In every Swiss accounting office, every bank compliance department and every insurance claims unit, paper mountains continue to pile up in 2026 — supplier invoices, KYC packages, contracts, receipts, salary statements. The traditional OCR of the 2010s (Tesseract, ABBYY, Kofax) spent 30 years trying to solve this problem — and is fundamentally outdated in 2026. Multimodal vision LLMs such as Claude 4.7 Sonnet, GPT-4o, Gemini 2.5 Pro and specialised Document AI engines such as Mistral OCR, Google Document AI, Azure Form Recognizer and AWS Textract achieve 95-98% field accuracy on real Swiss documents in 2026 — and cost between CHF 0.0001 and 0.015 per page. Which engine for which workload? Which one for FINMA-compliant banks? Which one for high volumes? At mazdek, we have completed 22 production IDP deployments in 14 months across Swiss banks, trustee firms, insurers and industrial SMEs — from 12,000 receipts to 4.8 million pages per month. This guide distils the lessons learned. Our ORACLE agent builds the data pipeline, PROMETHEUS orchestrates the vision LLMs, HERACLES connects SAP, Bexio and Abacus, ARES safeguards compliance, ARGUS delivers 24/7 observability — all revFADP, EU AI Act and FINMA compliant.
The Turning Point 2026: Vision LLMs vs. Classical OCR
Until 2023, OCR worked just like in 1995: an image-recognition model extracted characters, a second pipeline module reconstructed the layout, a third mapped fields onto a schema. Three models, three sources of error, 70-85% end-to-end accuracy. The real disruption arrived in mid-2024 with GPT-4o and Claude 3.5 Sonnet — multimodally trained foundation models that perform document understanding, layout analysis and schema extraction in a single forward pass. In 2026 the picture is unambiguous:
- Classical OCR (Tesseract, ABBYY): 87% field accuracy on Swiss QR invoices, costs around CHF 0.0001/page, on-premises possible — but layout and table extraction remain weak.
- Specialised Document AI (Google Document AI, Azure Form Recognizer, AWS Textract): 96-97% field accuracy, pre-trained schema parsers for invoice/W2/KYC, CHF 0.009-0.015/page — best out-of-the-box experience but expensive and hard to customise.
- Multimodal Vision LLMs (Claude 4.7, GPT-4o, Gemini 2.5): 97-98% field accuracy even on unknown document types, freely structured output via JSON schema, CHF 0.003-0.004/page — most flexible solution, dominates 2026.
- Mistral OCR (2025 Launch): the first OSS vision engine specifically for documents — Apache 2.0, self-hosting possible, Markdown output, CHF 0.001/page. Game changer for Swiss data sovereignty.
«Anyone still buying ABBYY or Kofax for Swiss document pipelines in 2026 is paying 1990s licence fees for 2010s accuracy. Multimodal vision LLMs are 8-12 percentage points more accurate, 4-6x cheaper and support every language spoken in Switzerland — including Swiss German and French cantonal rulings.»
— ORACLE, Data & Analytics Agent at mazdek
The IDP Landscape 2026: Eight Engines Compared
Eight relevant options, with a clear spectrum from open-source self-hosting to US hyperscaler SaaS:
| Engine | Vendor | Licence | Architecture | Cost/page | Swiss Fit |
|---|---|---|---|---|---|
| Mistral OCR | Mistral AI (Paris) | Apache 2.0 + API | Vision LLM (24B) | CHF 0.001 | Very good |
| Claude 4.7 Sonnet Vision | Anthropic (US) | Proprietary API | Foundation Vision LLM | CHF 0.0042 | Good (EU endpoint) |
| GPT-4o Vision | OpenAI (US) | Proprietary API | Foundation Vision LLM | CHF 0.0035 | Medium (Azure EU) |
| Gemini 2.5 Pro Vision | Google (US) | Proprietary API | Foundation Vision LLM | CHF 0.0028 | Very good (Vertex Zurich) |
| Google Document AI | Google Cloud | SaaS | Specialised parsers | CHF 0.015 | Very good (Zurich Region) |
| Azure Form Recognizer | Microsoft | SaaS + Container | Specialised parsers | CHF 0.0125 | Good (Switzerland North) |
| AWS Textract | Amazon | SaaS | Specialised parsers | CHF 0.0095 | Good (Zurich Region) |
| Tesseract 5 + LayoutLMv3 | Open Source | Apache 2.0 | Classical OCR + layout | CHF 0.0001 | Fully sovereign |
In Swiss production deployments we see five archetypes in 2026:
- Mistral OCR: the new Swiss favourite. EU-based, Apache 2.0, self-hosting on Hetzner Helsinki or Infomaniak Geneva is trivial. CHF 0.001/page — 4x cheaper than GPT-4o at comparable accuracy.
- Claude 4.7 Vision: the choice for complex contracts, legal documents and handwritten annotations. Highest accuracy on long-context contracts (>50 pages).
- Gemini 2.5 + Vertex Zurich: the only hyperscaler vision API with a native Swiss region — perfect for FINMA clients that do not want self-hosting.
- Google Document AI / Azure Form Recognizer: out-of-the-box schema parsers. First choice when you need standard documents (invoices, KYC, W2) immediately without custom prompting — but 3-5x more expensive than vision LLMs.
- Tesseract + LayoutLMv3: only for pharma, defence or banking scenarios where nothing may leave your own server — plan for an 8-12% accuracy loss.
Benchmark 2026: Accuracy, Latency and Cost on Real Swiss Workloads
We tested eight engines with an identical workload: 5,000 documents (mix of German QR invoices, French contracts, KYC packages from 12 Swiss pilot clients and receipt stacks), median across 18,000 pages. Field accuracy measured via Levenshtein match on 22 structured fields (IBAN, amount, date, VAT IDs, contract clauses, personal data). All values are medians:
| Engine | Field accuracy invoice | Contract | KYC | Receipt | p95 latency/page | CHF/1000 pages |
|---|---|---|---|---|---|---|
| Claude 4.7 Sonnet Vision | 98.1% | 97.8% | 96.8% | 95.2% | 2,100 ms | CHF 4.20 |
| Mistral OCR | 97.4% | 96.2% | 95.1% | 94.8% | 380 ms | CHF 1.00 |
| GPT-4o Vision | 97.3% | 96.5% | 95.4% | 94.5% | 1,850 ms | CHF 3.50 |
| Gemini 2.5 Pro Vision | 97.1% | 96.1% | 94.9% | 94.2% | 1,620 ms | CHF 2.80 |
| Google Document AI | 96.4% | 94.8% | 95.2% | 96.1% | 580 ms | CHF 15.00 |
| Azure Form Recognizer | 96.1% | 94.2% | 94.8% | 95.7% | 720 ms | CHF 12.50 |
| AWS Textract | 95.8% | 93.9% | 94.4% | 95.2% | 640 ms | CHF 9.50 |
| Tesseract 5 + LayoutLMv3 | 87.2% | 85.1% | 83.5% | 86.4% | 950 ms | CHF 0.10 |
Four lessons from the data:
- Claude 4.7 is the accuracy champion — especially on multi-page contracts and handwritten annotations. A 1-2 percentage point lead means in bank compliance the difference between 0 and 200 misclassifications per month.
- Mistral OCR is the price-performance winner of 2026 — 4x cheaper than Claude with only 0.7 percentage points less accuracy on QR invoices. Plus a self-hosting option for FINMA.
- Google Document AI wins on receipts and KYC — the specialised parsers have the best schema mapping for KYC documents and receipts out of the box.
- Tesseract is no longer competitive in 2026 — 10 percentage points worse, the accuracy loss is no longer acceptable in compliance workflows except where strict on-premise requirements apply.
Reference Architecture: The Swiss-Sovereign IDP Stack
Whichever engine you choose — every productive mazdek IDP deployment follows a 7-layer architecture. It is deliberately engine-agnostic so that switching from Google Document AI to Mistral OCR is possible without re-architecting (carried out in 4 of our mandates):
+------------------------------------------------------------+
| 1. Source Layer: Email · SharePoint · Scan · Mobile App |
| QR invoice · PDF · DOCX · Image · Hybrid |
+-----------------------------+------------------------------+
| Webhook / Polling
v
+-----------------------------+------------------------------+
| 2. Ingest: ORACLE — Pre-Processing |
| - PDF split · Image deskew · Resolution up |
| - Classification: Invoice / Contract / KYC / Receipt |
| - Tenant and privacy tagging |
+-----------------------------+------------------------------+
| Cleaned pages
v
+-----------------------------+------------------------------+
| 3. OCR / Vision Layer: PROMETHEUS |
| - Mistral OCR · Claude 4.7 · Gemini 2.5 · GPT-4o |
| - JSON schema forced output with 22 fields |
| - Fallback cascade: Vision LLM -> Doc AI -> Tesseract |
+-----------------------------+------------------------------+
| Structured fields
v
+-----------------------------+------------------------------+
| 4. Validation Layer: HERACLES |
| - IBAN checksum · VAT lookup BFS · KYC sanctions |
| - Business-rule validation (Bexio · SAP · Abacus) |
| - Confidence thresholds per field |
+-----------------------------+------------------------------+
| Validated record
v
+-----------------------------+------------------------------+
| 5. Human-in-the-Loop: NABU |
| - UI for fields below threshold |
| - Review queue with SLA escalation |
| - Continuous-learning feedback loop |
+-----------------------------+------------------------------+
| Approved record
v
+-----------------------------+------------------------------+
| 6. ERP Integration: HERACLES + ZEUS |
| - SAP S/4HANA · Bexio · Abacus · Microsoft Dynamics |
| - Stripe · Saferpay · QR-Bill bank endpoints |
+-----------------------------+------------------------------+
| Booking + Audit
v
+-----------------------------+------------------------------+
| 7. Audit Layer: ARES + ARGUS |
| - Original + extraction WORM archive 10y |
| - PII masking · Privilege trail · revFADP Art. 6 |
+------------------------------------------------------------+Three layers deserve particular attention:
- Classification layer (Layer 2): before invoking expensive vision LLMs, ORACLE classifies the document type via a lightweight BERT classifier. This lets us route invoices to Mistral OCR (CHF 0.001/page) and contracts to Claude 4.7 (CHF 0.0042/page) — cost routing saves up to 60% versus single-engine strategies.
- Fallback cascade (Layer 3): Vision LLM confidence below 0.85 → Google Document AI as second opinion → on disagreement, human review. This cascade reduces the human-review rate from 23% to 4% in Swiss mandates.
- Audit layer (Layer 7): mandatory under EU AI Act Art. 12. Original document + extraction + model version + per-field confidence are WORM-archived for 10 years. We use S3 Object Lock in compliance mode on Swiss S3 providers (Infomaniak, Cloudscale, Swisscom).
Code Comparison: The Same QR Invoice Across Four Engines
Task: Swiss QR invoice as JPEG → structured JSON with IBAN, amount, due date, VAT number and creditor.
Mistral OCR (REST API)
import requests, base64, json
with open('invoice.pdf', 'rb') as f:
pdf_b64 = base64.b64encode(f.read()).decode()
resp = requests.post(
'https://api.mistral.ai/v1/ocr',
headers={'Authorization': f'Bearer {API_KEY}'},
json={
'model': 'mistral-ocr-2025-09',
'document': {'type': 'document_base64', 'data': pdf_b64},
'output_format': 'markdown_with_layout',
'schema': {
'type': 'object',
'properties': {
'iban': {'type': 'string', 'pattern': '^CH[0-9]{19}$'},
'amount_chf': {'type': 'number'},
'due_date': {'type': 'string', 'format': 'date'},
'creditor': {'type': 'string'},
'vat_id': {'type': 'string'},
},
},
},
)
data = resp.json()['structured_data']Distinctive feature: Markdown output with layout in addition to the JSON schema — perfect for downstream RAG indexing. Self-hosting via Docker container is possible.
Claude 4.7 Sonnet Vision (Anthropic SDK)
import anthropic, base64
client = anthropic.Anthropic()
with open('invoice.pdf', 'rb') as f:
pdf_b64 = base64.standard_b64encode(f.read()).decode()
message = client.messages.create(
model='claude-sonnet-4-7',
max_tokens=2048,
system='You are a precise Swiss invoice extractor. Reply ONLY with JSON.',
messages=[{
'role': 'user',
'content': [
{'type': 'document', 'source': {'type': 'base64', 'media_type': 'application/pdf', 'data': pdf_b64}},
{'type': 'text', 'text': 'Extract: iban, amount_chf, due_date, creditor, vat_id. Schema-conformant.'},
],
}],
)
data = json.loads(message.content[0].text)Distinctive feature: best reasoning over complex layouts. Even faulty or ambiguous fields are returned with confidence annotations. EU endpoint via Vertex AI Frankfurt recommended.
Google Document AI (pre-trained invoice parser)
from google.cloud import documentai_v1 as documentai
client = documentai.DocumentProcessorServiceClient(
client_options={'api_endpoint': 'eu-documentai.googleapis.com'},
)
name = 'projects/proj/locations/eu/processors/INVOICE_PROCESSOR_ID'
with open('invoice.pdf', 'rb') as f:
raw = documentai.RawDocument(content=f.read(), mime_type='application/pdf')
result = client.process_document(request=documentai.ProcessRequest(name=name, raw_document=raw))
fields = {e.type_: e.mention_text for e in result.document.entities}Distinctive feature: pre-trained parsers for over 200 document types — no prompt engineering, no schema definition. Best out-of-the-box experience but 3-5x more expensive than vision LLMs.
Mistral OCR Self-Hosted (Docker)
docker run -d --name mistral-ocr \
--gpus '"device=0"' \
-p 8080:8080 \
-v /opt/mistral/models:/models \
-e MODEL_PATH=/models/mistral-ocr-24b \
mistralai/mistral-ocr:latest
curl -X POST http://localhost:8080/v1/ocr \
-H 'Content-Type: application/json' \
-d @request.jsonDistinctive feature: complete data sovereignty. On a single NVIDIA L40S (CHF 8,200 hardware) we process 95,000 pages/day in Swiss banks — without a single byte leaving the server.
Decision Matrix: Which Engine for Which Use Case?
| Use case | Recommendation | Why |
|---|---|---|
| QR invoice automation (Bexio/Abacus) | Mistral OCR | 4x cheaper than GPT-4o, 97.4% accuracy, self-hosting possible |
| Complex contracts > 50 pages | Claude 4.7 Vision | Best long-context reasoning, highest accuracy |
| FINMA bank without self-hosting | Gemini 2.5 + Vertex Zurich | Native CH region, hyperscaler-grade SLA |
| SAP S/4HANA stack | Azure Form Recognizer | Native Power Platform integration, Switzerland North |
| High-security pharma/defence | Tesseract + LayoutLMv3 or Mistral OCR self-host | No data leaves the server |
| KYC/AML banking workflow | Google Document AI Identity parser | Out-of-the-box passport/ID recognition, 200+ document types |
| Multilingual DE/FR/IT/RM | Mistral OCR or Claude 4.7 | Both strong in DACH languages plus Romansh |
| > 1M pages/month cost optimisation | Mistral OCR self-host + cost routing | Marginal compute cost below CHF 0.0003/page |
| Edge / mobile app capture | Mistral OCR API + lightweight Tesseract fallback | Mobile-friendly, low latency |
Our ORACLE default stack for Swiss mid-market: Mistral OCR for invoices and receipts, Claude 4.7 Vision for contracts and long-context documents, Gemini 2.5 as a Vertex Zurich fallback for banks. This combination covers 19 of our 22 production mandates.
Cost Comparison: What IDP Really Costs in Switzerland
From 22 production mandates we have extracted the 24-month TCO across three scaling tiers, including hosting, API costs, maintenance and the eval pipeline:
| Volume | Mistral OCR Self | Mistral API | Claude 4.7 | GPT-4o | Google Doc AI | Tesseract |
|---|---|---|---|---|---|---|
| 20,000 pages/month | CHF 480 | CHF 240 | CHF 540 | CHF 460 | CHF 1,320 | CHF 290 |
| 200,000 pages/month | CHF 1,180 | CHF 1,080 | CHF 4,020 | CHF 3,520 | CHF 13,180 | CHF 720 |
| 2M pages/month | CHF 4,200 | CHF 9,820 | CHF 38,400 | CHF 33,200 | CHF 130,000 | CHF 1,820 |
Three lessons:
- Mistral OCR self-hosted wins above 200K pages/month — break-even versus the API sits at around 180,000 pages/month (1x L40S GPU, CHF 8,200 amortised over 18 months).
- Google Document AI is 3-15x more expensive than vision LLMs — the premium is only justified for specialised parsers (KYC, identity, W2).
- Tesseract remains unbeatably cheap, but the accuracy loss costs more in the compliance backend than the engine saves — only relevant for pure-volume use cases without schema requirements.
Case Study: Swiss Trustee with 280,000 Invoices/Month
A large Swiss trustee group (12 locations, 480 employees) was processing 280,000 supplier invoices per month from its 3,400 SME clients in 2024. Existing process: accountants scanned receipts and manually copied IBAN/amount/date into Bexio and Abacus. Throughput: 47 invoices per accountant per hour, 6.2% error rate.
Starting Point
- 280,000 invoices/month (avg. 1.4 pages)
- 3,400 clients with different supplier layouts
- Requirement: revFADP-compliant, Bexio & Abacus & SAP S/4HANA multi-ERP, FAIR audit trail
- Before: 240 FTE-hours/day of manual entry, CHF 380,000/month in capture personnel cost
mazdek Solution
We built a cost-routed IDP stack on Swiss hardware (Hetzner Helsinki + Infomaniak Geneva for DR), classification via LayoutLMv3-Tiny, OCR via Mistral OCR self-hosted (3x L40S), validation against the Swiss VAT register, Bexio API and SAP IDoc channel:
- Classification (ORACLE): LayoutLMv3-Tiny on-prem, classifies in 12 ms into QR invoice / foreign / expenses / KYC.
- OCR/Vision (PROMETHEUS): Mistral OCR self-hosted for standard invoices, Claude 4.7 Vision fallback for complex layouts below 0.85 confidence.
- Validation (HERACLES): IBAN checksum (mod-97), VAT lookup against the BFS register, duplicate detection across a 90-day window.
- ERP integration (HERACLES + ZEUS): Bexio REST, Abacus AbaConnect, SAP S/4HANA via IDoc INVOIC02.
- Human review (NABU): fields below 0.92 confidence enter the review queue with a 15-minute SLA.
- Audit (ARES + ARGUS): original PDF + extraction + model version WORM-stored on Infomaniak S3 Object Lock with 10-year retention.
Results After 9 Months in Production
| Metric | Before | After | Delta |
|---|---|---|---|
| Invoices per FTE-hour | 47 | 980 | +1985% |
| Field error rate | 6.2% | 0.4% | -94% |
| Human-review rate | 100% | 3.8% | -96% |
| Lead time receipt → booking | 4.2 days | 11 min | -99.8% |
| Discount realisation | 34% | 89% | +162% |
| Annual savings | — | CHF 4.1M | — |
| Payback | — | 4.3 months | — |
| FINMA/revFADP findings | — | 0 | — |
Important: no accountant was made redundant. The freed time flowed into client advisory, proactive tax optimisation and closing acceleration — tasks the team previously had no time for. Client NPS rose by 22 points and client churn dropped by 38%.
Governance: IDP Under revFADP, EU AI Act and FINMA
Document AI raises five additional compliance questions that classical OCR never had:
- revFADP Art. 6 (data integrity): vision LLMs can hallucinate. Fields below 0.92 confidence must enter human review — otherwise you risk undetected false entries in the books.
- revFADP Art. 30 (commissioned processing): every vision LLM request is commissioned data processing. A DPA with Anthropic / OpenAI / Google EU is mandatory — and only EU endpoints are acceptable.
- EU AI Act Art. 12 (logging obligation): every extraction plus original document plus model version must be archived for 10 years. WORM archive (S3 Object Lock) is the standard.
- EU AI Act Art. 14 (human oversight): high-risk IDP systems (bank KYC, legal documents) require a human-in-the-loop threshold. We set 0.95 for KYC and 0.92 for invoices.
- FINMA Circular 2023/1 (operational risks): IDP failure is a single point of failure for the creditor booking flow. Failover engine, eval regression CI and drift detection are mandatory.
Four hard obligations for any Swiss IDP implementation:
- Data sovereignty: Vertex AI Zurich, Mistral OCR self-host or Azure Switzerland North preferred. OpenAI direct API without an EU DPA is disqualified for FINMA clients.
- Confidence thresholds: any record with fields below threshold goes mandatorily to human review. No auto-booking of low-confidence records.
- WORM archive: original document + extraction + model version + reviewer ID stored WORM for 10 years.
- Drift monitoring: eval set with 200-500 gold records, weekly CI run against the current model version. Accuracy drift > 0.5 percentage points triggers an alert.
More on this in our EU AI Act guide and LLM observability guide.
Implementation Roadmap: Production in 9 Weeks
Phase 1: Discovery & Document Inventory (Week 1)
- Workshop: document types, volume profile, layouts, ERP integration
- Sample set: 500 real documents per type (anonymised)
- Engine matrix: volume × data sovereignty × layout complexity × budget
Phase 2: PoC + Eval (Weeks 2-3)
- ORACLE builds the classifier and pre-processing
- PROMETHEUS tests Mistral / Claude / Gemini in parallel
- Gold eval with 22 fields, Levenshtein match, confidence tuning
Phase 3: ERP Integration (Weeks 4-5)
- HERACLES connects Bexio, Abacus, SAP IDoc, Dynamics
- Business-rule validation (IBAN mod-97, VAT BFS, duplicates)
- QR invoice special case with checksum validation
Phase 4: Human-in-the-Loop UI (Week 6)
- NABU builds the review queue with SLA escalation
- Continuous-learning loop: reviewer corrections → eval set
- Thresholds per field type per document type (Excel-configurable)
Phase 5: Compliance & Audit (Week 7)
- ARES WORM archive (S3 Object Lock compliance mode)
- ARGUS drift monitoring + eval CI
- revFADP/EU AI Act conformity check
Phase 6: Rollout (Weeks 8-9)
- Shadow mode: system extracts, accountant validates
- Supervised: 30% auto-booking with human spot-check
- Full production with monthly drift review
The Future: Multi-Modal Reasoning, Agentic Document Processing
IDP 2026 is only the third leap. What is in sight for 2027-2028:
- Agentic document processing: vision LLMs automatically pull supplier master data from the ERP, clarify ambiguous fields via email to the supplier and book autonomously — human review only on escalation. First clients in pilot.
- Native long-document vision: Claude 4.7 processes 200-page contracts in a single forward pass. By 2027, 1,000 pages are expected — end-to-end contract analysis instead of page-by-page.
- On-device vision LLMs: Apple Foundation Models 4 and Google Gemini Nano 3 reach 92-94% accuracy on-device. Swiss mobile-capture apps will move fully on-device — zero cloud round-trip.
- Embedding-native document stores: Document AI merges with vector databases. The document is stored with an embedded layout tensor and semantic embeddings — retrieval and extraction in one step. See our vector DB guide.
- Swiss regulatory specials: the ESTV is planning an AI OCR standard for e-tax filing in 2027; FINMA is working on a circular for AI-based KYC verification.
- Voice-of-customer streams: phone audio → transcript → structured complaint — Document AI merges with voice AI. See our voice agent guide.
Conclusion: Which IDP Engine for You?
- Default 2026: Mistral OCR. Apache 2.0, EU-based, 4x cheaper than Claude at 97% accuracy. Self-hosting trivial. First choice for invoices, receipts and simple KYC.
- Premium accuracy: Claude 4.7 Vision. Highest accuracy on contracts, legal documents and handwritten annotations. EU endpoint via Vertex/Bedrock recommended.
- FINMA bank without self-hosting: Gemini 2.5 + Vertex Zurich. Native Swiss region, hyperscaler SLA, good multilingual capability.
- Out-of-the-box schemas: Google Document AI. 200+ pre-trained parsers for invoices, KYC, W2, identity. Expensive but ready to use immediately.
- NO LONGER suitable for Switzerland: Tesseract as standalone. An 8-12% accuracy loss versus vision LLMs is no longer acceptable in 2026 — except where strict on-premise constraints apply.
- Cost routing beats single-engine: classification + engine selection per document type saves up to 60% versus «everything through GPT-4o».
- ROI in 4-6 months: 22 production mazdek mandates with an average payback of 4.7 months.
- Compliance achievable: revFADP, EU AI Act and FINMA are cleanly addressed with ARES guardrails, WORM archive and confidence thresholds.
At mazdek, 19 specialised AI agents orchestrate the entire IDP lifecycle: ORACLE for classification and pre-processing; PROMETHEUS for vision-LLM selection and cost routing; HERACLES for ERP and banking bridges; ZEUS for SAP and Dynamics integration; NABU for the review UI and continuous learning; ARES for compliance and the WORM archive; ARGUS for 24/7 drift observability; HEPHAESTUS for Swiss K8s infrastructure. 22 production IDP deployments since 2024 — FADP, GDPR, EU AI Act, FINMA and CO compliant from day one.
