Thirty to forty-five days, a fixed fee, and DocxIntel deployed inside your own infrastructure. The accuracy threshold on your document set, the success criteria and the price the licence converts to are all agreed in writing before the first container starts. Miss the threshold and you keep the report and owe nothing further.
The fee is set by the document set, the number of extraction schemas and the complexity of the target environment. It is quoted after a one-hour scoping call and does not change once agreed.
By the end of a Proof of Value you have DocxIntel running inside your perimeter, a measured accuracy figure on your own paperwork, and the numbers you need to write a business case. Not a slide deck.
Kubernetes or Docker Compose on your hardware or in your own cloud tenancy, behind your firewall, using your storage and your identity provider. Air-gapped installs are delivered on physical media and never open an outbound connection.
Accuracy is scored against ground truth your named reviewer adjudicates, on documents you selected, using the normalisation and scoring rules published on the accuracy page. Our benchmark set plays no part in the number you receive.
Read the methodologyOne price agreed in advance, one calendar window, one named engineer. No hourly billing, no scope creep and no pilot that quietly runs for nine months without producing a decision.
At least one live path is wired end to end — a watched folder, an object-storage bucket, an intake mailbox or a direct API call — so you see documents move through your own pipeline rather than through a demo portal.
We run your volume against the deployed footprint to establish pages per hour on your real document mix, then size the production estate from measured numbers rather than from a data sheet.
Sizing and throughputIf you do not proceed, the containers and model bundles are removed under your change control and the report stays yours. Nothing was uploaded anywhere, so there is no vendor-side copy of your documents to delete.
We charge for the Proof of Value because it consumes the same engineering that a production deployment consumes, and because a paid engagement is the only kind that reliably ends in a measured number. The fee buys commitment on both sides — ours to a written threshold, yours to a named reviewer and an environment.
The trade is straightforward. You pay a fixed fee; we accept a written accuracy threshold on documents we have not seen. If we miss it, you keep the report and owe nothing further.
Every argument at the end of an evaluation comes from something that was left vague at the beginning. These four are settled first, in the agreement, in plain language.
A field-level accuracy figure, per document category, measured against ground truth your reviewer adjudicates, with the normalisation rules and the all-fields denominator written out. Both parties sign the definition of correct before either sees a result.
The exact categories, the count per category and the volume set used for throughput testing are all named. The set cannot be quietly narrowed later to help the result, and it cannot be widened later to sink it.
Accuracy alone does not make a decision. The agreement also states the throughput target, the integration paths to be proven, the review-queue behaviour to be demonstrated and any security controls your team needs evidenced.
The annual licence figure for the production footprint is fixed before the Proof of Value begins, so a strong result cannot be used to reprice the deal and a weak one cannot be used to discount your way into a bad decision.
A 30 to 45 day window, structured so the accuracy study starts in week two rather than in the last three days.
One hour with an engineer to fix the document categories, schemas, threshold, throughput target and conversion price, followed by the signed agreement. Your team confirms the environment, GPU access, the named reviewer and the security path our engineers will work through.
The bundle is installed on your infrastructure, identity is wired to your OIDC or SAML provider, storage and one live ingest path are connected, and the first documents flow through the pipeline. Air-gapped installs follow the offline runbook from physical media.
Extraction schemas are configured against your document categories while your reviewer, using our adjudication tooling, establishes the ground truth for the evaluation set. These run in parallel, and the ground truth is settled without reference to any model output.
The scoring script runs against the adjudicated set and reports per-category, per-field results, along with what would have been routed to review at each confidence threshold. The volume set runs separately to establish pages per hour on your hardware.
You receive the accuracy report, the throughput measurements, the production sizing recommendation and the integration plan, and we walk your team through all four. Then the threshold is either met or it is not, and the pre-agreed outcome follows.
Nothing exotic, but every item on this list has delayed a pilot somewhere. Confirm them during scoping and week one becomes deployment rather than procurement.
If any prerequisite is genuinely blocked — no GPU available, no reviewer free until next quarter — say so during scoping. We would rather move the start date than run a Proof of Value that produces a number nobody trusts.
Free trials are excellent for answering whether a tool is pleasant to use. They are structurally incapable of answering whether it is deployable in a regulated environment at your accuracy and throughput.
Where your documents are processed
Documents you can use
Volume allowance
Accuracy result you end with
Contractual accuracy threshold
Throughput evidence
Security review satisfied
Cost
If the result is disappointing
Trial mechanics described here reflect publicly documented free tiers of managed document-parsing and IDP services as of 2026, including credit-capped free plans and 14-day trials of a few hundred to a few thousand pages. Verify against current vendor documentation before comparing.
Six things exist at the end of a Proof of Value regardless of the decision you make. All of them are yours.
DocxIntel running on your infrastructure with your identity provider, your storage and at least one live ingest path connected — not a demo tenancy that expires.
Per-category and per-field results against your adjudicated ground truth, the review-queue volume at each confidence threshold, and a plain account of which of your document types performed worst and why.
GPU workers, CPU, memory and storage for your target throughput, derived from measured pages per hour on your own document mix, with the resilience options priced separately.
How the output lands in your core systems, where the review queue sits in your operating model, which schemas need work, and what your team has to build versus configure.
If the contracted accuracy threshold on your document set is not met, you keep every deliverable above and owe nothing further — no licence obligation, no extension fee, no conversion.
Threshold met and you want to proceed: the licence converts at the price fixed before the Proof of Value began, the evaluated deployment becomes the licensed deployment, and the footprint in the order is what renews annually.
How the licence is pricedThe practical objections, answered by the people who would run it.
Because it is not a trial — it is a deployment. Engineers install DocxIntel inside your infrastructure, configure schemas against your document types, run a measured accuracy study on ground truth your reviewer adjudicates and hand back a sizing recommendation and an integration plan. Free pilots are staffed accordingly: they get whoever is spare, they slip, and they end without a number anyone will sign. A fixed fee buys named engineers, a fixed end date and a written result.
