---
title: "Copilot or custom AI: 2026 comparison"
description: "What Copilot actually reaches in October 2026 (connectors, agents, Copilot Studio), what stays out of its reach, what the licence costs and who answers for the result. One comparison table and five questions to decide."
url: "https://gilabs.fr/en/guides/copilot-ou-ia-sur-mesure/"
lang: "en"
author: "Yann Gilliot"
published: "2026-10-01"
tags: ["Microsoft Copilot", "Copilot Studio", "AI Agents", "Microsoft 365", "SMEs"]
---
# Copilot or custom AI: a comparison for deciding, one process at a time

By [Yann Gilliot](https://gilabs.fr/en/about/) · published on 2026-10-01

What Copilot actually reaches in October 2026 (connectors, agents, Copilot Studio), what stays out of its reach, what the licence costs and who answers for the result. One comparison table and five questions to decide.

## The short version

- In October 2026, Copilot is no longer limited to emails and Office files: synced or federated (MCP) connectors give it access to line-of-business applications. Someone still has to build and secure that connection.
- A declarative Copilot agent answers when a user asks, inside Microsoft 365. It triggers nothing on its own. Work that runs with nobody at the screen calls for a custom engine agent or for development inside the application.
- In France, Copilot Business is listed at €18.20 excl. VAT per user per month (€15.60 on promotion as of 1 October 2026), on top of the Microsoft 365 licence. Custom work is paid once, as a project, then in running costs.
- A declarative agent inherits Microsoft 365 compliance. A custom agent puts compliance on whoever builds it. The Anthropic models offered in Copilot are excluded from the EU Data Boundary.
- The right decision is made one process at a time, and the answer is often both: the same MCP server, placed in front of your database, serves Copilot and your own agents.

"We already have Copilot, why build anything custom?" In June I answered that objection with an article that came down to one sentence: Copilot reads your emails and Office files, it does not reach into your information system. This guide replaces that article, because that sentence is no longer accurate.

In October 2026, Copilot can query a line-of-business application. Microsoft has opened up connectors, agents and a studio to build them. So the question is no longer "can Copilot see my data?". It becomes: who builds the connection, who triggers the work, and who answers for the result. That is what this guide compares, with facts checked against Microsoft's documentation on the date shown at the top of the page.

## What "Copilot" means in October 2026

The word covers four different things, and most misunderstandings start there. Microsoft is renaming Microsoft 365 Copilot to "Microsoft Copilot"; licences and part of the documentation still carry the old name.

| Level | What it is | Who sets it up |
|---|---|---|
| Copilot Chat | An AI chat included in the Microsoft 365 subscription at no extra charge. Access to work data is optional and billed on usage. | Nobody: it is already there |
| Copilot (licence) | The chat grounded in your Microsoft 365 data, plus the assistant in Word, Excel, PowerPoint, Outlook and Teams. | The Microsoft 365 administrator |
| Extended Copilot | Declarative agents, connectors to your other systems, agents built in Copilot Studio. | A capable administrator or a developer |
| Custom AI | A custom engine agent, or an AI feature built into your application, with your orchestration and the model you choose. | A developer |

The first two rows are bought. The last two are built. The "we already have Copilot" objection is almost always about the second row, while the need often sits on the third or the fourth.

## What Copilot reaches today

By default, Copilot answers from Microsoft Graph: emails, documents, calendar, chats, meetings, contacts. It only shows a user what that user is already allowed to see. That is a real strength, and also a warning: if your SharePoint permissions are untidy, Copilot makes the mess visible.

To go beyond Microsoft 365, there are two connector models.

| | Synced connector | Federated connector (MCP) |
|---|---|---|
| Principle | External content is copied and indexed into Microsoft Graph | No copy: the source is queried live, when the question is asked |
| Good fit | Knowledge bases, document stores | Fast-changing data, regulated content that must stay in the source |
| Permissions | An access list to set on each indexed item, otherwise visible to the whole organisation | The user's own permissions in the source system |
| Licence | Indexing at no extra cost; use in Copilot requires the licence | A Copilot licence for every user who queries the source |
| Writing to the source | No | Create, update, delete: rollout announced from early October 2026, with user approval |

More than a hundred connectors exist off the shelf (Salesforce, ServiceNow, Confluence, HubSpot, Notion and others). For an in-house application or a SQL Server database there is none, by definition. Microsoft covers that case with the custom federated connector. Its documentation opens with the prerequisite that matters: an MCP server, reachable at a URL, exposing read-only tools, authenticated through Microsoft Entra or OAuth 2.0.

In other words, connecting Copilot to your line-of-business application starts with custom development. Someone has to write that server, decide what it exposes, and enforce the application's permissions in it. The licence does not do that.

One limit worth knowing on synced connectors: semantic indexing serves topic and keyword searches. Microsoft states that it does not help with multi-parameter queries (a topic and an assignee, for example) or with requests for counts. "How many overdue files for this client?" is a database question, not a document search question.

## The comparison

| Criterion | Copilot (licence only) | Extended Copilot (declarative agent, connectors, Copilot Studio) | Custom AI (custom engine or inside the application) |
|---|---|---|---|
| Data reached | Microsoft 365 | Microsoft 365 and the systems wired in through a connector | Everything the application can access |
| Who triggers | The user, with a question | The user, with a question | A user, an event or a schedule |
| Work with nobody at the screen | No | No for a declarative agent | Yes |
| Where it is used | Microsoft 365 apps | Microsoft 365 apps | Microsoft 365, your application, or both |
| Orchestration and model | Microsoft's | Microsoft's | Yours: business rules, steps, chosen model |
| Compliance | Microsoft 365 commitments | Inherited from Microsoft 365 for a declarative agent | Yours to ensure: security, responsible AI, hosting |
| Cost model | Per user, per month | Licence, or Copilot Credits on consumption | Fixed-fee project, then hosting and maintenance |
| Time to set up | Days | Days to weeks | Weeks to months |
| What you own at the end | A subscription | A configuration in your tenant | The code, the rules, the tests |

Two rows of this table settle most cases.

**Who triggers.** Microsoft says so in its own comparison: a declarative agent does not support proactive interactions, it waits for a user to ask. Yet a process that costs money is rarely a question typed into a chat. It is a document arriving, a file waiting, data re-keyed from one system into another. That work starts on an event, and it belongs to the right-hand column. I explained why a well-adopted licence can change nothing in the accounts in [Copilot rolled out, nothing moved](https://gilabs.fr/en/blog/copilot-deploye-rien-na-change/).

**Who validates.** Microsoft describes what Copilot produces as useful drafts and summaries that the user reviews, rather than full automation of the task. That is the right contract for office work. Writing into a business system needs something else: control rules, a confidence threshold, a human validation queue, a record of who approved what. Those things are designed; they are not ticked in a console.

## What the licence costs, and what it does not cover

Microsoft's public prices for France, checked on 1 October 2026:

| Item | Price | Conditions |
|---|---|---|
| Copilot Chat | Included | Eligible Microsoft 365 subscription |
| Microsoft 365 Copilot Business | €18.20 excl. VAT per user per month (€15.60 on promotion on the date checked) | Annual payment, up to 300 users, on top of an eligible Microsoft 365 licence |
| Microsoft 365 Copilot (enterprise) | €26.00 excl. VAT per user per month | Annual payment, on top of an eligible Microsoft 365 licence |
| Copilot Studio | On consumption, in Copilot Credits | Pre-purchase plan (up to 20% off) or pay-as-you-go; an Azure subscription is required |

For thirty people, Copilot Business comes to €546 excl. VAT a month at the regular price, or €6,552 a year. That is a useful order of magnitude, on two conditions.

First, count what the licence does not include. Putting SharePoint permissions in order before opening Copilot, training the teams, writing the MCP server if you want to reach an in-house application, hosting a custom engine agent (Microsoft lists Azure AI Foundry, App Service and Bot Service as possible cost items).

Second, compare like with like. A licence is paid per person, every month, whether the process has moved or not. Custom development is paid once, then in running costs, and its return is calculated on one specific process: hours of re-keying, a lead time, an error rate. Without measuring the process, neither expense can be judged.

## Compliance: who answers for what

Three facts, taken from Microsoft's documentation.

Prompts, responses and data accessed through Microsoft Graph are not used to train foundation models. For European Union customers, Copilot is an EU Data Boundary service.

Microsoft offers third-party models in Copilot, including Anthropic's, which the administrator can enable or not. Models provided by Anthropic are currently excluded from the EU Data Boundary. For a profession bound by professional secrecy, that is a setting to check before any rollout, not after.

Finally, Microsoft's comparison table is plain about responsibility: a declarative agent inherits Microsoft 365 compliance and security, while a custom engine agent requires whoever builds it to ensure their own. Custom work gives you control (choice of model, hosting in Europe or on premises, a log of every decision) and the burden that comes with it. The two are decided together.

## Five questions to decide

Ask them about one process, not about the whole company.

1. **Where does the data live?** In Microsoft 365: Copilot is probably enough. In a line-of-business application or a SQL Server database: a connector will be needed, hence development, whatever comes next.
2. **Who triggers the work?** A person asking a question: Copilot or a declarative agent. An event, an incoming file, a deadline: custom.
3. **What does a mistake cost?** A draft reread by its author costs nothing. A wrong entry in the accounts, the payroll or a client file calls for controls, human validation and a trail.
4. **Where do users work?** In Teams and Outlook: Copilot is in the right place. In the business application all day: the AI belongs in the application, not in a window beside it.
5. **How many people are involved?** An office-work need shared by everyone justifies a licence per head. A process run by three people that weighs hundreds of hours a year justifies a project.

If all five answers point to Microsoft 365, a question and a draft, the right answer is Copilot, and I tell clients so when that is the case.

## In practice: both

The opposition in the title is convenient for deciding; it is wrong at the finish. In a Microsoft environment, the architecture that holds combines the two.

The meeting point is the MCP server placed in front of the application or the database. Written once, with the application's permissions enforced on the server side, it serves two uses. Copilot connects to it through a custom federated connector, for the questions teams ask in Teams. Your own agents connect to it as well, for work that runs on events, with their rules and their human validation.

The principle is the same as for RAG: [permissions are enforced in the query](https://gilabs.fr/en/guides/rag-sql-server-2025/), on the user's identity, never in a parameter the model could supply. And on the .NET side, the agent layer is built today with [Microsoft Agent Framework](https://gilabs.fr/en/guides/semantic-kernel-agent-framework/), which calls the tools of an MCP server through the protocol's official C# SDK.

## Going further

Answering "Copilot or custom" means starting from the process, not the licence: where the data lives, who triggers, what a mistake costs. That is what the [30-day AI diagnostic](https://gilabs.fr/en/services/ai-diagnostic/) is for: processes mapped, the return costed for each priority case, and for each one the answer to this question, including when the answer is the tool you already pay for. When the answer is custom work, the next step is [AI agents and business applications](https://gilabs.fr/en/services/ai-agents-automation/) or [.NET AI integration](https://gilabs.fr/en/services/dotnet-ai-integration/).

## Sources and references

- [Data, Privacy, and Security for Microsoft Copilot (page revised 9 July 2026)](https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy), Microsoft Learn
- [Microsoft 365 Copilot connectors overview (synced and federated connectors)](https://learn.microsoft.com/en-us/microsoft-365-copilot/extensibility/overview-copilot-connector), Microsoft Learn
- [Federated connectors overview (MCP, write actions from October 2026)](https://learn.microsoft.com/en-us/microsoft-365/copilot/connectors/federated-connectors-overview), Microsoft Learn
- [Set up custom federated connectors (the organisation's own MCP server)](https://learn.microsoft.com/en-us/microsoft-365/copilot/connectors/set-up-custom-federated-connectors), Microsoft Learn
- [Prerequisites for deploying connectors (licensing table)](https://learn.microsoft.com/en-us/microsoft-365/copilot/connectors/prerequisites), Microsoft Learn
- [Compare declarative and custom engine agents](https://learn.microsoft.com/en-us/microsoft-365-copilot/extensibility/agents-overview), Microsoft Learn
- [Plan Licensing and Cost for Microsoft 365 Copilot Extensibility](https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/cost-considerations), Microsoft Learn
- [Copilot Studio licensing (Copilot Credits)](https://learn.microsoft.com/en-us/microsoft-copilot-studio/billing-licensing), Microsoft Learn
- [Microsoft 365 Copilot pricing, France (checked 1 October 2026)](https://www.microsoft.com/fr-fr/microsoft-365-copilot/pricing), Microsoft

Topics: Microsoft Copilot, Copilot Studio, AI Agents, Microsoft 365, SMEs

## Take it further

Do you know, process by process, what belongs to Copilot and what needs custom work?

Before you spend a euro on AI, you know what it will return. In 30 days: your processes mapped, the real end-to-end delay of each one, a return in hard euros per priority case and a payback measured in months. My method has three steps, map it, prioritise it, then automate it, in that order: automating a cluttered process does not speed it up, it multiplies it. That is why part of the gain I quantify for you cannot be bought, least of all from me. Quantified by the person who will build it.

[30-day AI diagnostic](https://gilabs.fr/en/services/ai-diagnostic/)

**Book a 30-minute call**: [Calendly](https://calendly.com/yann-gilabs/30min)