AI document processing with local models: a buyer’s guide for European teams
TL;DR
Define the job first: document families, monthly pages with the backlog counted separately, a 10–20 field schema, the delivery format and who works the exceptions. Without these you cannot score a vendor.
There are four ways to buy: a SaaS extraction tool, a hyperscaler document API, a managed pipeline on local models, or DIY on-prem. They differ most on data location, ops burden and exit.
Judge accuracy by field-level precision and recall, straight-through rate and human-touch rate on your own 200-page sample. 98% per field across 12 fields is about 78% per document.
At list prices, a routine invoice page costs about €0.007 in GPT-4.1 inference and €0.0015 of compute on a rented EU GPU; a managed pipeline costs more per page because it includes schema, validation and delivery.
For DACH, Nordic and Benelux buyers, where inference runs and which subprocessors touch the pages often decide the vendor before cost does. Get both into the GDPR Art. 28 agreement.
Questions people ask
What is AI document processing with local models?
It is extraction — invoices, contracts, claims or forms turned into CSV, JSON or API records — performed by open-weight models such as PaddleOCR-VL, DeepSeek-OCR or Qwen3-VL running inside the EU or on your own hardware. Pages never go to a US API, so there is no GDPR transfer to justify, and cost is driven by compute and pipeline work rather than per-token fees.
How do I evaluate document AI accuracy honestly?
Label a 200-page sample of your own documents with ground truth and score every candidate on it: field-level precision and recall per document family, the straight-through rate (documents needing no human edit) and the human-touch rate. Ignore document-level claims measured on the vendor’s demo set, and check whether per-field confidence actually predicts the errors.
What does AI document processing cost per page?
At list prices at the time of writing, a routine invoice page costs roughly €0.0014 on GPT-4.1 Mini, €0.007 on GPT-4.1 and €0.012 on Claude Sonnet 5 in inference alone, and about €0.0015 of compute on a rented EU GPU. Managed pipelines cost more — Ækora’s published plans work out to €0.067–€0.10 per page — because they include schema, validation, exception handling and delivery.
Which deployment option is best for a European company?
It depends on volume, engineers and residency. A SaaS tool fits a few thousand documents a month with no integration work. A hyperscaler API fits teams with engineers already in that cloud. A managed pipeline on local models fits steady volume where residency matters and nobody wants to run GPUs. DIY on-prem fits very high volume with a platform team and a hard residency rule.
What GDPR questions should I ask a document AI vendor?
Where inference runs and whether that location is in the Art. 28 agreement; the full subprocessor list and any transfers outside the EEA with their legal basis under Art. 44–49; whether your pages train models; retention and deletion at contract end; Art. 32 security measures; and, for special-category data such as prescriptions or claims, support for an Art. 35 impact assessment.
Does the EU AI Act apply to document extraction?
Usually as a low-risk use. Document extraction becomes high-risk only when its output feeds an Annex III decision such as creditworthiness, employment or access to essential services; those obligations apply from 2 August 2026. In the typical setup the vendor is the provider and you are the deployer, and the Art. 4 AI-literacy duty has applied to both since 2 February 2025.
Want this worked out on your documents?
We will price a real sample against OpenAI or Anthropic and tell you whether Bulk document processing or a local model on your hardware is the cheaper first step.
AI document processing with local models means running extraction — invoices to CSV, contracts to a dataset, claims to JSON — on open-weight models inside the EU or on your own hardware, instead of sending every page to a US API. For a European buyer the decision rests on four numbers: monthly page volume, the share of pages a human still touches, the all-in cost per page, and whether the pages may leave the EEA at all. On a routine invoice page a frontier API costs about €0.007–€0.012 in inference; a rented EU GPU about €0.0015; a managed pipeline more than either, because it includes the people and the plumbing. Below: how to define the job, four ways to buy it, how to test accuracy without being fooled, what each pricing model costs at 10,000, 75,000 and 250,000 pages a month, and a six-step pilot.
Define the job before you talk to a vendor
Five things need to be written down before the first demo.
Document families, not “documents.” Supplier invoices from your top 50 vendors are one family; phone photos of expense receipts another; a 40-page framework agreement a third. Each has its own layout variance, language mix and scan quality, and each will produce a different accuracy number. List the families with a monthly count.
Monthly volume in pages, backlog counted separately. Everything is priced per page, and a multi-page invoice is several pages. A backlog is a spike; the run rate is a floor.
Fields. Field name, type, required or optional, and what “correct” means (ISO date, amount in cents, IBAN without spaces). An invoice header is typically 10–20 fields plus a line-item table. If you cannot write the field list, you cannot score a vendor.
Delivery format. CSV to a folder, JSON per document, or an API call into the ERP, DMS or claims system — one row per document or per line item, with the source-file reference and per-field confidence.
Exception handling. Every pipeline produces a queue of pages it cannot finish. Decide who works it, in which tool, within what time. The exception queue is where the real cost hides.
Four ways to buy document processing
1. SaaS extraction tool
Parseur, Rossum, Klarity and similar: forward an email or upload a file, the vendor’s models extract, you correct in their interface, you export. The fastest start, and the right fit for a few thousand documents a month handled by one team. Check which subprocessors sit behind it — many run on US hyperscalers — and whether your corrections can leave with you.
2. Hyperscaler document API
Azure AI Document Intelligence, Amazon Textract and Google Document AI. You choose a region (EU regions exist), pay per page, and build everything around the call: ingestion, validation, the exception queue, delivery. A good fit when you already have engineers in that cloud. Two trade-offs: the pipeline is yours to run while the model is theirs, and a US parent company keeps the CLOUD Act question open even when the region is Frankfurt.
3. Managed pipeline on local models
This is what we do with bulk document processing: open-weight models of the PaddleOCR-VL, DeepSeek-OCR and Qwen3-VL class, run as a pipeline that includes the schema, validation rules, exception queue and delivery as CSV, JSON or API, priced as a fixed monthly plan with a page allowance. Dedicated EU infrastructure or a deployment on your own hardware is scoped up front. The fit is steady volume, no appetite for running GPUs, and residency that matters. The trade-off is honest: a managed pipeline costs more per page than raw inference, because it includes the work raw inference leaves to you.
4. DIY on-prem
Your GPU, your open-weight models, your engineers. Cheapest per page at high volume, most control, most work: model selection, drivers, a queue, monitoring, on-call. It suits organizations above roughly 100,000 pages a month with a platform team and a hard residency rule. The stack we deploy is described under local document models; the concept itself is covered in what on-premise document AI is.
The four deployment options compared on the dimensions European buyers ask about most. Qualitative; verify every cell against the specific vendor’s contract and your own sample.
Option
Data location
Cost model
Setup time
Ops burden
Accuracy control
Exit / lock-in
SaaS extraction tool
Vendor cloud; region and subprocessors vary — ask
Per document or per seat, tiered
Days
Low
Vendor models; corrections in their UI
Data exports; templates and corrections rarely portable
Hyperscaler document API
Region you choose (EU available); US parent company
Per page or per 1,000 pages, plus the pipeline you build
Weeks
Medium — you run the pipeline
Prebuilt and custom models in their console
Pipeline is yours; custom models are not portable
Managed pipeline on local models (Ækora)
Vendor-run open-weight models; dedicated EU infrastructure or your hardware as an option
Fixed monthly plan with page allowance, per-page overage
Weeks — pilot on your sample first
Low — vendor runs it
Field-level scores on your sample; validation rules; exception queue
Open-weight models, your schema and rules are portable; output is plain CSV/JSON
DIY on-prem
Your data center or an EU GPU you rent
Hardware or GPU rental plus engineering time
Months
High — you own everything
Full control, full responsibility
None — you own it
The four deployment options compared on the dimensions European buyers ask about most. Qualitative; verify every cell against the specific vendor’s contract and your own sample.
How to evaluate accuracy honestly
“99% accuracy” is not a number until you know the denominator — percent of what, measured on whose pages? Four measurements on your own 200-page sample replace the claim.
Field-level precision. Of the values the system emitted, the share that were correct. On 200 invoices with 12 header fields there are 2,400 values; if the system emitted 2,300 and 2,185 were right, precision is 95%.
Field-level recall. Of the values that exist on the pages, the share found correctly. Same sample: 2,185 of 2,400 is 91%. Low recall means blank cells in the ERP; low precision means wrong ones, which is worse.
Straight-through rate. The share of documents that pass every validation rule and reach the target system with no human edit. If 150 of 200 do, it is 75%.
Human-touch rate. The share of documents where a person changed at least one field. It drives your staffing, and it is the number vendors quote least.
Field accuracy and document accuracy are different metrics, and the gap is arithmetic: a system that is 98% correct per field across 12 independent fields produces an error-free document about 78% of the time (0.98 to the twelfth power). That is how a headline per-field score coexists with a 25% human-touch rate. Score per family, normalize before comparing, and ask for per-field confidence — then check whether low confidence predicts the errors. If it does, the threshold is your exception router; if not, it is decoration.
Pricing models and what they cost at three volumes
Document processing is sold four ways. Per page (hyperscaler APIs, some SaaS tools): predictable, linear. Per token (OpenAI or Anthropic behind a pipeline you build): cheap at low volume, sensitive to image resolution and retries. Per seat (review-centric SaaS): cheap when few people touch it, wrong for bulk. Fixed plan (managed pipelines): a monthly fee for a page allowance plus overage, pipeline included.
A routine invoice page is about 2,200 input and 400 output tokens; a hard page (dense form, stamped scan) about 8,000 and 800. At list prices at the time of writing, converted at $1 = €0.92, a routine page costs about €0.0014 on GPT-4.1 Mini, €0.0070 on GPT-4.1 and €0.0116 on Claude Sonnet 5. Retries and second passes add 1.3–2.0×, and high-resolution scans push pages into the hard profile. Local compute: an EU A100-class rental at about €1.80 an hour, running a compact document VLM at roughly 1,200 pages an hour, is €0.0015 per page; a reserved GPU line is about €600 a month. Our published plans are the fixed-plan example: Pilot at €999 a month for 10,000 pages (€0.15 per extra page, one document type), Scale at €4,999 for 75,000 pages (€0.09 per extra page, multiple document types, API delivery), Enterprise on request.
Illustrative scenario, assumptions stated here. Routine invoice profile (2,200 input + 400 output tokens); list prices at the time of writing, $1 = €0.92. API rows are inference only — no pipeline, validation or people. The GPU row is a €600/month reserved A100-class card including power and a slice of ops, excluding the engineers who build and run the pipeline. Ækora rows are published plan prices and include schema, validation, exception queue and delivery.
Route
10,000 pages / month
75,000 pages / month
250,000 pages / month
GPT-4.1 Mini, sticker
€14
€105
€350
GPT-4.1, sticker
€70
€525
€1,750
GPT-4.1 with 1.5× retries
€105
€788
€2,625
Claude Sonnet 5, sticker
€116
€870
€2,900
Reserved EU GPU, compute + ops slice
€600
€600
€600
Ækora managed pipeline, published plans
€999 (Pilot)
€4,999 (Scale)
€20,749 at Scale overage; Enterprise on request
Illustrative scenario, assumptions stated here. Routine invoice profile (2,200 input + 400 output tokens); list prices at the time of writing, $1 = €0.92. API rows are inference only — no pipeline, validation or people. The GPU row is a €600/month reserved A100-class card including power and a slice of ops, excluding the engineers who build and run the pipeline. Ækora rows are published plan prices and include schema, validation, exception queue and delivery.
The GDPR and residency questions to ask
For many DACH, Nordic and Benelux buyers, residency decides before the spreadsheet does. Get the answers into the contract, not the sales deck.
Where does inference run? Country, provider, and whether that location is written into the data processing agreement (GDPR Art. 28). “EU region” of a US provider is an answer; it is not the same answer as a European provider or your own rack.
Which subprocessors, and are any outside the EEA? Art. 28(2) requires your prior authorization. For anything outside the EEA, what is the transfer basis under Art. 44–49 — standard contractual clauses, or the EU-US Data Privacy Framework for a certified US firm — and what changes if that basis falls, as Privacy Shield did in Schrems II in 2020?
Are our pages or extracted values used to train anything? Plus retention, deletion or return at contract end (Art. 28(3)(g)), and the Art. 32 measures: encryption in transit and at rest, access logging.
Special-category data. Prescriptions, claims and HR files carry Art. 9 data. A data protection impact assessment under Art. 35 is likely; ask whether the vendor supplies the inputs for it.
EU AI Act role. Most document extraction is not high-risk; check Annex III if the output feeds decisions on credit, employment or access to essential services, since the main high-risk obligations apply from 2 August 2026. Usually the vendor is the provider and you are the deployer; the AI-literacy duty in Art. 4 has applied to both since 2 February 2025.
Local models on your own hardware remove the transfer question rather than answering it; a managed pipeline on dedicated EU infrastructure keeps the pages inside the EEA. Ask for either in writing.
A six-step pilot plan
Pick one document family and freeze the schema. The highest-volume family, not the hardest. Write the 10–20 fields, their types and the normalization rules. Change nothing until step 6.
Pull a 200-page sample that mirrors reality and label it. Same proportion of vendors, languages, scans and phone photos as the live stream. Your own people key the ground truth. Do not send it to the vendor.
Run every candidate on the same sample and score it. Field-level precision and recall, straight-through rate, human-touch rate, per family. Log tokens or pages so you can price the result, and record whether confidence predicted the errors.
Write the validation rules and the exception route. VAT arithmetic, IBAN checksum, date order, PO match, duplicate check. Define who sees an exception, in which tool, within how many hours.
Wire delivery and run one live month in shadow mode. CSV to a folder or JSON to the API, alongside the current process. Compare what the pipeline produced with what the team actually booked.
Decide on numbers, then widen. All-in cost per page including exceptions and people, human-touch rate, residency answered in the contract. Only then add the second family.
Common mistakes
Buying on a document-level accuracy claim measured on the vendor’s pages instead of field-level scores on yours.
Leaving the exception queue out of the business case, then discovering it needs two people.
One frontier model for everything. The most common budget failure. The design that holds is hybrid: local models for the routine ~95%, a frontier API or a human for the ~5% exceptions, gated by confidence.
No exit clause. If the export, the schema and the corrections cannot leave with you, the discount was not a discount.
Measure four things this quarter: pages per month by family, human-touch rate on a labeled sample, all-in cost per page including the people who work the exceptions, and where each family is allowed to be processed. Those numbers tell you which of the four routes fits, and when it stops fitting. If the harder question is which of your workflows belong on local models at all — not only documents — that mapping is what local AI consulting is for.