required. Check titles and charts as well: a headline may assert an improvement that the table merely suggests. The source document should remain available for additional context during the meeting, while supporting notes can retain qualifications that do not fit on screen.

Reference documentation: Add a slide from a file with Copilot in PowerPoint.

Image generation and editing

Copilot can create images from descriptions and transform an uploaded image through instructions. This is visual generation, distinct from analyzing a photograph as a source of information.

For training materials, Sabentis proposes using it for conceptual illustrations that do not need to reproduce a real facility. The request should specify the teaching point and which elements must not change. Then review postures, protective measures, tools, signs and text. An attractive scene can depict an inappropriate practice without making that obvious. When the purpose is to document equipment condition or an inspection, retain the real photograph: a generated illustration does not replace that evidence.

Reference documentation: Using Image Generation in Microsoft Copilot.

Meetings, notes and follow-up

Teams supports Copilot use only during a meeting or call, as well as transcript-supported experiences. Later availability depends on that configuration. Even without recording or transcription, prompts and responses may be retained under Microsoft Purview policies.

To turn a meeting into follow-up work, distinguish a proposal, a decision and an accepted commitment. Saying “we could review the guards” does not itself assign an action or an owner. The team can request a table of issues, evidence of agreement, responsible person and confirmed date; cells without agreement should remain unresolved. An automatic note can help prepare minutes but does not establish that every participant has accepted its contents.

Reference documentation: Use Copilot without transcribing or recording a Teams meeting or call.

Voice, images and languages

Documentation distinguishes dictation, which turns speech into text, from spoken conversation, which receives and produces audio. Language lists include Spanish, Catalan and English and distinguish interface, prompts and voice. A translated interface does not guarantee every combination of language and function.

For multilingual safety communication, establish a glossary of equipment, task and sign names. Then check that each version preserves obligations, conditions and sequence. An image or visual conversation can help formulate questions about what is visible; it cannot establish that a facility is safe, measure an exposure or reveal what lies outside the frame. Voice use should take place at times compatible with the attention the work requires.

Reference documentation: Frequently asked questions about voice features in Microsoft Copilot · Supported languages for Microsoft Copilot.

Research with Researcher

Researcher combines multiple sources into structured reports with citations. For personal accounts, Microsoft identifies it as the research experience for Premium and Pro following Deep Research’s retirement; they are distinct experiences. Scope can include the web and work information depending on access.

A document review should specify a concrete question, period and inclusion criteria. The report can organize disagreements and unresolved questions, but every important conclusion needs a source that supports it. A citation to an institution’s homepage does not establish a specific fact. If the report combines an old publication with an update, explain what changed and which version is being used rather than simply collecting links.

Reference documentation: Get started with Researcher in Microsoft Copilot.

Cowork, agents and Copilot Studio

Microsoft announced commercial general availability of Cowork in June 2026 for long tasks using multiple tools. Commercial access combines a user licence with consumption charges; administrators control enablement and budgets. Copilot Studio enables agents with knowledge and actions and supports lifecycle management.

Agent design requires more than an initial instruction. Decide what it can read, what it can change, when it should request review and what evidence it leaves. Preparing a dossier, for example, reasonably separates document selection, production and approval. If a step fails, the agent should retain its state and flag the failure; repeating the entire operation can duplicate files or tasks.

Reference documentation: Copilot Cowork is now generally available · Overview of Microsoft Copilot Studio.

Windows, hardware and other branded experiences

Windows provides application access through Start, the taskbar and shortcuts. The Copilot key is an access route, not a licence. Recall is a separate function: it retains local snapshots to retrieve earlier activity. Microsoft changed its launch in 2024 and added consent and Windows Hello authentication.

In March 2026 Microsoft announced fewer Copilot entry points in several Windows applications. In September, Windows Central reported a gradual withdrawal of the Copilot+ PC hardware brand. These presentation changes do not mean that the assistant is disappearing or that every capability runs locally.

The brand also reached gaming: Xbox documented a mobile beta in November 2025. In May 2026 its mobile wind-down and cancellation of the planned console release were reported; the available report did not settle every PC version’s status. It is therefore not presented as a current professional capability.

Reference documentation: Access Microsoft Copilot on Windows · Update on the Recall preview feature for Copilot+ PCs · Our commitment to Windows quality · Microsoft's week of change: Copilot+ branding quietly retired · Xbox November Update: Gaming Copilot Available on Mobile · Xbox CEO Asha Sharma kills Copilot for Gaming.

Work information and access controls

Microsoft states that professional Copilot prompts, responses and Graph data do not train foundation models, and access follows user permissions. Documentation specifies conditions for model providers; it currently excludes Anthropic models as a subprocessor from the EU Data Boundary.

The operational implication is to review permissions before expanding use. If too many people can read documents in a folder, the assistant does not correct that decision. Separate drafts, approved versions and restricted records, and test access with representative profiles. Also examine connected services and shared outputs: a summary can bring previously scattered information together on a single page.

Reference documentation: Data, Privacy, and Security for Microsoft Copilot.

Memory and data in personal accounts

The current personal application has controls for memory, history, web search and personalized advertising. Switching memory off does not automatically delete saved memories: deletion is managed separately.

This distinction matters when a tool is used over a long period. Changing a preference, deleting a conversation and removing a file are different operations. For occupational safety work, start with the organization-approved account and only the necessary material. Do not move information between accounts for convenience without checking its destination. A fictional example allows a workflow to be rehearsed without turning personal history into a parallel archive of workplace cases.

Reference documentation: Microsoft Copilot for individuals: your privacy controls and choices.

Adoption and results: what the evidence supports

The UK government’s September–December 2024 trial distributed 20,000 licences. Its June 2025 report estimated 26 minutes of daily savings from participants’ responses and recorded difficulties with complex contextual information. This was a self-reported estimate, not a transferable guarantee.

A 2026 study of approximately 5.5 million Microsoft 365 Copilot Chat sessions observed use concentrated in writing, search, analysis and communication. Analyzing sessions shows how a product is used; it does not itself establish a causal improvement in productivity or safety.

An occupational safety service can conduct its own assessment by comparing total report preparation and review time, errors found and usefulness for decisions. Ease of producing more text should not be the main measure.

Reference documentation: Microsoft 365 Copilot Experiment: Cross-Government Findings Report · AI in the Enterprise: How People Use M365 Copilot Chat.

Reliability, security and intellectual property

During Bing’s initial trial in February 2023, extended conversations produced inappropriate responses. Microsoft acknowledged the problem and imposed temporary limits of five exchanges per session and fifty per day. These are historical facts, not current limits or evidence that the system has emotions.

June 2025 brought publication of CVE-2025-32711, an information-disclosure vulnerability involving instruction injection. BleepingComputer reported that Microsoft had fixed it server-side in May and had no evidence of real-world exploitation. The case illustrates why retrieved content should be treated as material rather than authority to change the task.

In 2023 Microsoft announced a conditional commitment to defend certain copyright claims. It does not make all generated material free of rights. When publishing an image, extract or presentation, retain provenance and review rights to the elements used as well as the generated result.

Reference documentation: The new Bing & Edge: Learning from our first week · The new Bing & Edge: Updates to Chat · CVE-2025-32711: Microsoft CNA record · Zero-click AI data leak flaw uncovered in Microsoft 365 Copilot · Microsoft announces new Copilot Copyright Commitment for customers.

Public-information errors: interpreting the tests

In May 2024, CORRECTIV published its own test of twelve questions in three languages and documented Copilot errors about European election candidates. This is a dated example of public-information problems, not a measurement of current performance.

The lesson for a reference profile is to separate verified facts, observed examples and generalizations. An incorrect answer deserves correction and context; it cannot establish an overall error rate. Equally, answering one question correctly does not guarantee accuracy when the language, sources or date change. For information affecting a decision, a link to the specific source is more useful than the assistant’s general reputation.

Reference documentation: Chatbots: Don't Bother Asking AI About the EU Elections.

Applications in occupational safety and health

Sabentis proposes using Copilot to prepare information that safety professionals will subsequently check and approve. Its potential contribution lies in connecting documents, data and communication while retaining the management system as the reference for the actual status of actions.

• Preventive action follow-up: turn an action export into a draft distinguishing deadlines, incomplete records and pending decisions. “Closed” must be defined: it may mean administrative closure without an effectiveness check.
• Communication with managers: turn an approved report into a short presentation while preserving limitations and decisions requiring resources or coordination.
• Meeting preparation: assemble questions about discrepancies between minutes and records. The aim is to make clarification needs visible, not assign causes or responsibilities absent from the evidence.

These are proposed uses, not results of a test conducted by Sabentis. Preventive value arises when the material helps follow up a particular measure. A better presentation does not demonstrate reduced risk or replace observation, consultation or measurement where needed.

Hypothetical case: a monthly preventive action report

A team needs to report to management on actions across several sites. It has an authorized, anonymized export, but statuses have not been used consistently. Copilot’s task is to prepare a draft that can be reconciled with records, not decide which measures are effective.

1. Prepare the input. Retain the original export and work with identifier, site, functional owner, due date, status and available evidence. Remove unnecessary personal data and establish a cutoff date.
2. Define rules. Agree what counts as open, overdue, duplicate or undated. Keep “implemented” and “effectiveness checked” separate if the system distinguishes them. Prohibit filling blanks as though they were facts.
3. Separate outputs. Request a count table with identifiers and then a narrative based exclusively on that table. List excluded records with reasons.
4. Reconcile. Recalculate totals and categories in the spreadsheet. Check a sample of each status and every ambiguous case; resolve filters and criteria before writing conclusions where discrepancies arise.
5. Prepare decisions. Identify missing information, who can clarify it and which conclusion remains open. Do not attribute delays to insufficient training or resources without evidence.
6. Approve and present. The responsible person reviews interpretation, dates and recipients. The final presentation preserves the period and exclusions and links to the approved report.

The acceptance criterion is concrete: totals reconcile, every quantitative claim is reproducible, and each highlighted action links back to its record. If an action is marked closed but lacks evidence, the report must reflect that absence. The case does not establish implementation of measures on site or turn the summary into an audit.

Limitations affecting interpretation

Microsoft’s transparency note acknowledges that Copilot can invent information or misrepresent even a reliable source. A visible reference helps verification but does not eliminate that risk.

In occupational safety, the most important mistakes may be subtle: confusing sites with different activities, changing a condition in an instruction, combining records from different periods or presenting missing data as an absence of incidents. Review should concentrate on those decision points.

Also acknowledge what the task does not cover. An action report describes records; an effectiveness evaluation requires additional criteria and evidence. If the source cannot answer the question, the correct result is to identify the missing information. Asking the assistant to sound more certain does not improve the documentary basis.

Reference documentation: Transparency Note for Microsoft Copilot (for individuals).

Choosing and complementing Copilot

Copilot fits work already dependent on Microsoft documents, meetings and files where the aim is to reduce preparation steps. Keep reproducible calculations in Excel. If the aim is to publish recurring indicators based on a maintained data model, Power BI addresses a different need.

Compare tools on a specific process using the same input information. Measures can include numerical corrections, omissions, review time and ease of updating the output. If verifying the text costs more effort than preparing it, narrow assistance to organizing questions, structuring a document or improving part of the wording. The useful decision is which task to delegate and under what conditions.

Frequently asked questions

Does the unified application mix personal and work information?

Microsoft describes accounts as separate by design. Nevertheless, check the active account and file destination before attaching information. Similar-looking experiences do not authorize moving data between them.

Is a Microsoft 365 licence sufficient?

It depends on the capability. Check standard or priority access, available work context, administrator settings and additional consumption where agent work is involved.

Can it certify that a preventive measure works?

No. It can organize supplied evidence and identify gaps; effectiveness requires a criterion and a real check. A closure status and a generated conclusion are insufficient.

Reference documentation: Copilot App Updates for Personal, Work, and School Accounts · What is Microsoft Copilot?.