Confidential Memorandum — March 2026
Lemnia: local-first AI for Italian professional studios
Lemnia translates Italian commands into compliance workflows — FatturaPA, FSE 2.0, PCT, PEC — executed entirely on the studio's hardware. No data leaves. No cloud. Law 132/2025 compliant by architecture.
Round: €1.6M–€1.8M · Seed · 50% engineering, 25% go-to-market, 15% accreditations + infrastructure, 10% IP/legal
The Problem
Italy's 119,952 commercialisti, 233,260 avvocati and 50,000+ private medical structures waste 15-25% of staff time on manual, repetitive compliance workflows that require only rule knowledge — not professional judgment.
- 1. Compliance volume is unmanageable
- 2.4 billion FatturaPA invoices transmitted annually via SDI. A studio with 200 clients generates 800 LIPE transmissions/year, hundreds of PCT deposits, thousands of PEC messages. All manual, all repetitive.
- 2. Existing tools don't speak Italian compliance
- UiPath and Power Automate require certified developers — impossible for a 5-15 person studio. TeamSystem and Zucchetti added AI as generic chatbots, not compliance workflow executors. ChatGPT and Copilot don't know FatturaPA schema v1.2.3, the FSE 2.0 gateway, or PCT XML specs.
- 3. Cloud AI is legally excluded from clinical and legal data
- Law 132/2025 mandates Garante pre-notification for any AI system processing Italian clinical data — with blocking power. The Codice Deontologico Forense (Art. 45) sanctions lawyers who upload client files to cloud AI without complete DPA and explicit consent with 1-3 year suspension. The CLOUD Act compels US providers to hand over data on government request regardless of server location.
The Solution
Lemnia is an AI-native RPA replacement for Italian professional studios. Studio staff issues a command in Italian. Lemnia executes the compliance workflow on the studio's hardware, calling government gateways and cloud accounting systems directly, with a full local audit trail.
"Reconcile passive invoices for March and prepare the VAT summary"
→ reads passive invoices from FattureInCloud
→ validates CF/PIVA via AdE
→ matches SDI receipts
→ computes IVA credit/debit
→ generates VAT summary draft
Time: ~8 seconds. Audit trail: local SQLite.
7-stage NLU pipeline, 7 specialized ONNX INT8 models + Qwen3.5-4B (2.74GB), CPU execution. Total: ~4.6GB on disk, ~5GB RAM. No GPU required.
The Regulatory Moat
We didn't choose professional studios because it's a hot market. Regulation chose this architecture for us. Seven overlapping regulatory layers make cloud AI structurally non-compliant for Italian clinical, legal, and fiscal data.
Any AI system processing Italian clinical data must notify the Garante Privacy 30 days before starting. The Garante can issue a blocking order within that window. Requires disclosure of all sub-processors, including US cloud providers.
Cloud vendors: Microsoft, Google, OpenAI must notify before every clinical AI use with unpredictable outcome. The Garante has already demonstrated blocking power: ChatGPT (March 2023), DeepSeek (January 2025).
Regulated professionals (lawyers, doctors, accountants, notaries) must disclose AI tool use to clients in clear, simple, exhaustive language. AI is only permitted as support — human intellectual work must always predominate. First national AI law in the EU.
Cloud vendors: Creates personal legal liability for every professional using cloud AI on client data without explicit consent and complete DPA. CNF already published the disclosure template (October 2025).
AI systems in healthcare are classified as high-risk. Requirements: training dataset governance (Art. 10), automatic event logging with 6+ month retention for healthcare (Art. 12), human oversight with override capability (Art. 14), mandatory conformity assessment before deployment. Penalty: up to €15M or 3% of global turnover.
Cloud vendors: Dataset governance requirements (Art. 10) are incompatible with cloud multi-tenancy: models trained on mixed client data cannot satisfy required traceability. Logging requirements (Art. 12) require per-client verifiable audit infrastructure.
Establishes national Data Access Bodies (DAB) for every EU member state. Secondary use of health data (including AI training) requires explicit authorization from the Italian DAB. Extra-EU transfers of health data for secondary use are prohibited without certified equivalent protection.
Cloud vendors: From 2027, cloud AI vendors cannot use Italian health data for training without explicit DAB authorization. Transfer to US infrastructure: prohibited by default. Permanently closes the door to cloud AI on clinical data.
The EU-US Data Privacy Framework survived its first challenge (September 2025) but Schrems/NOYB announced Schrems III based on Trump administration dismantling of PCLOB oversight and FISA 702 (US cloud server surveillance) reauthorized without EDPB-required reforms. Schrems I (2015) and Schrems II (2020) both invalidated prior frameworks.
Cloud vendors: US cloud vendors operate on a historically unstable transfer framework. Each invalidation created operational chaos for European businesses. European companies processing health or legal data on US cloud operate in a permanent legal gray zone.
Violation of professional secrecy carries 1-3 year suspension from the bar. CNF published AI Guidelines (2025) explicitly prohibiting upload of client files to cloud AI systems without DPA identifying all sub-processors and explicit client consent. The disclosure template per Art. 13 Law 132/2025 is already available.
Cloud vendors: Any lawyer using ChatGPT, Copilot, or any cloud AI on client files without explicit consent and complete DPA risks personal suspension. The risk is individual, not corporate — an immediate deterrent for 233,000 lawyers.
FNOMCeO declared uploading patient data to public AI tools constitutes a violation of medical confidentiality and GDPR. Physicians must verify the AI tool does not use patient data for model training and guarantees reasonable output explainability.
Cloud vendors: OpenAI violated this condition (Garante: Provvedimento n. 755, November 2024 — €15M fine). The only structural guarantee that patient data never enters training is local inference. Microsoft/Google enterprise contracts do not offer this guarantee in a verifiable way.
The window: 18-36 months. Microsoft, Google, and OpenAI could theoretically redesign their architecture for local inference. In practice: re-engineering Azure AI for on-device execution without cloud round-trips requires 2-3 years. SDI accreditation requires 4-12 weeks of mTLS certificate exchange with Agenzia delle Entrate. FSE 2.0 accreditation requires approval of a PR to a ministerial GitHub repository with test plan v8.2.4 (ministero-salute/it-fse-accreditamento). Lemnia is already building these accreditations.
Garante enforcement record
Mar 2023ChatGPT blocked — world's first LLM ban
Nov 2024€15M fine to OpenAI for unauthorized training on Italian user data (Provv. n. 755)
Jan 2025DeepSeek blocked — second AI system banned in Italy in 2 years
2024835 collegial measures, 468 sanctions, €24M total fines
H1 2026Garante inspection plan (Provv. n. 797): AI tools included in inspection scope
Technical Moat
Every component is designed to make the system progressively harder to replicate.
- Italian NLU specialized (8+ months of training)
- 7-stage pipeline with 7 specialized ONNX INT8 models (shared XLM-RoBERTa backbone, ~170MB). Trained on Italian professional compliance data: FatturaPA, FSE 2.0, PCT, PEC, legal and fiscal terminology. No generalist model knows the PCT XSD schema or FSE 2.0 gateway structure.
- Government accreditations (durable barrier)
- SDI/FatturaPA accreditation requires mTLS certificate exchange with Agenzia delle Entrate (4-12 weeks). FSE 2.0 accreditation requires PR approval on the ministerial GitHub with test plan v8.2.4. These processes start during roadmap Phase B — they are permanent entry barriers, not temporary operational advantages.
- Knowledge graph-constrained generation
- The system only generates valid operations against the customer's management systems, constrained by a knowledge graph built from connector OpenAPI specs. Zero hallucinations: entity, endpoint, and parameter references are correct by construction — impossible with a generic LLM.
- Durable execution with HITL
- Step-level persistence (Temporal-style): workflow resumes from last completed step on failure. 5 mandatory HITL gates on mutation operations. Local SQLite audit trail for every operation — required by EU AI Act Art. 12 and EHDS.
- Privacy by architecture
- Credentials in OS keychain. 100% local processing in MODE 1. Zero data transits to Lemnia. Not a company policy — a technical constraint that simultaneously satisfies Law 132/2025, GDPR, Art. 45 CDF, and FNOMCeO.
Market Opportunity
Italian professional studios are the ideal launch market: high concentration of repetitive compliance workflows, regulatory barriers that exclude cloud, an established software market demonstrating willingness-to-pay.
DACH/Europe expansion
The local compliance model replicates in DACH and France with equivalent e-invoicing mandates and GDPR architecture. Estimated European SAM: €400-500M. XLM-RoBERTa models natively support 100+ languages — zero language porting cost.
The FSE 2.0 mandate (October 2025) created a regulatory moment analogous to the 2019 FatturaPA mandate. Every Italian private medical structure must comply. The window for first movers is open now.
Competitive Position
RPA competitors require certified developers. AI competitors are cloud-only and legally excluded from clinical and legal data. Italian management systems offer AI as generic chatbots, not workflow executors.
| UiPath | Power Automate | MS Copilot | Lemnia | |
|---|---|---|---|---|
| Setup | Certified developer | Power Platform | M365 + licenza | 15 min, self-serve |
| Italian NLU | None | Generic | Generic | Compliance-specialized |
| Data processing | Cloud | Azure | Azure | 100% local |
| L. 132/2025 | Non-compliant | Non-compliant | Non-compliant | Compliant by architecture |
| SDI/FatturaPA | Manual scripting | 3rd-party only | Not supported | Native (Phase D) |
| FSE 2.0 | Not supported | Not supported | Not supported | Native (Phase D) |
| PCT | Manual scripting | Not supported | Not supported | Native (Phase D) |
| Cost (5-30 studio) | €10K+/anno | €15/utente/mese | €30/utente/mese | €99–399/mese |
TeamSystem (€1B revenue, 2.5M Italian clients) and Zucchetti (1.5M clients) are the ERP/accounting incumbents but offer no AI compliance workflow automation. Lemnia is complementary to existing management systems, not a replacement.
Business Model
Fixed monthly fee per studio, zero consumption costs. No GPU costs in production — gross margin improves with scale.
Pricing
1-5 staff
5-15 staff
15-30 staff + MODE 3 on-prem
Annual billing: 2 months free. One-time onboarding: €500-1,500.
Unit Economics
Cost structure
- No GPU costs in production (local execution)
- Cloud fallback <10% of queries (Tier 2 Qwen3.5-4B)
- Low churn: compliance software has high switching cost
- Gross margin improves with every new studio acquired
Traction & Milestones
Full NLU pipeline, specialized models trained, KG-constrained workflow generation operational. FattureInCloud, WooCommerce, PrestaShop, Mago4 connectors working. First pilots in conversation phase.
Current status: <1s end-to-end latency, >90% queries processed locally, 3 target verticals, 4 active connectors.
Months 0-6
Validation & first 10 paying studios
- 10 paying studios (at least 1 per vertical)
- SDI + FSE 2.0 accreditation process started (parallel to dev)
- Datev Koinos + Aruba Fatturazione + InfoCert PEC connectors
- AdE CF/PIVA validation (universal entity resolution)
- Engineering: 3 hires
Months 6-12
50 studios · €25K MRR · SDI + FSE 2.0 accredited
- SDI/FatturaPA accreditation completed
- FSE 2.0 accreditation completed (ministero-salute/it-fse-accreditamento)
- SDI + FSE 2.0 + Firma digitale + PCT busta connectors
- Self-serve onboarding for commercialisti
- Go-to-market: 2 hires
Months 12-18
100+ studios · €60K MRR · Series A
- Sistema Tessera Sanitaria (600M prescriptions/year)
- DACH expansion (equivalent regulatory framework)
- Series A closed
The Ask
Seed round of €1.6M–€1.8M to complete Italian compliance connectors, acquire the first 50 paying studios, and complete SDI and FSE 2.0 accreditations.
Use of funds
Why now
FSE 2.0 mandate active. All Italian private medical structures must transmit to FSE 2.0 from October 2025. Every day without a solution is direct cost for the studio. The window for first movers is open now.
Law 132/2025 built the moat. Cloud AI cannot legally process Italian clinical data without Garante notification and risk of blocking. Lemnia's local architecture is not a product choice — it's the only architecture that simultaneously satisfies all seven regulatory layers.
Technical convergence. Qwen3.5-4B (2.74GB) on CPU via llama.cpp makes enterprise-grade AI inference on consumer hardware without GPU possible for the first time. This didn't exist 18 months ago.
Team
Experience from Google, Nike, and Meta in machine learning, systems engineering, and B2B go-to-market.
Core competencies
- Italian NLU specialized: fiscal, legal, clinical compliance
- Distributed systems & edge architecture (local inference)
- Privacy-first system design
- Direct B2B sales to Italian professional studios
Technical milestones reached
- 7 specialized ONNX INT8 models trained
- 7-stage NLU pipeline operational (<1s latency)
- 4 connectors operational (FattureInCloud, WooCommerce, PrestaShop, Mago4)
- Tauri desktop app with Rust gRPC sidecar
For the full pitch deck, detailed metrics, or a live demo.
investors@lemnia.app