How to Build a Custom AI Assistant on Your Company's Data: Platforms Compared

If you want an assistant that answers from your own documents, tickets and systems rather than the open web, there are two routes: switch on the company-knowledge features of an AI platform your team already pays for, or have an assistant built around your systems. This guide compares the main platforms as they stood on 27 September 2026, using each vendor's own documentation, and says plainly when a custom build is the better call and when it is not.

Two naming changes to know before you search: Microsoft has renamed Microsoft 365 Copilot to Microsoft Copilot, and Amazon Q Business is no longer open to new customers. AWS now points new customers to Amazon Quick.

The short answer

Start from where your company's information already lives. The platform that already holds your files and email is usually the fastest one to try.

• Your company runs on Microsoft 365: look at Microsoft Copilot Studio first.
• You run on Google Workspace: look at Gemini Enterprise.
• Your team already works in ChatGPT or Claude: their business plans can search the apps you connect (ChatGPT's Company Knowledge, Claude's enterprise search).
• Your data is in AWS: look at Amazon Quick.
• Your data sits in a database or system none of these reach, the assistant has to face customers, or it must follow your own approval steps: that is when a custom build earns its cost.

What "works with your company's data" actually means

Every vendor says its assistant uses your data. Three questions separate the real capability from the marketing line.

First, how does it reach the data? Most platforms use connectors: ready-made links to apps such as SharePoint, Google Drive or Slack. For anything without a ready-made connector, several now support MCP, the Model Context Protocol, which AWS's documentation describes as an open standard that defines how AI applications communicate with external tools and data sources. In practice, MCP is how an assistant reaches an in-house system: someone runs an MCP server in front of it.

Second, does it respect the permissions people already have? A good assistant only answers from what the person asking is allowed to open. ChatGPT, Claude, Copilot Studio and Amazon Quick all describe permission controls in their documentation; the details differ, and the table shows them.

Third, does it show its sources? An answer you cannot trace back to a document is hard to trust with company information. ChatGPT, Claude and Copilot Studio describe cited answers in the pages we checked.

The platforms side by side

What each vendor's documentation says about reaching company data, checked on 27 September 2026.

Summarised from each vendor's documentation on 27 September 2026 - see Sources at the end. Features, plan names and prices change often; confirm on the vendor's site before you decide.

Is your data used to train their models?

This is the first question most businesses ask. What the vendors' own pages said when we checked:

• Anthropic: by default it will not use inputs or outputs from its commercial products, such as Claude for Work and the Anthropic API, to train its models.
• Google: for the Gemini Enterprise app, prompts and outputs are not used to train Google models or models for any other customer.
• Microsoft: for Microsoft Copilot, prompts, responses and data accessed through Microsoft Graph are not used to train foundation models.
• OpenAI: publishes its commitments for business plans on its own business-data page. We could not load that page for this check, so read it directly before you connect company data.
• Amazon: we did not find a statement on training in the Amazon Quick documentation we read. Ask AWS before you connect company data.

When a custom build is the better choice

The platforms above are the quickest way to try an assistant on data that already lives in their ecosystems. A custom build is worth the extra cost when one of these is true:

• Your data sits somewhere none of them connect to - an in-house database, an ERP, an old line-of-business system - and nobody will maintain an MCP server for it.
• The assistant has to face customers, not only staff. Copilot Studio can publish agents to web and messaging channels; the ChatGPT, Claude and Gemini features above are for people inside your own workspace.
• It must follow your process rather than chat freely: draft, route to the right person, wait for approval, then write back to your system.
• Only a few people need it now and then, so a subscription for every user costs more than the work saves.
• You want to own the code, keep prompts and logic in your own repository, and change the underlying model later without rebuilding.

And when not to build: if your information is already in Microsoft 365 or Google Workspace and the job is "find and summarise what we already have", switch on the platform's own feature first. It is faster to try, and a pilot there tells you exactly what a custom build would need to do better.

What drives the cost

• Seats: the Gemini Enterprise app is billed per seat, per month; check ChatGPT's and Claude's business-plan pricing pages for theirs.
• Usage: Copilot Studio is sold as prepaid Copilot Credit packs or pay-as-you-go, so cost follows how much the agents are used.
• Subscription tier: in Amazon Quick, setting up integrations needs the Enterprise subscription; Professional users can use integrations shared with them.
• Custom build: a one-off build scoped per role, then model usage, hosting, and keeping the integrations working as the connected systems change.

Questions to ask any vendor before connecting company data

• Does it only answer from what the person asking is already allowed to see?
• Does every answer cite the document it came from?
• Is our data used to train models - by default, and can we switch that off?
• Where is our data processed and stored, and can we choose the region?
• Can administrators decide which apps and sources are connected, and remove access when someone leaves?
• What happens to our connectors and prompts if we move to another platform?

Sources

Checked against each vendor's own pages on 27 September 2026. Products and plans change often; confirm on the vendor's site before you buy.

• OpenAI Help Center - Company knowledge in ChatGPT
• Claude Help Center - Use enterprise search
• Claude Help Center - Use connectors to extend Claude's capabilities
• Anthropic Privacy Center - Is my data used for model training?
• Microsoft - Microsoft Copilot Studio
• Microsoft Learn - Knowledge sources summary (Copilot Studio)
• Microsoft Learn - Data, privacy and security for Microsoft Copilot
• Google Cloud - Gemini Enterprise app FAQs
• AWS - Amazon Q Business availability change (and Amazon Quick)