A purpose-built reasoning engine for Permanent Establishment risks. We started with the legal reasoning — the AI only executes it.
Pilot conversations under way with Benelux advisory firms and Fortune 500 in-house tax teams. Named references will be added as each pilot confirms public attribution.
Cross-border activity, remote work and digital business models keep producing new taxable footprints. Tax teams are forced to choose between two flawed options — neither of which scales.
Days or weeks of senior hours per question. Deep analysis but margin-thin for the firm, prohibitively expensive for the client, impossible to scale across a portfolio.
A one-line answer with no source trail and no defensibility. The exposure does not disappear — it surfaces years later, in a tax audit, with penalties and profit-attribution adjustments attached.
Days of legal analysis. Minutes of execution. Every memo grounded in OECD Commentary, doctrine, treaty text and case law — with sequential footnotes a partner, a tax director or an inspector can verify.
Three structural realities make PE work uniquely difficult — and uniquely poorly served by both traditional advisory and broad AI tools.
Every PE outcome turns on fact patterns mapped across treaty text, OECD Commentary, doctrine, case law and country-specific guidance. No two cases reach the conclusion via the same path — until the reasoning itself is encoded.
Operational decisions taken today surface in tax audits years later — when commercial models have moved on and exposures are no longer reversible. By the time the issue is visible, the cost is already real.
General-purpose models save 50–60% of consultant time, but produce inconsistent reasoning, invented citations and no audit trail. They look credible. They are not defensible. The heavy lifting — and the liability — stays with you.
We did not start with a large language model and bolt legal context onto it. We started with the legal reasoning itself — and only then added an AI execution layer on top.
Authored by senior PE practitioners and academics. Mirrors the analytical path an experienced specialist actually walks — step by step, condition by condition.
Each node carries explicit instructions: how the condition is understood, which nuances matter, which fact patterns are decisive. The engine never improvises interpretation.
Every node points to specific sources — treaty articles, OECD Commentary, doctrine, case law, rulings. Precise references for a precise legal question.
A controlled language model walks the predefined path, applies the mapped sources, and articulates the conclusion in clean professional output. The same answer, every time.
The legal reasoning sits in one tier. The AI sits in another. That separation is what makes the output dependable.
Decision structure, interpretive layer, source mapping. Authored and maintained by senior PE specialists. The legal reasoning lives here.
A controlled LLM articulates the conclusion in clean professional language. It cannot invent legal interpretation — only execute one.
→ Two layers. One guarantee: the engine cannot invent answers.
Every exPErtus analysis produces a structured legal memorandum — sequential footnotes, risk classification, do's and don'ts, recommended next steps. The same shape a senior advisor would hand you, at a fraction of the cost.
exPErtus covers the Permanent Establishment concepts that dominate international tax practice, with dedicated reasoning paths for each PE archetype.
Horizontal AI tools, tax governance platforms and internal AI initiatives each solve a slice of the problem. None of them deliver complete, source-referenced PE analysis end-to-end.
exPErtus is built by a combination of senior international tax practice, legal academia, and deep-tech delivery experience — a lockstep expertise combination that is rare in the market.
15 years of international tax experience with large (a.o. Big 4) consulting firms. Leads commercial, client relationships, and go-to-market.
30 years of PE expertise. Tax Partner and Professor at the University of Antwerp. Architect of the proprietary reasoning framework.
Second-time founder with a background as a professional researcher in AI. Previously co-founded and successfully sold a service-based AI firm. Leads end-to-end technical vision and delivery.
Work from the team on how the reasoning engine is built, how source material maps onto practice, and where the PE discipline is heading. We'll post here as we go.
Professional PE work needs the same answer every time, with every step traceable to a source. We walk through why a structured reasoning layer — not a bigger prompt — is what turns AI output into something a partner, an in-house tax director or an auditor can actually rely on.
The 2025 OECD Commentary update sharpens the tests for home office PE — "at the disposal", business connection, the role of employer instruction. We compare the new framing with how home office cases have historically been analysed, and unpack what genuinely changes for cross-border employers today.
A close read of the BMF draft guidance of 13 February 2025 and the positions German tax authorities are taking in practice — what they consider a home office PE, where they push back, and what this means for Belgian, Dutch and other EU employers with staff working from Germany.
Designed from the ground up for GDPR and the EU AI Act. Your data stays in the EU, under controlled access, and is never used to train models.
EU-hosted infrastructure · assessment data and training data strictly separated · sequential footnotes on every output.
A PE is the international-tax concept that defines when a cross-border company creates a taxable presence in another country. Once a PE exists, the consequences are real: corporate tax registration, profit attribution, payroll obligations and VAT implications — often surfacing years later in a tax audit, when corrective action is no longer possible. PE analysis is inherently fact-driven, treaty-specific and locally interpreted, which is exactly why it absorbs so much senior advisory time. See the full explainer with worked example →
They all share an architecture that exPErtus does not. ChatGPT and Claude generate text from probabilistic patterns. Specialised legal-AI platforms and tax-specific tools sit on top of a frontier LLM with retrieval-augmented generation (RAG) over a curated legal corpus — the LLM decides per query what to retrieve and how to synthesize it. That works for fast research, but two structural problems remain for high-stakes PE work: the reasoning path varies between runs, outputs contain errors and are rarely fully complete, and sources are too often hallucinated. See the full architectural comparison →
exPErtus inverts the architecture. The reasoning path for every PE question is predefined in a proprietary decision framework, authored by senior PE specialists, with each node pre-mapped to its authoritative sources. The LLM only executes that path — it never decides what to retrieve, interpret or cite. The result: deterministic reasoning, bounded sources, and a complete audit trail.
A structured legal memorandum: introduction and scope, our understanding of the facts, the fiscal query, the applicable legal framework, condition-by-condition application to the facts, and a conclusion with risk classification, do's and don'ts and recommended next steps. Every reference is footnoted to the specific OECD Commentary paragraph, treaty article, doctrine or case law it relies on. The same memo can be re-rendered as a high-level email advice, a ruling request or a defence file.
By architecture, not by guardrails. Generic AI is well-known for fabricating citations — Mata v. Avianca set the precedent (a New York lawyer was sanctioned for filing a brief with non-existent cases generated by ChatGPT). Even purpose-built legal RAG tools still hallucinate 17–33% of the time according to Magesh et al. (2025), Journal of Empirical Legal Studies, because the LLM still chooses what to retrieve and synthesize per query. exPErtus removes that risk at the source: every citation in our output is predefined at the decision node where it is used. The LLM cannot discover, invent, or fabricate sources — they are baked in. If a source appears in a memo, it exists in the framework, has been verified by senior PE specialists, and applies to that specific legal question.
Two interaction modes. A structured chat interface mirrors how advisors already gather facts: it identifies what's needed, asks targeted questions, and produces the analysis. Or an email-based mode: the advisor forwards a client email to exPErtus, the platform extracts facts and sends targeted follow-ups to the client (on the advisor's behalf), and delivers the final memo to the advisor for review and approval before it reaches the client. The advisor stays in control throughout — only the speed and consistency change.
No. It transforms what they can do. For advisory firms, exPErtus dramatically reduces lead time and cost per analysis — freeing senior hours for higher-value work like planning, structuring, audit defence and client relationships. The trusted advisor relationship is preserved; the engine simply makes the work faster, more consistent and more profitable.
Yes. Every output follows orthodox PE reasoning methodology and carries a complete audit trail. Each conclusion is traceable to the specific treaty provision, OECD Commentary paragraph, case law or doctrine that supports it. The output is suitable for internal decision-making, sharing with external advisors and defending positions during a tax audit.
The reasoning engine is modular and source-mapped, so legal changes are integrated through a structured maintenance process overseen by our expert team. exPErtus has a structural collaboration with a world's leading independent provider of cross-border tax expertise. Treaty changes, OECD Commentary updates, relevant case law and legislative developments are systematically channelled and pre-processed — the engine always operates on the most current authoritative sources.
Data is processed in a controlled, EU-hosted environment. Client inputs are never used to train models — training data and assessment data are strictly separated. exPErtus is designed from the ground up for GDPR (data minimisation, purpose limitation, storage limitation) and aligned with the EU AI Act's transparency and human-in-the-loop requirements. ISO 27001 certification is in progress.
The MVP is ready end of May 2026. A limited number of pilot slots open in June 2026. Commercial launch is Q4 2026. Join the waitlist below to be contacted about a pilot or early-access engagement.
We're opening a limited number of pilot slots ahead of commercial launch. Email us to start the conversation — no mass mailings, no drip campaigns.