ChatGPT Enterprise and Microsoft Copilot Seats or a Document Processing Pipeline: What a Regulated Company Needs

ChatGPT Enterprise, ChatGPT Business and Microsoft Copilot give staff an assistant that can read a contract or a claim form and answer questions about it. These per-person subscriptions, often called seats, suit documents that people read and decide on one at a time. When the same kinds of documents arrive every day and their fields have to land in another system, you need a document processing pipeline. It reads every document with a large language model (LLM), the kind of AI behind ChatGPT. Then it checks each field against your rules, sends doubtful ones to a person and writes the rest into your systems. In a regulated company, volume and where your documents may be processed decide between the two.

What do ChatGPT Enterprise and Copilot seats give a regulated company?

Seats give each employee an assistant, priced per person and covered by business terms on how the vendor uses company data. OpenAI's pricing page, read on 16 October 2026, says its paid ChatGPT plans, Business and Enterprise included, “are priced per user per month”. Microsoft's page on licence options describes Microsoft Copilot, the product it used to call Microsoft 365 Copilot, as “an add-on plan” to a Microsoft 365 subscription, with a licence assigned to each user.

OpenAI's enterprise privacy page, updated on 8 January 2026, says “We do not train our models on your data by default”. It gives that commitment for ChatGPT Business, ChatGPT Enterprise and its API, among other products. The API is the interface your own software uses to call OpenAI's models. Microsoft's privacy page for Copilot, read on 16 October 2026, says “Prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLMs, including those used by Microsoft Copilot.” Microsoft Graph is the layer through which Copilot reads your company's mail, chats and files.

The documents still go to the vendor's cloud. For EU customers, Microsoft calls Copilot an EU Data Boundary service, one whose customer data it commits to store and process inside the EU and EFTA, and adds: “Models provided by Anthropic as a subprocessor are currently excluded from the EU Data Boundary.” Microsoft's page on Anthropic models says that customers in the EU Data Boundary and in the UK “have Anthropic models disabled by default”. If your rules say certain documents must stay on servers you control, seats are out for those documents. For an assistant that runs on your own servers, see the article on a private ChatGPT, linked above.

Where do seats stop being enough for document work?

The first limit at volume is people's time: each document needs someone to upload it, ask, read the answer and copy the result. OpenAI also caps the pace. Its help article on file uploads, read the same day, says: “You can upload up to 80 files every 3 hours.” It adds: “OpenAI may lower these limits during peak hours.”

Two colleagues who word the same question differently can get different values from the same form. One fixed written instruction per document type helps. Even then, someone has to measure how often each field comes out right, by scoring the answers against an acceptance set. That is a set of real documents with the correct value of every field written down by your staff.

In a chat, the result stays on the screen, and someone types it into the claims, ticket or onboarding system, with a chance of a new error each time. A pipeline checks each value against your rules, such as an amount within a limit, a date that exists or a client number that matches your records, and writes it in itself.

Seats come with records of conversations. OpenAI says ChatGPT Enterprise workspace admins “can access an audit log of conversations and GPTs through the Enterprise Compliance API”, and Microsoft stores “the user's prompt and Copilot's response”. A document process may need one record per document instead, showing the instructions and model version behind each field, the checks it passed and who approved it. Ask your compliance team which of the two your process needs.

Can Copilot agents or custom GPTs do the pipeline's job?

Partly. Custom GPTs and Copilot agents are assistants set up for specific tasks. Connected to your systems, they can take over some of the manual copying, and for some document types that may be enough.

OpenAI's privacy page says workspace members “can build and share GPTs internally”. The same page says apps “enable ChatGPT to send and retrieve information from connected internal sources and third-party applications”, and admins choose which apps are switched on. Microsoft lets Copilot use agents and connectors to other tools. Its privacy page says “Admins have full control to select which agents are allowed in their organization”. It also advises companies using an agent to “check the privacy statement and terms of use of the agent to determine how it will handle your organization's data.”

Test an agent as you would test a pipeline. Run it on your acceptance set, find out where each document is processed, including at the agent's own supplier, and look at the record it keeps for each document. Use it only for the document types where it passes.

What does a document pipeline add?

A pipeline runs the same steps on every document, and a person sees only the documents that fail the checks:

• Intake collects documents from email, an upload page or a scanner and sorts them by type
• A language model reads each document and returns the fields you ask for, with the place in the document each value came from
• Your rules check each value, and anything doubtful goes to a review queue, the list of documents waiting for a person
• Checked values are written into the system that uses them
• A record for each document keeps the model version, the instructions, the check results and the name of any reviewer

The model can be one your staff already use. OpenAI's privacy page separates the two uses: “The easiest way to get started with OpenAI is to deploy ChatGPT Business or ChatGPT Enterprise for your employees,” and “If your engineering teams wish to build custom solutions using our technology, try our API Platform”. A pipeline can call a provider's API under its business terms, or run an open-weight model on servers you control. An open-weight model is one whose files you can download and run yourself. The article on API, self-hosted and hybrid setups, linked above, compares those choices.

API terms differ from seat terms, so read them separately. The same OpenAI page says: “Except for certain endpoints and features listed in our platform documentation, OpenAI may securely retain API inputs and outputs for up to 30 days to provide the services and to identify abuse.” It adds that you can request zero data retention “for eligible endpoints if you have a qualifying use-case”.

How do we decide which documents go where?

List your document types and write down for each: how many arrive a month, how many fields you take from each document, which system those fields go to, and whether the documents may leave your perimeter (the servers and cloud accounts your company controls).

• Seats fit varied documents that a person reads anyway to decide something, such as a contract read before a meeting, when nothing from them goes into another system
• A pipeline fits types that repeat, such as claim forms, invoices or account opening files, when their fields feed another system and an error has a cost you can name, such as a wrong payment
• Documents that must stay inside your perimeter need a system on servers you control, a pipeline or a private assistant, whatever their volume

Seats and a pipeline can run side by side. Staff keep their seats for reading and drafting, and the steady flow of standard documents moves to the pipeline.

How do we compare the two on our own documents?

Score both on one acceptance set, made of recent documents of each type. The article on buying document AI software or building a pipeline, linked above, explains how to build the set and keep part of it for the final score.

• Give two or three staff members the seats and one written instruction per document type, and time them while they work through a sample of the set
• Score their answers field by field against the correct values
• Have the pipeline process the same sample where your rules allow documents to be processed, and score it the same way
• Compare accuracy for each field, staff minutes per document, and how many results reached the target system without retyping

If the seats score as well as the pipeline on a document type and the staff time is acceptable, keep that type with the seats.

What does each option cost as volume grows?

With per-person pricing, the seat bill follows headcount, plus any extra usage credits you buy. If volume doubles and you hire people to handle it, you buy more seats as well. A pipeline's bill follows the documents: model usage or servers, plus the staff who work the review queue and the people who run the system. Ask your seat vendor and each firm that offers a pipeline to price the same written volumes for three years, and count your own staff time on both sides.

Can we start with seats and add a pipeline later?

Yes, and the months with seats can prepare the pipeline. While staff work with them, collect for each document type the list of fields you take from it, the written instruction your staff settled on and the documents where staff had to correct the result. The field list and the instruction become the start of the pipeline's specification. The corrected documents, with their correct values, go into the acceptance set.

Keep this material in your company's own files rather than scattered across chat histories, so that any vendor or firm you talk to later can be tested against it.

What should we ask a firm that offers to build the pipeline?

• Which document pipeline have you taken to production, into daily use, and what happened to the documents it could not read?
• Will you score your prototype on our acceptance set, where our rules allow, before we sign for the full build?
• Which of our document types would you leave with staff and their seats, and why?
• What does the record for each document contain, and can our compliance team read it without your help?
• Who owns the code, the instructions given to the model and the acceptance set when the work ends?

Where does amBrain fit?

The one language model project amBrain describes in public is this: “We have taken an LLM integration to production inside a client's FinTech perimeter: extracting and normalising unstructured broker and venue notices — corporate actions, instrument and margin changes — into structured records the trading system consumes.”

This article is not a case study, and the client is not named. The quoted line does not say how that project started, which model it used, where the model ran or which of the steps above it included. Nothing here says that amBrain replaced seats with a pipeline, compared the two for a client, built systems for claim forms, invoices or account opening files, or has deployed ChatGPT Enterprise, ChatGPT Business, Microsoft Copilot or the OpenAI API for anyone. OpenAI and Microsoft are named as examples from their own published pages, and the article claims no relationship between them and amBrain. No amBrain prices, timelines or terms for a prototype or a test before signing are given.

amBrain has been building software since 2019. It works in 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.” If amBrain is on your shortlist, give it the same acceptance set and the same questions as any other firm.