Does Microsoft 365 Copilot train on our data?

Microsoft states that prompts, responses and data accessed through Microsoft Graph are not used to train foundation models, and that Copilot data stays within Microsoft's contractual commitments. Read the current statement on learn.microsoft.com — our point is who operates the endpoint, not that Microsoft's terms are bad.

Can we keep Copilot and still use the add-in?

Yes. They are separate endpoints. Copilot stays licensed per user; the add-in points at your gateway. Usage simply appears in two places — Copilot in the Microsoft 365 admin center, the add-in in your gateway ledger.

What does the add-in still depend on?

Microsoft Entra for sign-in and Microsoft Graph for the hosts you enable (Outlook needs mail and calendar permissions). Those dependencies do not disappear — self-hosting moves the model endpoint, not the Office platform.

Is this a drop-in replacement?

No. Copilot is embedded across the suite with tenant-wide grounding; the add-in is a task-pane experience in the four Office hosts, grounded in the current document, attachments and the MCP servers you enable. Expect a different product, not a skin.

GateLLM vs Microsoft 365 Copilot

Both bring AI into Excel, Word, PowerPoint and Outlook. The difference is where inference runs, who chooses the model, and whether the data path stays inside your network.

Only GateLLM does this

Structural differences, not degrees of quality. Copilot is strong on Microsoft integration — that appears in the section below, not here.

Capability	GateLLM (Office add-in)	Microsoft 365 Copilot
Add-in served from your own originThe manifest points at a web app you host; requests go to the gateway you operate. Copilot is delivered as a Microsoft service and has no self-hosted deployment option.	Yes	No
Model and provider choice, including self-hosted modelsRoute to any provider you configure, or to a local vLLM / Ollama upstream. Copilot runs on Microsoft-selected models; Copilot Studio can add Azure AI Foundry models, but inference never moves onto your own infrastructure.	Yes	No
One gateway for Office and developer AIThe add-in calls the same gateway your apps, agents and coding tools use, so usage, quotas and audit sit in one ledger. Copilot usage is reported separately in the Microsoft 365 admin center.	Yes	No

Both do these well

A settled baseline — either option covers this layer. Do not choose on these rows.

Capability	GateLLM (Office add-in)	Microsoft 365 Copilot
In-app AI in Excel, Word, PowerPoint and Outlook	Yes	Yes
Sign-in with Microsoft Entra ID	Yes	Yes
Respects Microsoft Purview sensitivity labels	Yes	Yes
Admin rollout from the Microsoft 365 admin center	Yes	Yes
Usage reporting for administrators	Yes	Yes

Different in kind

Not better or worse — different products solving an overlapping problem. Weigh these directly.

Capability	GateLLM (Office add-in)	Microsoft 365 Copilot
Where inference runs	Inside your network, on the gateway you operate	Within the Microsoft 365 service boundary
Model choice	Any provider you configure, including self-hosted models	Microsoft-selected models; Copilot Studio can add Azure AI Foundry models
Grounding	The current document, attachments and the MCP servers you enable	A tenant-wide Microsoft Graph semantic index, honouring user permissions
Data terms	GateLLM never receives prompts; your provider contracts govern downstream use	Microsoft states tenant data is not used to train foundation models and stays within its commitments
Pricing model	Per-instance gateway licence, with no per-seat fee	A per-user licence on top of an eligible Microsoft 365 plan
Remaining dependencies	Microsoft Entra and Graph for Office sign-in and mail / calendar	Microsoft 365 licensing, Graph, and Bing web grounding (which admins can disable)

Microsoft statements are summarised from learn.microsoft.com; GateLLM facts are verified against docs.gatellm.io. Both products change — check the sources before procurement.

Non-Gateway Alternative

Where the difference actually is

Copilot is a Microsoft service; the add-in is software you run. Three consequences follow — and none of them is "Copilot trains on your data".

The add-in is served from your origin

How GateLLM does it The manifest points at a web app you host, and that app calls the gateway you operate. There is no vendor relay and no vendor control plane in the path.

What Copilot does differently Copilot is delivered as part of the Microsoft 365 service. There is no self-hosted deployment option, so the service boundary is Microsoft's by design.

You choose the model and where it runs

How GateLLM does it Requests route to any provider you configure, including a local vLLM or Ollama upstream, and policies can pin a model per access key or key group.

What Copilot does differently Microsoft selects the models. Copilot Studio can attach Azure AI Foundry models, but inference remains inside Microsoft's infrastructure rather than yours.

One ledger for Office and developer AI

How GateLLM does it The add-in calls the same gateway as your applications, agents and coding tools, so token usage, quotas and audit records are unified and attributable to a team.

What Copilot does differently Copilot usage is reported in the Microsoft 365 admin center, separately from whatever governs your developer AI traffic.

Moving from Copilot to a self-hosted add-in

You are not replacing Microsoft 365 — only the AI endpoint behind it. Copilot licensing, Entra sign-in and Graph stay exactly where they are.

1Deploy the gateway in your VPC and connect the providers you use, or a local vLLM / Ollama upstream.
2Deploy the add-in web app and register it in your Entra tenant; the manifest points at your origin.
3Grant the Graph permissions the enabled hosts need — Outlook requires additional admin consent.
4Map Purview sensitivity labels to access policies so restricted documents cannot open the add-in.
5Roll out through the Microsoft 365 admin center and watch usage in the gateway's own ledger.

What stays, what changes

Microsoft 365 licensing, Entra and Graph remain Microsoft's. What changes is the inference endpoint: it moves to a gateway you operate, model choice becomes yours, and Office usage lands in the same ledger as the rest of your AI spend.

Cost structure

•GateLLM is licensed per gateway instance; there is no per-seat fee on top of it.
•Under BYOK you pay model providers directly, and the licence does not scale with token volume.
•Microsoft 365 Copilot is a per-user licence on top of an eligible Microsoft 365 plan.
•The honest comparison is per-seat AI licence versus per-instance gateway plus model spend — they are not the same shape of cost.

See how these numbers are computed →

When to choose which

Choose Microsoft 365 Copilot when

You want a managed product with no infrastructure to run, your data-residency requirements fit Microsoft's service boundary, and tenant-wide Graph grounding is the feature you actually need. It is the lower-effort path, and for many organisations the right one.

Choose a self-hosted Office add-in when

Inference has to stay on your own infrastructure, you already route developer AI through a gateway, or you need to choose models per team — including self-hosted ones. You take on deployment and upgrades in exchange for that control.

Both together is legitimate

Several organisations run Copilot for general staff and the add-in for teams whose documents cannot leave the network. The two are not mutually exclusive; only the AI endpoint differs.

FAQ

Microsoft statements summarised from learn.microsoft.com; GateLLM facts verified against docs.gatellm.io. Both products change — verify before procurement.

View pricing