One AI Platform for the Whole Company: Single Vendor, Multi-Model Workspace, Gateway, Cloud Catalogue or Build Your Own

Many companies reach the same point. Staff already use ChatGPT or Microsoft Copilot, and now someone asks for a different model, a shared assistant for one team, or a record of who asked what. A model is the trained program that writes the answers, and a large language model (LLM) is a model that works with text. You can stay with one vendor's product, buy a workspace with several models, or put a gateway in front of several providers. A gateway is a control point between your software and the models. You can also rent models from a cloud catalogue or build your own. Most mid-sized companies should start with the product their staff already use and add pieces only when a named need appears.

Why would a company want more than one AI model, and when is one enough?

Models differ in what they do well and what they cost, and vendors change their line-ups often. Even a single product mixes models. Microsoft's page on Copilot says “By default, Copilot uses a real-time router to adjust the underlying model it uses based on your prompt.”

Three reasons justify choosing models yourself. A team may need a model your product lacks. A routine task may suit a smaller, cheaper one. Or you want a second provider ready for an outage. One vendor is enough when staff use AI for drafting and search, nobody builds AI into your software, and the admin controls meet your rules.

What do ChatGPT Enterprise and Microsoft 365 Copilot give us, and where do they stop?

For a ChatGPT Enterprise workspace, ask OpenAI in writing for the current admin controls, data terms and model list. OpenAI's API documentation says “data sent to the OpenAI API is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us)”, but that page covers the API, not the ChatGPT apps.

A custom GPT is an assistant that an employee sets up inside ChatGPT with written instructions and attached documents. OpenAI describes it as a way to customise ChatGPT “by providing instructions, attaching documents as knowledge”.

Microsoft says “Microsoft offers third-party AI models in Microsoft Copilot, such as Anthropic and OpenAI models.” It adds: “As an admin, you can decide whether to use these models to power your experiences, and additional terms may apply.”

The limits show up when you build assistants. Microsoft says declarative agents “use Copilot's AI infrastructure, model, and orchestrator”, so you set instructions, knowledge and actions but keep Microsoft's model choices. A custom engine agent, where you “bring your own orchestrator and models”, may “Require additional hosting outside of Microsoft 365”. Microsoft says setup “Varies from simple setups in Copilot Studio to advanced pro-code implementations”, and the pro-code end of that path is already a build.

What is a multi-model workspace, and what do we give up?

A multi-model workspace is a chat application for staff that offers models from several providers behind one login. Two examples that you run on your own servers: the Open WebUI documentation mentions “support for Ollama and OpenAI-compatible APIs”, and LibreChat calls itself “a self-hosted web application”. Hosted products of this kind exist too.

What you give up is depth. Microsoft says Copilot “only surfaces organizational data to which individual users have at least view permissions”, because it reads your Microsoft 365 permissions directly. A separate workspace needs its own links to your files and systems, and you must map who may see which file. Self-hosting needs someone to run it, and a hosted product adds another company that handles your prompts.

What does an LLM gateway do, and who needs one?

A gateway is a service between your software and the model providers. Your software sends every request to one address, and the gateway forwards it to the chosen model. A token is the small unit of text that a model reads and writes and that providers count and bill. Anthropic's glossary says tokens “can correspond to words, subwords, characters, or even bytes”.

The open-source LiteLLM project describes its gateway as “a self-hosted server” offering “one OpenAI-compatible endpoint for 100+ LLM providers, MCP tools, and A2A agents”. It says to “Give each user, team, and project a virtual key”, each with “models, budgets, and rate limits”. The real provider keys then stay in one place.

Others cover similar ground. Portkey says it has “open sourced” its gateway. Cloudflare AI Gateway, a hosted service, says you can “View metrics such as the number of requests, tokens, and the cost it takes to run your application.” Kong says its AI Gateway can “Automatically fail over to a different provider or model when one is slow or unavailable.”

You need a gateway when several teams or products call models from code and you want one place for keys, budgets, routing and logs. Staff who only use chat apps do not need one. A gateway gives them no chat window, and its logs can hold prompts (the text staff type in), depending on its settings, so protect them.

What does a cloud provider's model catalogue add?

Three examples are Microsoft Foundry (formerly Azure AI Foundry), Amazon Bedrock and Google's Gemini Enterprise Agent Platform (formerly Vertex AI). They let your engineers call many providers' models through the cloud account you already hold. Microsoft says the Foundry catalogue “includes over 10,000 models”, and AWS says Bedrock supports “100+ foundation models”. Google says Model Garden holds models “from Google and Google partners”.

The gains are your existing security setup, one invoice and a choice of processing location. Microsoft says “you can choose global, data zone, or regional processing based on your compliance requirements.” That applies to serverless deployments. For Claude models hosted on Anthropic's own infrastructure, Microsoft says data “might be processed outside of Azure including outside of your selected Azure geography”. The cost is that this is a toolbox for builders. Any chat screen in its console is for testing by builders, so staff get no ready-made chat product, and you still own testing and model choice. Microsoft also says older models “might be retired in the Foundry model catalog”.

How do we move our custom GPTs, prompts and knowledge files, and what stays behind?

Check OpenAI's timetable first. Press reports from autumn 2026 say OpenAI plans to retire custom GPTs in December 2026 and move them to plugins, with later dates for some Enterprise workspaces. Follow the notice for your own workspace. List every assistant with its owner and use. Rebuild the busiest first and test each copy with the same questions. Three things move well: the instruction text, the source documents if you still hold the originals, and your test questions. Four things stay behind. Conversation history stays on the old platform, and how much you can export depends on its admin tools. Sharing settings must be set again. Actions, the calls from an assistant to your other systems, must be set up again on each platform. An OpenAPI description of the API can be reused where the platform accepts one. And answers will change, because each platform splits and searches files in its own way and the model underneath differs.

Who is allowed to see what, and how does each route handle our data?

Ask every route whether your data trains the provider's models and where it is processed. The vendors say:

• OpenAI API: “data sent to the OpenAI API is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us)”.
• Microsoft 365 Copilot: prompts, responses and data read through Microsoft Graph (the Microsoft 365 data service) are not used to train its foundation models. “EU traffic stays within the EU Data Boundary while worldwide traffic can be sent to the EU and other countries or regions for LLM processing”, and “Models provided by Anthropic as a subprocessor are currently excluded from the EU Data Boundary.”
• Amazon Bedrock: “AWS and the third-party model providers will not use any inputs to or outputs from Amazon Bedrock to train Amazon Nova, Amazon Titan, or any third-party models.” AWS's general privacy FAQ adds “You choose the AWS Region(s) in which your content is stored.” With a global cross-Region inference profile, Bedrock can process a request in “Any supported AWS commercial Region worldwide”.
• Gemini Enterprise Agent Platform (formerly Vertex AI): Google states that it does not train models on your data without your prior permission or instruction. It says data at rest stays in the location you select, but requests to a global endpoint “may be processed in any Google Cloud location around the world”.
• A workspace or gateway you host: data stays in your environment until it is forwarded, and then the model provider's terms apply. A hosted gateway adds a company.

For personal data, the GDPR uses the word processor for a company that handles data on your behalf. In plain terms, Article 28 requires a written contract and action only on your documented instructions, unless the law requires otherwise. Another processor may be added only with your prior written authorisation. That authorisation can be general, and then the processor must tell you about changes so that you can object. Microsoft's Foundry page says the Azure Machine Learning service “Acts as the data processor for prompts submitted and content output for serverless deployments.” For Claude models in Foundry, Microsoft says Anthropic “acts as an independent data processor”, so check who the processor is for each model. Have your own counsel read each contract.

Inside the company, an assistant should show a person only what that person could already open. On any route, including Copilot, test this with one login on a folder that person must not see.

What does governance need in practice?

Roles come first: who owns each assistant, who approves a new one and who can switch a model. NIST's AI Risk Management Framework says “Roles and responsibilities and lines of communication related to mapping, measuring, and managing AI risks are documented and are clear to individuals and teams throughout the organization.”

Approvals come second. Sending an email, changing a record or paying an invoice should wait for a person, as the article on human approval, linked above, explains. Logs come third: who asked, which model answered, which version of the assistant ran, and what it cost. A gateway can write most of that record for every model you call, if your software sends the person and the assistant version with each request.

The fourth item is a register of use cases. NIST's playbook says “Mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities.” A shared spreadsheet is enough to begin. Give each row a use case, an owner, the model, the data it touches and a review date.

How does per-seat pricing differ from per-token pricing?

Per-seat pricing charges a fixed amount for each named user. Per-token pricing charges for the text your software sends and receives. Ask each vendor for the current seat price in writing, with every minimum and add-on.

For tokens, Anthropic publishes a price list in US dollars. It shows Claude Sonnet 5.5 at 2 dollars per million input tokens and 10 per million output tokens. Claude Haiku 5.5 costs 0.10 and 0.50 dollars for prompts up to 100,000 tokens.

Take an invented workload of 1,000 requests a day, each sending 3,000 tokens and receiving 500, which is 3 million input and 0.5 million output tokens. On Sonnet 5.5 the cost is 6 dollars for input plus 5 for output, 11 dollars a day. On Haiku 5.5 it is 0.30 plus 0.25, 55 cents a day. The workload is an assumption, and cloud resellers may charge differently.

Output tokens cost more than input tokens, model choice moves the bill by a large factor, and long attached documents raise the input side. Set a gateway budget from the first day.

What does lock-in look like, and what should we keep so we can leave?

Lock-in sits where the effort went, in assistants built in one vendor's editor, files stored and prepared for search in its system, actions written in its format, and history only it holds. A gateway can become lock-in too if your routing rules live in a format only it reads.

Keep every assistant's instruction text in a repository, with the source documents, your test questions and an export of logs. Call models through an OpenAI-compatible endpoint (a request format that many tools copy), and build actions on open formats such as an OpenAPI description (a standard description of an API) or the Model Context Protocol, which describes itself as “an open-source standard for connecting AI applications to external systems”. Agree export, deletion and model-retirement notice terms in the contract.

When is building our own platform justified?

Building is justified when AI is part of the product you sell, because your clients are then the users and no staff licence covers them. It is also justified when one process needs logic no product offers, such as routing by client contract. Rules on where data may be processed can also exclude every hosted option.

It is hard to justify for staff chat alone. The products above already do that, and you would take on updates, access control, logs and on-call duty.

What should we ask a vendor?

• Which models can we choose, and who decides when one is replaced or retired?
• Is our data used for training on any plan, and where does the contract say so?
• In which region is data processed, and does that change with the model we pick?
• Is the vendor our GDPR processor, and which sub-processors does it use?
• Can an admin see and export logs of prompts and answers?
• What does one month of our real usage cost, with every minimum and add-on included?
• What can we take with us if we leave, and in what format?

What should we ask a firm that would build it?

• Which parts will you build, and which will you configure from existing products?
• Will it run in our cloud account, and who holds the keys?
• How will you test each assistant before launch, and who on our side signs off?
• What do we receive at the end: code, configuration, test sets, logs, operating guides?
• Who runs it after launch, and what does that cost each month?

Where does amBrain fit?

amBrain publishes this blog. On its own product, it says: “Our own product, shyash.ai, runs on AI agents that research, write, check and measure from one shared memory of company facts. A person approves everything that goes out.”

This article is not a case study and describes no client work. It does not claim that amBrain has built a multi-model platform or a gateway for any client. The line above is about amBrain's own product, and the examples are general.

The companies and projects named here are cited from their published pages, read on 10 November 2026. The article claims no relationship between amBrain and any of them and gives no amBrain prices or timelines.

amBrain has been building software since 2019. On how it works with clients, it says: “Three formats: full delivery, a dedicated team, or engineers embedded in your team.” Its line on ownership is “The client keeps full ownership of the product and the code, except our reusable components.” Ask for the list of those components by name before you sign.

If amBrain is on your shortlist, ask it the same questions as any other firm.