Most firms treat this as a quality question: which one writes better? That’s the part of the decision that ages fastest. The models leapfrog each other every few months, and whichever is ahead today will not be ahead when your licence comes up for renewal.
The questions that survive renewal are duller. Where client data goes and who holds it. What your existing Microsoft agreement already covers. What your compliance officer can produce when a regulator or a client asks what the firm has been doing with AI. And whether anyone will actually use the thing you bought.
For a firm already on Microsoft 365, those four questions point fairly firmly in one direction. Here’s the working.
The four products you are comparing
There are four different products behind the two brand names, and mixing them up is where most of the confusion starts. Note that Microsoft has renamed two of them: its documentation now states that “Microsoft 365 Copilot is now named Microsoft Copilot, and Microsoft 365 Copilot Chat is now named Microsoft Copilot Chat”, with no change to the security, compliance and privacy position.
- Microsoft Copilot Chat, available with a work account. A capable assistant that does not read your firm’s files.
- Microsoft Copilot, the paid per-seat licence. The one inside Word, Excel, Outlook and Teams, grounded in your own mail and documents through Microsoft Graph. Microsoft’s licensing documentation lists it as included in Microsoft 365 E7 or sold as an add-on, with a Microsoft Copilot Business add-on for organisations of up to 300 users on an annual commitment.
- ChatGPT on a business tier. OpenAI’s trust portal names ChatGPT Enterprise and ChatGPT Edu, alongside the API, as the services inside its certification scope.
- ChatGPT on a free or personal account. Not a procurement option. It is where incidents come from, and it is already in your firm whether you approved it or not.
The two provider guides go deeper on each: Microsoft Copilot and ChatGPT. What follows is only the comparison between them.
The comparison that matters
| Microsoft Copilot (paid) | ChatGPT (business tier) | |
|---|---|---|
| Where processing happens | Inside your Microsoft 365 tenant, under your existing agreement verified 22 Oct 2026 | OpenAI’s platform, under a separate contract you negotiate and sign |
| Training on your data | Microsoft states prompts, responses and Graph data “aren’t used to train foundation LLMs” verified 22 Oct 2026 | Not verifiable on OpenAI’s own pages at review. Require the exclusion in writing, in the DPA, naming your tier |
| Reads your firm’s own files | Yes, through Microsoft Graph, limited to what the signed-in user can already see verified 22 Oct 2026 | No, unless you build it. Staff work by pasting |
| UK and EU data handling | EU traffic stays inside the EU Data Boundary, with Anthropic models currently excluded from it verified 22 Oct 2026 | API data residency options include the United Kingdom for storage, with regional processing in the US, Europe and UAE verified 22 Oct 2026 |
| Audit trail | Prompts and responses land in the unified audit log, with eDiscovery and retention through Purview verified 22 Oct 2026 | ChatGPT Enterprise is a supported enterprise AI app in Purview, and OpenAI provides its own admin and compliance logging |
| Independent assurance | ISO 27001 and ISO 42001 among Microsoft’s stated certifications for Copilot verified 22 Oct 2026 | SOC 2 Type 2 covering July 2025 to June 2026, plus ISO 27001, 27017, 27018, 27701 and 42001 verified 22 Oct 2026 |
| Licence shape | Per seat, annual, on top of a qualifying Microsoft 365 plan | Per seat for the app, per use for the API |
| Characteristic failure | Licences bought in volume, used by nobody | Staff on personal accounts because no approved tier was bought |
Prices are deliberately absent. Microsoft’s own licensing page warns that “licensing availability and product terms can change”, OpenAI’s commercial pages refused automated access on the day of this review, and a figure I cannot verify today is worth less to you than none. Get both quotes in writing, for your tenant, in sterling.
Where the client data goes
Client data either stays inside your Microsoft tenant, goes to OpenAI under a new contract, or leaks into a personal account. A compliance lead should read this section twice, because it decides whether either tool is defensible under UK GDPR, and the ICO expects a data protection impact assessment before new processing of this kind begins, not after [1][2].
Inside the Microsoft tenant
Copilot’s argument is that nothing new leaves the building. Microsoft’s documentation states that prompts, responses and data accessed through Microsoft Graph are not used to train foundation models, that interactions are encrypted at rest, and that Copilot “only surfaces organizational data to which individual users have at least view permissions”. For EU users it commits that traffic stays within the EU Data Boundary, with the honest caveat, in Microsoft’s own text, that models provided by Anthropic as a subprocessor are currently excluded from that boundary. If your DPIA names the EU Data Boundary as a control, that exclusion belongs in it.
The risk Copilot brings is inward rather than outward: your existing permissions become searchable. A folder somebody was given access to in 2021 stops being obscure the moment a natural-language question can find it. Tidy sharing on sensitive libraries before switch-on, not after the awkward discovery.
Inside OpenAI
Here you are signing a new processor relationship rather than extending one. OpenAI’s developer documentation states that data sent to the API “is not used to train or improve OpenAI models” unless you opt in, that abuse-monitoring logs are retained for up to 30 days by default, and that zero data retention is available subject to its prior approval. Data residency is a project configuration option, and the United Kingdom appears in its region list, though for storage rather than regional processing.
For the ChatGPT app tiers, OpenAI’s own commercial pages would not load for verification on 22 October 2026, so treat the training exclusion the way you would treat any processor claim you have not read yourself: ask for it in the data processing agreement, in writing, naming your tier. Any supplier that will not put it there has answered the question.
Inside somebody’s personal account
The third option is the one already running. A paraplanner pastes a client’s circumstances into a free account at 7pm because it is quicker than the alternative nobody provided. Neither vendor caused that, and no contract fixes it. An approved-tools policy naming a real tool fixes most of it, and the technical backstop is described in Microsoft’s documentation: endpoint data loss prevention policies that warn or block users sharing sensitive information with third-party generative AI sites in a browser.
What your compliance officer can evidence
With Copilot, your compliance officer can evidence most AI use from tooling you already own. The FCA’s position is that AI isn’t separately regulated and existing obligations apply from the moment a firm uses it [3]. So what counts is evidence: having a policy matters less than being able to show what happened.
- Copilot interactions are discoverable by default. Microsoft documents that prompts and responses are captured in the unified audit log, stored in the user’s mailbox, retrievable through eDiscovery content search with a Copilot activity condition, and subject to Purview retention policies. Your SM&CR-accountable individual can answer “what has the firm been doing with this” from tooling you already own.
- ChatGPT Enterprise is not outside that net. Microsoft’s Purview documentation groups it as an enterprise AI app alongside Microsoft Foundry and Anthropic Claude Enterprise, with supported data security and compliance capabilities. A firm with Purview can govern both tools from one place, which is a stronger argument for running both than any feature comparison.
- Consumer AI use is visible, but only through a different lens. Purview categorises plain ChatGPT, Gemini and consumer Copilot as other AI apps detected through browser activity in the Defender for Cloud Apps catalogue. Discovery, not control.
- Communication compliance and insider risk extend to prompts. Purview includes a risky AI usage policy template, and communication compliance policies can be configured to cover generative AI interactions.
The licensing footnote on all of this matters more than the capability. Several of the controls above sit in Microsoft 365 E5 or in Purview add-ons rather than in a Business Premium subscription. Before you tell a board that AI use is fully auditable, confirm which of these your tenant is actually licensed for, rather than which of them a feature page lists.
Licensing, and the licences nobody uses
Most firms asking this question already pay Microsoft for Exchange, SharePoint, Teams and Office. Copilot is an add-on to that, which makes the marginal procurement decision small and the marginal governance decision almost nil, because the processing sits under terms your firm signed years ago. ChatGPT is a new supplier, a new contract, a new DPA, and a new entry in your outsourcing and third-party register. That difference is worth real money in legal and compliance time, and it rarely appears in the comparison a vendor shows you.
Then there is the failure that loses the money. A firm buys Copilot for everyone because the per-seat price looks modest against the headcount, and four months later a handful of people use it weekly. Nobody cancels, because cancelling means admitting the decision. I think this is the most common way AI spend gets wasted in a professional firm, and the software is never the cause.
Buy for the roles where you can name the task. Expand on evidence.
Copilot or ChatGPT: which to choose
- Choose Copilot if your firm runs on Microsoft 365, the work you want help with is email, documents, spreadsheets and meetings, and your compliance function needs the audit trail to sit inside existing tooling. This is most UK advice firms, most accountancy practices, most small law firms.
- Choose ChatGPT if your work happens outside the Office apps, if you want the strongest standalone assistant for research, analysis and long-form drafting, or if your firm is not a Microsoft shop and Copilot’s proximity argument therefore does not apply to you.
- Use both if you have a small group doing different work from everyone else, and you can govern the second tool properly: on the approved register, on a business tier, with a signed DPA and, ideally, inside the same Purview visibility as the first.
- Use neither yet if nobody in the firm can name three recurring tasks either tool would take over. Call it sequencing rather than caution. Work out the processes first and the tool choice usually answers itself, which is the whole argument of the providers overview.
A one-month pilot that settles it
A month-long pilot with six people answers the question better than any procurement exercise. It costs a month and one person’s attention.
- Week zero, before any licence. Pick six people who are heavy users of email and documents, not the six keenest. Have each list the ten things they do most often in a week. Review sharing permissions on your most sensitive SharePoint libraries, because Copilot will make any drift visible.
- Weeks one and two, Copilot. Licence the six. Ask them to use it for the listed tasks only, and to mark each one faster, unchanged, or newly possible. Resist training them beyond fifteen minutes, because you are testing whether it survives normal working conditions.
- Week three, ChatGPT. Same six people, same task list, business tier. The difference in what they reach for when both are available tells you more than any feature table.
- Week four, the evidence. Pull the audit trail for the Copilot interactions. If you cannot produce it, you have learned something more valuable than the productivity numbers. Then count the tasks that genuinely moved and licence those roles, not the firm.
The result is usually untidy, and that is the point. Two or three roles get real value, several get a little, and some get none. A decision made on that evidence survives a renewal conversation. A decision made on a demo does not.
This page is for informational purposes only and does not constitute regulated financial advice, legal advice, or a compliance opinion. Consult a qualified compliance professional for advice specific to your firm.
Sources
[1] Information Commissioner’s Office, ‘Data protection impact assessments’. Available at: https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/accountability-and-governance/data-protection-impact-assessments-dpias/
[2] Information Commissioner’s Office, ‘Guidance on AI and data protection’. Available at: https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/
[3] Financial Conduct Authority, ‘AI and the FCA: our approach’. Available at: https://www.fca.org.uk/firms/innovation/ai-approach