Every document AI decision comes down to the same architectural question: where does the page get read, and who holds the model that reads it. This page lays out the three real answers, the honest trade-offs of each, and how to run an evaluation that survives contact with your own documents.
Figures are list rates published by the named vendors as of 2026, multiplied out. They are illustrative arithmetic, not vendor quotations, and they ignore negotiated discounts, committed-use terms and plan-included credits. Verify current pricing with each vendor before budgeting.
Marketing language blurs these together. Architecturally they are different products with different failure modes, different cost curves and different answers to a regulator's first question.
You post a document to a vendor endpoint and get structured output back. Excellent developer experience, no infrastructure to own, and you pay per page or per credit for every page you send.
Vendor-controlled container images run inside your own cloud subscription. Your data stays in your tenant, which answers most residency questions — but it is normally gated behind the top commercial plan.
The full stack — services and the proprietary DocxIntel models — is installed on infrastructure you own. It can be disconnected from every network and keep working identically. This is where DocxIntel starts, not where it upgrades to.
Not a scorecard. These are the properties that change what you can promise a regulator, an auditor or a CFO — and the ones where the answer is genuinely uncomfortable for us are marked as such.
Where the page is read
Runs with no network egress
Who holds the model weights
Pricing model
Illustrative cost at 1M pages/month
Reprocessing 2M archived pages
Vendor-side copy of your data
Self-hosting available on
Vendor's role under data-protection law
Arabic as a primary script
Time to first working extraction
Infrastructure you must operate
Compiled from publicly published vendor pricing and documentation as of 2026. Vendor pricing, plan structure and deployment options change frequently — verify against current vendor documentation before making a purchasing decision. Where a vendor does not publish a figure, this table says so rather than estimating.
The metered model is genuinely cheap while you are learning. The problem is that the price of reading a page never falls, so the cost of your document estate becomes a permanent function of how many documents you hold.
Illustrative list-rate arithmetic using published 2026 vendor rates, before discounts and committed-use terms. Your negotiated rate will differ — run the same multiplication on the rate you are actually offered.
The words that get blurred in this market — self-hosted, BYOC, hybrid and truly air-gapped — defined precisely, with a due-diligence checklist you can put in front of any supplier, including us.
What each term actually means for where a page is read and who holds the models.
Read the guideSixteen due-diligence questions, written to be pasted straight into an RFP.
See the checklistFive exercises that test the architecture, not just the accuracy.
Read the methodMost document AI bake-offs are decided by a demo on the vendor's documents. That measures the wrong thing. Five steps that measure the right one.
Assemble 300 to 500 real documents in the mix you actually receive — phone photographs, handwritten Arabic, stamps over printed text, multi-generation faxes, multi-page packets that need splitting. A sample of clean native PDFs will rank every vendor as excellent and tell you nothing.
Adjudicate a ground truth once, then score each field independently across every candidate. A 99% character accuracy rate can still mean a wrong number in one invoice total in three. The field is the unit your business acts on, so the field is the unit you should score.
Include the pages you reprocess, the pages that fail and get retried, and the archive you will want to run again after the next model upgrade. Multiply by the published rate for the parse quality your accuracy target actually requires — not the cheapest tier on the pricing page.
Install the candidate in an isolated segment, pull the outbound route, and process a batch. Then read your own firewall and DNS logs. A product that phones home for licensing, telemetry or a model fetch will show up in those logs in minutes, whatever the datasheet says.
Get the model manifest in writing: named model, version and licence for every component. Then ask the continuity question — if the vendor is acquired or the contract lapses, does your deployment keep running on the artifacts already in your data centre, or does it stop?
Copy these into your evaluation document. They are the questions that separate the deployment archetypes from each other — and every vendor, including DocxIntel, should be able to answer them in writing.
These questions are deliberately vendor-neutral. Ask them of DocxIntel too — the Proof of Value exists so that the answers can be verified in your environment rather than believed.
Including the ones where the honest answer does not favour us.
Often, yes. If you are prototyping, processing a few thousand pages a month, have no data-residency constraint and already run a cloud-native stack, a metered API gives you working extraction in an afternoon with no infrastructure to own. The architecture only becomes a problem when volume, residency law or audit scope makes per-page metering and vendor-side processing untenable.
