Documents in. Verified data out.
Faster. More accurate. Lower cost. Proven on public and customer-specific benchmarks. Upload a business document and receive structured, validated data with field-level evidence.
P50 · P95 · verified intakes/min
critical-field escapes · review rate · evidence
GPU + APIs + retries + human review
No numbers without a manifest.
Every future figure will open its dataset, hardware, date, commit and methodology, with downloadable results. Until real BiVelio-measured results exist, this page shows Benchmark in progress — never simulated numbers or pre-recorded demos presented as real.
Does it preserve text, layout, tables, formulas and reading order?
Benchmark in progressIs the output useful for agents and workflows?
Benchmark in progressDo the concrete fields come out correct at comparable cost and latency?
Benchmark in progressDoes it generalize across invoices, templates and languages?
Benchmark in progressDoes it win on the documents, languages and fields we sell?
Benchmark in progressWhat throughput and P50/P95 does the real hardware deliver?
Benchmark in progressFully Automatic
No human intervention: the document goes in and the result is published exactly as it left the system, errors included.
Verified Outcome
With review-by-exception included: what we would actually deliver in production, measured end to end.
Transparency: accuracy is always published together with Review Rate and Abstention Rate — never on its own.
Five steps, one CTA.
Describe what you need and the system picks the fastest route above the quality floor. Basic users never build schemas or pipelines, and engine names stay hidden on purpose.
PDF · image · email
the fields you need or your business goal
the fastest route above the quality floor
evidence + validators + exceptions
JSON · export · webhook · BiVelio
Simplicity rule: a single CTA and zero mandatory configuration. Nobody needs to know which engine processed their document — only that the result arrived verified, with its evidence.
A benchmark that leaves the product running.
Seven steps, from first upload to activation. The same run works as benchmark, as onboarding and as a sales tool — with no extra steps.
- 01Upload 5–20 documentsreal ones, from your own operation
- 02Define the outcomewhat must come out the other end
- 03Describe the fields"name, NRT, date, total, IBAN and line items"
- 04See the evidenceevery value opens the exact snippet it came from
- 05Confirm the errorsyou decide what counts as a failure
- 06Receive the proofresults on your documents, not a demo
- 07Activate itthe same flow stays configured in production
One run, three jobs
- Benchmark. Measured on your documents and with your error criteria, not on a dataset that flatters us.
- Onboarding. By confirming the errors you have already configured the product: activating it is a click, not a project.
- Sales tool. The final report is the evidence you can put in front of whoever signs.
Initial scope, stated honestly: Supplier & Invoice Intake — invoices, purchase orders, receipts, credit notes, supplier bank forms and tax certificates — in ES · CA · EN · FR. We do not promise "any document".
One unit. Four tiers. Zero surprise multipliers.
Everything is measured in Intake Units, and the tiers are published without figures on purpose: the number will arrive once it is measured, not before.
One input processed to the agreed state. Classify, split, OCR, extract, validate and route never appear as surprise multipliers.
Free / Sandbox
To try the benchmark with no commitment.
- Sample documents included
- Limited volume
- No card
Developer PAYG
Pay as you go, to build on top.
- Intake Units on demand
- API and webhooks
- Transparent limits
Operate
The operations layer, as a separate tier.
- Cases and policies
- Approvals
- Actions
Enterprise
Offered once implemented — not before.
- Reserved throughput
- Region and retention
- SLA and support
Honesty note: definitive pricing is published only after measuring P95 variable cost and review exposure under real load.
Evidence first. Unknown is a valid output.
A value without its evidence is a model's opinion, not data. Intake is built the other way round: proof first, schema second.
Evidence first
Every critical field opens its exact evidence: the snippet of the document the value came from.
Abstention rule
Unknown is a valid output — Intake never invents a value to complete a schema.
Isolation per schema
Each customer's data lives isolated per schema: no tables shared across tenants.
EU hosting
Data hosted and processed in the European Union, with retention controlled by contract.
API-first
JSON, export, signed webhooks and idempotent delivery: everything the UI does can be done via API.
Controlled retention
You decide how long we keep your documents — including the option of not keeping them at all.
Bring 5–20 documents. Leave with proof.
The benchmark runs on your documents and with your error criteria. No card, and with sample documents if you cannot share yours yet.