A practice manager I work with showed me three vendor decks she had been sent in a fortnight. Every one of them used the word copilot in the first slide. By the third deck she said, half-laughing and half-frustrated, “What does any of them actually mean by it?”
It is the right complaint. By 2026 the word copilot has become a category label that nearly every productivity vendor uses, and it has stopped sorting the genuine integrations from the rebranded chatbots. The plain-English version is worth knowing, because the per-seat fees and the data implications are real even when the marketing is not.
What is a copilot?
A copilot is an AI assistant that lives inside the software you already use, rather than a separate window you log into. Microsoft 365 Copilot sits inside Word, Excel, Outlook and Teams. GitHub Copilot sits inside the developer’s code editor. Salesforce Einstein Copilot sits inside the CRM. Google’s Gemini for Workspace sits inside Gmail, Docs and Sheets. The defining trait is integration at the interface layer, not a different underlying AI engine.
The contrast worth holding onto is with a chatbot. A chatbot like ChatGPT is a separate interface that only knows what you paste into it. You copy your spreadsheet data in, the chatbot answers, you copy the answer back. A copilot reads the spreadsheet directly, sees the structure, the formulas and the headers, and works on the live thing. Same underlying AI in many cases. Different relationship to your work.
That difference matters because the per-seat fees are priced as if integration is the product. A genuine copilot saves the round-trip and the retyping. A chatbot dressed up in a side panel does not.
Why it matters for your business
The first thing it changes is friction. Tasks that involve switching tabs, retyping context and pasting answers back disappear when the assistant is already inside the tool. An accounts team using Excel with a copilot can ask it to flag duplicates or summarise spend without leaving the spreadsheet. The same team using a standalone chatbot has to export, paste, wait, and paste back. Across a team of ten people doing several of those round-trips a day, the time difference is material.
The second is fit. A copilot that reads your real CRM records, your actual customer history and your own document templates can offer suggestions calibrated to your business, not to a generic SME. The quality jump is biggest where your team has clean data and consistent processes, and smallest where everything is ad-hoc. A copilot inherits the discipline of the host tool. It cannot impose discipline that is not there.
The third is cost honesty. Microsoft 365 Copilot adds roughly £20 to £30 per user per month to your existing Microsoft subscription. GitHub Copilot is around £8 per developer. Gemini for Workspace lands at £18 to £24 per user. For a ten-person team that is between £200 and £300 a month, or £2,400 to £3,600 a year. The maths works easily if the copilot reclaims even one full day a month per person. It does not work at all if the team treats the side panel as an occasional novelty.
Where you will meet it
You will meet the word in nearly every productivity, CRM, helpdesk and developer-tool pitch you take. “We have a copilot for that” is the standard answer to almost any feature question in 2026. Sometimes the feature is a deeply integrated assistant that reads your live data. Sometimes it is a chat panel pinned to the side of an existing screen. The marketing language is the same in both cases.
You will also meet it inside products you already pay for. Microsoft 365 Copilot is bundled into the higher Microsoft 365 business plans. Salesforce Einstein Copilot is included at the Professional tier and above. Slack AI, Notion AI, Zendesk Copilot and HubSpot AI all use the same framing. The likely state of affairs in your business is that several copilots are already available to you and your team is using them unevenly.
The most useful place to meet the word is in a vendor demo. Ask the vendor to show the copilot working on your kind of data, in the host tool, without you pasting anything into a chat box. The genuine ones do this comfortably. The marketing ones suddenly want to walk you through a slideshow.
When to ask about it, when to ignore it
Ask hard questions when the copilot will read employee, customer or financial data. The legal frame in the UK is unambiguous: if a tool processes personal data, you remain the controller and you owe transparency, a lawful basis and a Data Processing Addendum with the vendor. The ICO published updated guidance on AI and on workplace monitoring in 2025 covering exactly this scenario. ACAS guidance on workplace AI adds the employee-relations angle. The fact that a copilot is bundled with software you already pay for does not remove any of those obligations.
Ask hard questions on integration depth too. “Does the copilot read my live document, or do I need to provide context in a chat panel?” If the answer is the second, the per-seat fee is buying you a chatbot under another name and a free ChatGPT or Claude subscription would do the same job at a lower price.
Ignore the term entirely when the work is one-off, low-stakes and easy to check. A team using a copilot to draft an internal status update does not need to interrogate the architecture. The question worth asking is whether the suggestion saves time and the team trusts it enough to use it. If yes, the label is irrelevant. If no, no amount of vendor framing will change that.
There is one trap worth flagging. A copilot that takes actions on your behalf, sending an email, updating a CRM record, scheduling a call, is doing more than suggesting. Some vendors are starting to call this “agentic” capability. Treat it as a different category. The governance question shifts from “did the suggestion help” to “what just happened on my behalf and can I undo it.”
Related concepts
A chatbot is the contrasting category. Same underlying AI in many cases, separate window, no native context. ChatGPT, Claude and the standalone Gemini app are chatbots in this sense. Useful, but a different shape of tool.
Embedded AI is the broader umbrella that copilots sit inside. Spell-checkers, recommendation engines and writing-style suggestions are all embedded AI. What makes something a copilot rather than just embedded AI is usually that it offers proactive drafts and suggestions, not only passive checks.
An agent is the next layer up. A copilot suggests; an agent acts. Vendors are beginning to use the word agentic in 2026 to describe copilots that can send the email, update the record or book the call without per-action approval. That changes the risk profile and deserves its own conversation.
System prompt is the standing instruction set behind a copilot’s behaviour. Enterprise versions of most copilots let you customise this so the assistant follows your house style and stays inside your guardrails. Worth asking the vendor whether you control the system prompt or whether it is locked.
Function calling is the technical capability that lets a copilot trigger actions in other systems, send the email, file the record, run the workflow. It is what turns a copilot from a draft-helper into something closer to an agent. Helpful when you need it, governance work when you have it.
The point of the vocabulary is not to make you fluent in vendor architecture. It is to give you enough purchase that the next time someone says “we have a copilot for that” you can ask the question that separates the integration from the wrapper.
