Buy Document AI Software or Build a Custom LLM Pipeline: How a Regulated Company Decides and Tests Both

If your documents cannot leave your perimeter (the servers, networks and cloud accounts your company controls), you can buy document AI software that installs inside it, have an engineering firm build a custom pipeline around a language model there, or combine the two. Decide by scoring every candidate, bought or built, inside the perimeter on an acceptance set: your real documents, with the correct value of every field written down by your own staff. For the build route, shortlist engineering firms that can show a system they took to production inside a client's perimeter. Each firm should also agree to be scored on your documents before you sign.

What are we choosing between?

Document AI software, which vendors call intelligent document processing or IDP, is a licensed product. It sorts incoming documents by type, extracts fields such as dates and amounts, and sends unclear documents to a person. Some document AI software installs inside your perimeter, and the vendor keeps the code.

A custom pipeline is code an engineering firm writes for your document types and rules. One step is a large language model (LLM), a model trained on large amounts of text. It reads each document and returns a value for every field you ask for. Other steps check those values against your rules, send doubtful ones to your staff and pass the rest to your systems.

Which document AI software can run inside our perimeter?

Check this first, because some document AI software is sold as a service in the vendor's cloud, and every page you send it leaves your perimeter. Amazon Textract's FAQ, read on 15 October 2026, lists the AWS regions where Textract is available and describes no version you can install yourself.

The same FAQ says Textract accepts calls over a private connection from your own AWS network (AWS PrivateLink), so they “avoid using the public internet”, but Amazon's service still processes the pages. It also says that, unless you opt out, “some portion of content processed by Amazon Textract may be stored in another AWS region” to improve Amazon's machine-learning services.

Other vendors say on their own pages, read the same day, that their software can run inside your perimeter. Naming them here is not a recommendation, and the list is not complete.

Microsoft sells Azure Document Intelligence as a cloud service and, for some of its models, as containers, packaged software you run on your own machines. Its disconnected containers page says they “are intended for scenarios where no connectivity with the cloud is needed for the containers to run”. To use them, you “must first fill out and submit a request form and purchase a commitment plan”.

ABBYY's page for its Vantage platform says “Vantage is designed as a cloud-first platform”, then adds: “Alternatively, flexible deployment options are available for running Vantage on premises or within your private cloud instance on Microsoft Azure.” Hyperscience's Hypercell platform page lists on-premises deployment (on your own servers) and a “Customer Private Tenant (AWS, Google, Microsoft Azure)” among its deployment options, next to Hyperscience SaaS, a version the company hosts on AWS or Google Cloud.

Get the list of features that run inside your perimeter in writing, because a vendor's own pages can disagree. Microsoft's FAQ says “disconnected container usage is available for Document Intelligence custom and invoice models”, while its disconnected containers page lists Read, Layout, ID document, receipt and other containers.

Which of our documents suit bought software, and which need a custom pipeline?

Buying fits when most of your volume is in document types the software already reads, such as invoices or identity documents. Your rules for those types should also be close to other companies' rules. You can then test the software's ready-made models on your documents before paying for any build.

A custom pipeline suits document types specific to your company, such as internal forms, letters from counterparties or free-text support tickets. It also fits when the results must follow rules only your company has, or feed a system no vendor connects to. The pipeline holds those rules as code, while bought software lets you set them up only as far as the vendor's tools allow.

A combined setup splits the work by document type: the software reads the standard forms it knows, and a custom pipeline handles your own types, your rules and the hand-off to your systems. The test results decide which type goes where. The parts that join the two systems need an owner named in the contract. They include the step that routes each document, the review queue (the list of documents waiting for a person) and the audit record, the log of what happened to each document.

How do we test both on our own documents before signing?

Build the acceptance set before you meet vendors. Take recent documents in the mix you receive, scans and phone photos included, and have your own staff write down the correct value of every field. Include the types you expect the software to read and the ones you expect a custom pipeline to handle, so one set scores every route.

Run every test inside the perimeter: the software as a trial of the version you would license, and the firm's prototype under accounts you control.

A trial can still connect to the vendor. Microsoft's disconnected containers need an approved request and a commitment plan first, so a test before you commit may run as connected containers. Its FAQ says connected containers “don't send customer data, such as the image or text that's being analyzed, to Microsoft”, and that “internet connectivity is required to send billing information to Azure.” Your security team decides whether to allow that connection.

Before anyone sees the set, split it in two. The vendor and the engineering firm may use the first part to set up their systems, and the second part stays with you. Final scores come only from that second part, on four measures:

• Accuracy for each field within each document type, such as the date on a contract or the amount on an invoice
• The share of documents each candidate sends to a person at the same pass mark (the lowest accuracy you accept for fields that no person checks)
• Reviewer time, measured with a timer while your own staff check a sample on each candidate's review screen
• What happens to documents the candidate cannot read, such as a photo taken at an angle or an unfamiliar type: whether each one reaches a person, and how many wrong values pass through with no warning

The pass mark has to be agreed before the test starts, because a candidate can look more accurate by sending more documents to people. The article on what a COO should ask, linked above, covers timing reviewers and setting the pass mark.

What do we pay for on each route?

On the buy route you pay for a licence. Microsoft sells disconnected containers in yearly commitment tiers, “for a fixed fee” in the words of its disconnected containers page. There are separate tiers for prebuilt (ready-made) models, for custom extraction and for Read, its text recognition model.

Microsoft's pricing page, read on 15 October 2026, lists a prebuilt tier for up to 1.2 million pages a year, a projected 100,000 pages a month. Its public price list shows that tier, “Commitment Tier Pre-Built Disconnected 100K”, at $8,640 a year in the Central US region.

ABBYY's general subscription terms, last updated on 2 January 2026 and read on 15 October 2026, say a subscription term “shall automatically renew for an unlimited number of twelve (12) months renewal terms”. That happens unless your price quote says otherwise or either party gives written notice at least 45 days before the current term ends. For volume above what you bought, the terms say that, unless your price quote says otherwise, ABBYY “will charge Licensee for the subsequent amounts due for amounts in excess of the Volume Units purchased at the end of the Subscription Term”.

Bought software still has to be connected to the place documents arrive and to your own systems. It also needs monitoring, staff for its review queue and an owner on your side, so get that work priced before you sign.

When a firm builds the pipeline, that work should be written into its contract. You pay for the build, then for servers and any GPUs (the processors that run the model), for the people who watch the system, and for updates when your documents or the model change. Bought software runs on your servers too: Microsoft's disconnected containers page recommends 8 processor cores and 24 GB of memory for each container it lists.

To compare the routes, give every vendor and firm the same written volumes (documents a month, pages per document, mix of types) and ask each for a three-year cost. The cost should include your servers and your staff's review time.

What do we own at the end of each route?

When you buy, you own the right to use the software while the licence runs. Your configuration, such as document types, fields and trained models, stays in the vendor's format, so ask whether you can export it.

The licence can stop the software from working, even inside your perimeter. Microsoft's disconnected containers need a licence file, and Microsoft says that file “also contains an expiration date, after which the license file is invalid to run the container.” ABBYY's terms say that if you do not pay for use beyond the volume units you bought, “ABBYY Software functionality will not be available for Licensee.”

The code of a custom pipeline becomes your company's if the contract transfers it, so read what the contract excludes, such as the firm's reusable components (code the firm reuses across projects). Put in writing that your acceptance set and every test result belong to your company on every route. They let you score any replacement on the same documents later.

What should we ask a software vendor and an engineering firm?

For a software vendor:

• Which version are we licensing, and can you list in writing which of its features run inside our perimeter?
• Does any part of the software connect to you from inside our perimeter, for licences, updates, usage reports or support?
• Can the trial run inside our perimeter, in the version we would license, and be scored on our acceptance set?
• What stops working if we do not renew, and in what format can we export our configuration and trained models?

For an engineering firm:

• Have you taken a system to production inside a client's perimeter, and what can you show us about it?
• Will you build the prototype inside our perimeter, under accounts we control, and be scored on the part of the acceptance set you have not seen?
• For which of our document types would you recommend buying software, and why?
• Who runs the system after launch, what does the contract transfer to us, and which reusable components does it exclude by name?

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 whether that project used bought document AI software, where its model ran or which of the practices above it used. Nothing here says that amBrain compared document AI software with a custom pipeline, in that project or any other, or that it has built systems for invoices, identity documents or support tickets. No amBrain prices, timelines or terms for a prototype or a test before signing are given.

amBrain has been building software since 2019. It describes how it works with clients in one line: “Three formats: full delivery, a dedicated team, or engineers embedded in your team.” On ownership, its line 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 questions as any other firm, and ask for the list of those components by name.