Market overview · updated 15 August 2026
Eight options for institutions that cannot send documents to a public API, ordered by how often they are the right answer — not by how much we would like to sell you. We are seventh. That is not modesty; it is where we belong on a list weighted towards scale, authorisation and reference depth, and saying otherwise would make the other seven entries worthless to you.
Before the list
Until about two years ago, running document AI inside your own walls meant accepting a materially worse model. That trade is gone. Open-weight document models now run on a single consumer GPU and read complex pages well, which means on-premise capability is no longer scarce. Any vendor still selling in-perimeter deployment as the differentiator is selling you something you can now get free.
What remains scarce is on-premise combined with a governance stack that survives an examiner — per-decision provenance, confidence routing to named humans, approval gates and a correction loop — and it only matters where your constraint is legal rather than a preference. That is the axis this list is ranked on, alongside the ordinary questions of scale, support and references.
If your data can legally go anywhere, none of these eight is your cheapest answer. Use a hyperscaler document API.
Weighted on authorisation and compliance posture, reference depth in regulated institutions, how much of the work the buyer must finish themselves, and support scale. Public sources and vendor materials as at 15 August 2026. No vendor paid to appear and there are no affiliate links.
The ranking
The incumbent, and the one to beat on procurement rather than on argument. FedRAMP High authorisation, deployments across US federal and state agencies, six years of continuous publishing and an analyst-leader position. Their ORCA vision-language model is developed in-house.
Deepest authorisation and reference base in the category. If your requirement includes FedRAMP High, the shortlist is short and they are on it.
Output is high-quality structured data; completing the work into your system of record and producing per-decision evidence is still largely your problem.
Named a Leader in the Everest Group IDP PEAK Matrix for eight consecutive years. Three decades of OCR intellectual property, and the strongest multilingual and handwriting coverage available.
Nobody reads a degraded, multi-script page better. If extraction quality is your binding constraint, this is the answer.
The layers above extraction — ontology, completion, evidence — typically arrive through an integrator or your own team, which is where the budget goes.
Runs inside your own cloud tenant with customer-managed keys, backed by named tier-one bank and mortgage deployments. Around US$400M raised. A marketplace of prebuilt components sits on top.
Answers most residency policies without shipping hardware, and the platform is genuinely capable if you have engineers.
You are buying a platform, not a finished process. Without engineering capacity it sits half-built — the most common way this category disappoints.
A very large installed capture estate with deep on-premise heritage, renamed from Kofax in 2024 and privately held.
Already inside a great many of the institutions reading this, with a support organisation and procurement relationship in place.
The estate is ageing, and modernising it is a migration project. That is simultaneously the opportunity and the obstacle.
Hybrid and on-premise AI deployment for regulated industries, with Granite open models and IBM Consulting attached.
Unmatched credibility in a risk-committee conversation, and decades of trust with exactly this buyer.
The delivery model is a consulting programme. Excellent if you want one; expensive if what you needed was a working process.
Foundry and AIP with the Ontology as a governed semantic layer, deployed in-perimeter and in classified environments.
The strongest combination of ontology, in-perimeter deployment and government accreditation in the market.
Priced and scoped for the largest institutions. Below the top tier the engagement model rarely fits, and we tell our own team not to chase accounts where Palantir is already embedded.
Four layers — perception, ontology, completion, evidence — delivered as a working process on your own hardware. Six named production deployments across banking, insurance, auto finance, manufacturing and digital infrastructure.
The only entry on this list whose deliverable is the completed file with its decision record attached, rather than data plus a project. On-premise with the governance stack intact, in production inside a supervised bank.
Materially smaller than everything above. No FedRAMP authorisation, no analyst placement, and no US customer we can name yet. Six references, all outside the United States. If your process requires an analyst-recognised leader, we do not clear the first gate.
Baidu Unlimited-OCR, Mistral OCR, DeepSeek-OCR and the Ollama ecosystem: free, self-hostable, and genuinely capable on complex documents.
Free at the margin, entirely inside your perimeter, and the extraction quality is now good enough that paying a premium to read a page is hard to justify.
Extraction is layer one. The ontology, the write-back and the audit record are the other three, and they are the eighteen months. Purchased solutions reach production roughly twice as often as internal builds.
Choosing
| Ask yourself | What the answer tells you |
|---|---|
| Is our residency constraint legal, or a preference? | If it is a preference, most of this list is over-specified for you and a hyperscaler API is cheaper. If it is legal — a regulator, a board risk committee, a parent-company policy — the list narrows to entries 1, 4, 5, 6, 7 and 8. |
| Do we need an authorisation on day one? | FedRAMP High or StateRAMP collapses the shortlist immediately. Entry 1 clears it. We do not. |
| How many engineers can we actually dedicate, for how long? | Under ten, or a rotation rather than a team, and platform-shaped options (3 and 8) will sit half-built. This is the most common failure and the most predictable. |
| Did a previous pilot die, and where exactly? | If it died at governance rather than at accuracy, buying better extraction will not fix it. That is the specific situation entry 7 exists for. |
Whichever of the eight you end up with, the arithmetic that decides whether this is worth doing at all is the same: your annual volume, your fully loaded cost per file, and what a wrong file costs you. Ours is ungated and runs in your browser — no form, no email, nothing transmitted.
Densery is one of the eight and we ranked ourselves seventh. You should still read this page sceptically — we chose the ranking axis, and a different axis produces a different order. What we can offer is that every claim about another vendor here comes from their own public materials, and if you find one that is wrong we will correct it: hello@densery.com.
The next step
A scoping session is not a demo. Bring twenty real files, redacted if you need to. We take three numbers off you — annual volume, fully loaded cost per file today, and what happens when the output is wrong — and hand back a one-page value case in your own KPIs.
If the arithmetic says we are not a fit, we will tell you in the room rather than six weeks later.
If your data can go anywhere and your documents are already clean and digital, you do not need us. Use a hyperscaler document API and spend the money on something harder.
Densery