25 RFP questions for choosing a document AI vendor in Europe
TL;DR
Five answers settle the shortlist: where inference runs, the subprocessor list, accuracy on your own 200-page sample, exit terms, and a GDPR Art. 28 agreement. A missing answer on any of them is a stop, not a weakness.
Ask which models do the work and whether they are open-weight. Open-weight models can be re-hosted if the vendor disappears; proprietary ones cannot.
Score accuracy only on your labeled sample, per field and per document family: 98% per field across 12 fields is about 78% per document.
Price the all-in cost per page at your real volume — plan, overage in a backlog month, retries if per-token, and the hours spent on exceptions — not the rate card.
Weight the six groups before reading answers, score each question 0–3, apply four knock-outs, and let a paid 30-day pilot with the top two pick the vendor.
Questions people ask
What questions should I ask a document AI vendor?
Group them: where our pages are processed and by which subprocessors; which models are used and whether they are open-weight; how accuracy is measured and whether they will run our labeled sample; pricing unit, overage and exit terms; SLA, throughput, exception handling and monitoring; and the GDPR Art. 28 agreement, EU AI Act roles, logging and AI-literacy support. This article lists 25 with model answers.
How do I compare document AI vendors fairly?
Send every bidder the same 200 labeled pages drawn from your real mix, keep the ground truth private, give everyone the same schema, normalization rules and deadline, and score per field with one script. Count any human edit as a touch. Do not accept results on the vendor’s own sample, and do not let a vendor choose or re-label pages.
Is the vendor the provider or the deployer under the EU AI Act?
In the usual setup the vendor that builds and supplies the system is the provider and the buyer that uses it is the deployer. If you substantially modify the system or put your own name on it, provider duties can shift to you under Art. 25. Most document extraction is not high-risk unless its output feeds an Annex III decision.
What should a document AI contract say about exit?
A notice period of 90 days or less, a full export of extracted data as CSV or JSON, written confirmation of deletion at contract end as required by GDPR Art. 28(3)(g), and hand-over of schemas, validation rules and corrections as documents. Ask whether the models are open-weight, since that determines whether the pipeline can be re-hosted.
Should I ask whether a document AI vendor uses OpenAI or Anthropic behind the scenes?
Yes, explicitly and per pipeline stage. Many products route low-confidence or hard pages to a frontier API as a fallback, which is a transfer outside the EEA even when the main processing is European. A good answer names the rule that triggers the fallback and offers a switch to disable it.
How should I score RFP answers for document AI?
Weight the six groups first — for example data location 25%, accuracy 25%, pricing and exit 15%, operations 15%, provenance 10%, compliance 10% — score each question 0–3 by specificity and evidence, treat a zero on data location, subprocessors, export or the DPA as a knock-out, and let the sample test on your own pages carry the accuracy group.
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.
Choosing a document AI vendor in Europe comes down to 25 RFP questions in six groups, and the answers to five of them — where inference runs, which subprocessors touch the pages, whether accuracy is measured on your sample, what leaving costs, and whether a GDPR Art. 28 agreement exists — settle the shortlist before any demo. This is the RFP we would want to receive. Each question comes with why it matters and what a good answer looks like, followed by a scoring method and the rules for a fair sample test. The tables apply equally to a SaaS tool, a hyperscaler API, a managed pipeline on local models and a DIY build.
Before you send it
An RFP without a defined job produces answers you cannot compare. Write down your document families, monthly page volume (backlog separate from run rate), the field list with normalization rules, the delivery format and who works the exceptions — the method is in our buyer’s guide to document processing with local models. Label a 200-page sample with ground truth, and weight the six groups below before you read a single answer; otherwise the best pitch sets the weights.
Group 1: data location and subprocessors
Where the pages go decides whether GDPR Chapter V applies at all, and whether the US CLOUD Act can reach the provider. If you buy in the Nordics or the Baltics, check which residency questions actually have legal force there before you weight this group.
Group 1 — data location and subprocessors. Questions 1–5. Answers should end up in the data processing agreement, not only in the RFP response.
Question
Why it matters
What a good answer looks like
1. In which country, and on whose infrastructure, does inference on our pages run?
Determines whether a GDPR transfer exists and whether a non-EU parent company is in the chain.
A named country and provider written into the DPA. “EU region of a US provider” stated as exactly that.
2. List every subprocessor that can access our pages or extracted data, with location.
Art. 28(2) requires your prior authorization; the list is where undisclosed API calls surface.
A complete list with locations, a notification duty for changes and a right to object.
3. If any processing happens outside the EEA, what is the transfer mechanism, and what is the fallback if it falls?
Schrems II (2020) ended Privacy Shield; SCCs and the 2023 EU-US Data Privacy Framework can also be challenged.
SCCs or DPF certification named, plus a contractual option to move processing EU-only.
4. Can the system run on our hardware, or in an EU data center we choose?
Removes the transfer question instead of answering it; often decisive for Art. 9 data.
A clear yes with a reference deployment, or a clear no. “We are looking into it” is a no.
5. Are our pages, extracted values or corrections used to train or improve models for anyone else?
Purpose limitation (Art. 5(1)(b)); your vendor data could surface elsewhere.
A contractual no by default, or an explicit opt-in with a documented scope and duration.
Group 1 — data location and subprocessors. Questions 1–5. Answers should end up in the data processing agreement, not only in the RFP response.
Group 2: model provenance
Provenance tells you whether the capability can be re-hosted if the vendor disappears, and whether a hidden frontier-API fallback is quietly moving 5% of your pages to the US.
Group 2 — model provenance. Questions 6–9. Open-weight models are the difference between an exit plan and a rebuild.
Question
Why it matters
What a good answer looks like
6. Which models perform OCR and extraction, by name and version?
You cannot assess portability, hardware needs or change risk without names.
Named models and versions (open-weight examples: PaddleOCR-VL, DeepSeek-OCR, Qwen3-VL) with a change-notice policy.
7. Are the models open-weight or proprietary, and under which license?
Open-weight models can be re-hosted by you or another vendor; proprietary ones cannot.
License named. For proprietary models, an escrow or documented migration path.
8. Does the pipeline call any third-party API — OpenAI, Anthropic, Google, Azure — for any page, including as a fallback?
The “hard 5%” fallback is a transfer too and is often undisclosed.
Yes or no per pipeline stage, the routing rule, and a switch to disable it.
9. What hardware do the models need, and can they be swapped when a better open-weight model appears?
GPU class sets DIY cost; swappability protects you against a stale model.
A GPU class named and a swap procedure that includes re-scoring on your sample.
Group 2 — model provenance. Questions 6–9. Open-weight models are the difference between an exit plan and a rebuild.
Group 3: accuracy measurement
Every vendor claims 99% of something. The only comparable number in the RFP is the result on your labeled sample, scored per field under identical conditions.
Group 3 — accuracy measurement. Questions 10–13. Insist on per-field results on your sample; treat a summary slide as a non-answer.
Question
Why it matters
What a good answer looks like
10. How do you define accuracy — per field, per document or per character — and on which sample?
98% per field across 12 fields is about 78% per document; vendor demo sets flatter.
Per-field precision and recall, per document family, measured on the buyer’s sample.
11. Will you run our 200-page labeled sample under the same conditions as every other bidder and return per-field output?
It is the one number you can compare across vendors.
Yes, with raw per-field output and the pages or tokens consumed, so you can price it.
12. What straight-through and human-touch rates do you expect on our sample, and how are they measured?
These set your staffing, not the headline score.
Definitions and a method. Declining to guess before the test is a good sign.
13. Do you output per-field confidence, and does it predict errors?
Confidence is the exception router; if uncorrelated with errors it is decoration.
Yes, with evidence from the sample that errors concentrate below the threshold.
Group 3 — accuracy measurement. Questions 10–13. Insist on per-field results on your sample; treat a summary slide as a non-answer.
Group 4: pricing and exit
Per-token bills swing with image resolution and retries; per-page plans do not. Either works if the unit is defined in writing and the exit is priced at signing, not discovered at renewal. If a vendor also offers an on-premise or air-gapped edition, ask for the disconnected container price, not the cloud rate: the two can be priced very differently.
Group 4 — pricing and exit. Questions 14–17. Ask for a worked price at your stated volume, including a backlog month.
Question
Why it matters
What a good answer looks like
14. Is pricing per page, per document, per token, per seat or a fixed plan — and what exactly counts as a page?
Per-token cost moves with resolution and retries; “page” definitions differ between vendors.
A unit definition in writing and a worked price at your volume, per document family.
15. What is the overage rate, and what happens in a backlog month?
A backlog is a spike; overage terms can exceed the plan itself.
The overage rate stated, plus an option to step up a tier for a single month.
16. What does the price include — schema setup, validation rules, exception handling, delivery, support?
Raw inference and a managed pipeline look alike on a rate card and are not the same product.
An itemized scope with what is extra and at what price.
17. Contract term, notice period, and on exit: export format, deletion timeline, portability of schemas, rules and corrections?
Lock-in is decided at signing. Deletion at contract end is an Art. 28(3)(g) duty.
Notice of 90 days or less; full export as CSV or JSON; written deletion confirmation; schema and rules handed over as documents.
Group 4 — pricing and exit. Questions 14–17. Ask for a worked price at your stated volume, including a backlog month.
Group 5: operations
Bulk processing has deadlines — month-end close, claims clocks — and templates change without notice. The exception queue is most of the ongoing cost; ask how it works before what it costs.
Group 5 — operations. Questions 18–21. Throughput and monitoring are what keep a cleared backlog at zero.
Question
Why it matters
What a good answer looks like
18. What SLA covers availability and turnaround, and what is the remedy?
Late extraction misses the close; an SLA without a remedy is a wish.
Turnaround per batch size in writing, with service credits or termination rights.
19. What throughput can you sustain, and how do you handle a 10× spike?
Backlogs and seasonality are normal; queues without a policy silently delay everything.
Pages per hour, a queueing policy, and priority handling for time-critical families.
20. How are exceptions handled — who sees them, in which tool, with what reason codes?
The queue is where the people cost lives; reason codes are how it shrinks.
A queue with reason codes, assignable to your staff, exportable, with a review cadence.
21. What monitoring do we get — volume, straight-through rate per family, exception age, accuracy drift?
A supplier changing its template shows up in the ledger unless someone is watching.
A dashboard or data feed with alerts, plus a scheduled re-score on a fresh labeled sample.
Group 5 — operations. Questions 18–21. Throughput and monitoring are what keep a cleared backlog at zero.
Group 6: compliance
Without a DPA the processing is unlawful on your side too. The EU AI Act adds roles: usually the vendor is the provider and you are the deployer, and most document extraction is not high-risk unless its output feeds an Annex III decision. Some public-sector buyers also work to national security criteria. For PiTuKri, if you buy for the Finnish public sector, ask the vendor to state which requirements land on them, which on you, and which on both.
Group 6 — compliance. Questions 22–25. GDPR articles cited in brackets; EU AI Act high-risk obligations apply from 2 December 2027 under Regulation (EU) 2026/1744, the AI-literacy duty since 2 February 2025.
Question
Why it matters
What a good answer looks like
22. Provide your data processing agreement (GDPR Art. 28) and your technical and organizational measures (Art. 32).
Instructions, confidentiality, subprocessor flow-down, audit rights and deletion are mandatory contents.
A signed-ready DPA covering all Art. 28(3) items; encryption in transit and at rest, access control and logging described.
23. Under the EU AI Act, do you consider yourself the provider and us the deployer, and what changes if we modify or re-brand the system?
Roles set obligations; a substantial modification can shift provider duties to you (Art. 25).
A clear role statement, and conformity documentation if any of our uses fall under Annex III.
24. What logs are kept, for how long, and can we access them?
Accountability under GDPR Art. 5(2); for high-risk uses, deployers must keep the system’s logs (Art. 26).
Logs of inputs, outputs and routing decisions, a stated retention period, exportable on request.
25. What do you provide to support our AI-literacy duty (Art. 4) and a DPIA (GDPR Art. 35) where one is required?
Both duties are yours, not the vendor’s, but you cannot meet them without their inputs.
Documentation on model behavior and limitations, a data-flow description, and a DPIA input pack.
Group 6 — compliance. Questions 22–25. GDPR articles cited in brackets; EU AI Act high-risk obligations apply from 2 December 2027 under Regulation (EU) 2026/1744, the AI-literacy duty since 2 February 2025.
How to score the answers
Weight the six groups before reading anything. A reasonable European starting point: data location 25%, accuracy 25%, pricing and exit 15%, operations 15%, model provenance 10%, compliance 10%. Move weight toward data location and compliance for Art. 9 data.
Score every question 0–3. 0 is no answer or evasion; 1 is a generic claim; 2 is a specific answer; 3 is specific and evidenced — a contract clause, a document, a test result.
Mark the knock-outs. A 0 on question 1, 2, 17 or 22 removes the vendor regardless of total. No location, no subprocessor list, no export or no DPA is a stop, not a weakness.
Score the sample test separately and let it carry the accuracy group. Per-field precision and recall on your ground truth, plus the human-touch rate, replace whatever was written in the response to questions 10–13.
Compute all-in cost per page at your real volume. Plan fee, overage in a backlog month, retries if per-token, plus the hours your people spend on the exception queue at the measured human-touch rate. Compare that number, not the rate card.
Run a paid 30-day pilot with the top two, on one document family, with exit terms agreed before it starts. The RFP picks the finalists; the pilot picks the vendor.
Running the sample test fairly
Send every bidder the same 200 pages, drawn in proportion to your real mix of vendors, languages, scans and phone photos — ugly ones included. Keep the ground truth to yourself. Give every vendor the same schema, the same normalization rules (ISO dates, amounts in cents, IBANs without spaces) and the same window. Score per field and per family with one script, count any human edit as a touch, and ask for per-field confidence so you can check whether it predicts the errors. Do not let a vendor choose or re-label pages, and do not accept results on their sample in place of yours.
Sent to us, this RFP gets these answers: open-weight models named per document family, a fixed plan with a stated overage, field-level results on your sample before any contract, dedicated EU infrastructure or your own hardware scoped up front, and plain CSV or JSON on the way out. That is the shape of bulk document processing; the stack is on local document models; and if the real question is which workflows should leave frontier APIs at all, local AI consulting is where that starts. Send the 200 pages first — questions 10–13 are worth more than the other 21 combined.