Copilot vs ChatGPT for Office: which licence earns its keep

Most people in a Hong Kong office never chose their AI tool. The firm has a Microsoft agreement, IT switches on Copilot, and the question reaches you as a line on a budget: US$21 per user per month for Copilot Business on an annual subscription, before any promotional discount and on top of a qualifying Microsoft 365 licence. So the real question in Copilot vs ChatGPT for Office work is not which model is cleverer this quarter. It is which one anybody will still be using in March.

The short version. Copilot wins on reach into your own material and on where the data sits. A standalone chat tool wins on the quality of the thinking. If your week is assembling things out of documents that already live in SharePoint, Outlook and Teams, the Microsoft licence earns its keep. If your week is analysis, argument and drafting from a blank page, Copilot will disappoint you and you will quietly keep a second tab open.

What you are actually buying

Copilot is not a better model. Underneath, it is broadly the same class of model you would reach on a consumer site. What you are paying for is a retrieval layer and a permissions boundary: it can see the files, mail and meetings your account can already see, inside your firm's tenancy, and it inherits your access rights rather than granting new ones.

In a regulated firm that boundary is the whole argument. It is the difference between a tool your InfoSec team can reason about and a website your staff are pasting client names into. Nothing about the answers is better. Everything about the approval conversation is.

There is a catch that nobody mentions in the demo. Copilot inherits your permissions mess as faithfully as your permissions. If a folder of comp spreadsheets was shared to the whole company three years ago and nobody noticed, the tool that reads everything you can read will find it and quote it back in a meeting. Firms that turn Copilot on before they audit their sharing discover their oversharing the interesting way.

Copilot vs ChatGPT for Office tasks: the split by job

Two rows deserve their own sentence.

Excel is where finance people expect the most and get the least. On a clean demo table Copilot is impressive. On a real workbook, with merged cells, eleven tabs, three years of someone else's naming conventions and a circular reference somebody disabled in 2019, it struggles. The useful pattern is the opposite of the advertised one: paste the formula or the block into a chat tool and ask what it does and where it breaks, rather than asking anything to operate the model for you.

PowerPoint generation flatters itself. You get a structurally sound deck with the argument sanded off. For an internal update that is a fine trade. For anything going in front of a client, the fastest route is still a board deck built from your own skeleton, with the machine doing the assembly underneath it.

The Hong Kong complication

Here the comparison stops being about features. As of 4 October 2026, Hong Kong is absent from the published supported-location lists for both ChatGPT and Claude. Some staff still find ways to use them through personal accounts or a VPN, outside the firm's approved route. That creates a policy and data-handling question before it creates a productivity gain.

Copilot can be procured through the Microsoft 365 stack a firm already governs, subject to its own licensing and data review.

That asymmetry can decide a Hong Kong corporate purchase before any benchmark does. It is also why the honest answer for a lot of firms is both: Copilot as the licensed tool inside the perimeter, and a properly procured route to a stronger model for the small number of people whose work is analysis rather than assembly. An enterprise route through a cloud provider's hosted models is the usual answer there, and it is a procurement exercise, not a credit card.

Two caveats on all of that. Availability and enterprise terms move, sometimes quickly, so check the current position before you plan a year around it. And none of it settles what your staff may paste into anything, which is a separate conversation entirely, covered in whether ChatGPT is safe for work data.

What about Gemini

Same logic, different suite. If your documents and mail live in Google Workspace, Gemini has the grounding advantage that Copilot has in Microsoft, and the comparison collapses to the same question: which suite holds the material you work on all day.

Very few Hong Kong financial firms are on Workspace, which is why this is a paragraph and not a section. For the professional services firms and startups that are, read every Copilot argument above with Gemini in its place and the conclusion barely changes.

The licence is not the decision. The decision is whether anyone is still using it in March.

The price is not the licence

Forty Copilot Business seats at the US$21 annual-subscription list price cost US$10,080 a year before any discount or local taxes. Microsoft lists a separate enterprise Copilot plan at US$30 per user per month. Check the current Hong Kong offer before you budget. That is not a big number for a bank and it is not the number that matters.

The number that matters is what share of those seats are live in month six. AI licence budgets often assume adoption is automatic, because the licence is the part with a price on it and the habit is the part with nobody's name against it. A seat that goes dark in November costs the same as one that saves four hours a week.

This is the same failure as the prototype-trap workshop, arriving through finance instead of through training. Something gets bought, a launch email goes out, and nobody owns the Tuesday.

Decide it in a week, not in a committee

Pick five tasks your team genuinely repeats: the Monday pack, the client follow-up, the credit summary, the variance commentary, whatever yours are. Run each one twice, once in Copilot and once in a standalone tool, on real work with real constraints. Write down the minutes and whether the output needed rewriting.

You will get an answer in a week, and it will be specific to your firm rather than to a vendor's benchmark. A likely pattern is that Copilot takes the assembly jobs and a standalone tool takes the thinking jobs. How much of your team's week falls into each is the question the test answers. Yours may land differently, which is the reason to run it rather than argue it.

Thirteen years at a bank left me with one conviction about tools: the licence decision is the easy half. The half that decides whether the money was well spent is whether five specific tasks got faster, and that is a habit question, not a procurement question. If you want that tested on your own work rather than argued in a paper, the corporate AI training page sets out how those sessions run, and the prompt library is free if you would rather start on your own.