Precept

A precept is a rule that governs action.

Sovereign AI strategy and governed AI engineering

Decisions under human authority, data and models under your control.

Practice

Three areas, one line of delivery: from strategy to systems in production.

AI strategy & roadmaps

Identify and prioritise AI value cases on measurable value. Set pilot-to-production roadmaps with executive leadership, and own delivery through to production.

Data foundations

Audit data pipelines and data quality. Re-engineer technical debt and corrupt data streams into a reliable foundation for analytics and production AI.

Governed AI & agents in production

Design, build and deploy auditable AI and agents: bounded LLMs paired with deterministic rules, with audit trails built for risk and compliance sign-off. Delivery owned end to end.

Principles

Principles

Data first

Data foundations are engineered first, not retrofitted.

Governed AI cannot rest on pipelines that cannot be trusted.

Governance by design

Every output is constrained, auditable and reconstructible.

Language models operate inside deterministic architecture: they draft, extract and summarise. Rules, scores and decisions stay in code and under human authority.

Confidentiality

Sensitive data stays within a perimeter you control.

Where data is processed is an architectural decision, not a vendor default. Indexing and retrieval run locally; what reaches an external model is explicit, limited and confirmed.

Production delivery

The measure is a system running in production.

Operating protocols, privacy and security standards are in place from the start, not after the pilot.

End-to-end ownership

Delivery is owned end to end, from strategy to production.

Strategy, design and delivery are not handed off between parties.

Grounded in applied research

The approach is formalised in applied research.

Precept Research studies governed AI in regulated domains, with applications in investment decision-making and case-law research. It led to a fully documented reference architecture that informs every engagement.

Research & Work

Applied research on governed AI.

In regulated domains, reliance on AI output requires traceability, reproducibility and verification; a general-purpose integration does not provide these by default. Two applications, in finance and in law, of the same principle: AI prepares; what it produces is traceable, reproducible and verified before anyone relies on it.

Working paper: finance

Sovereignty in Institutional Investment

Trust as a matter of architecture: AI drafts and flags, deterministic engines score and allocate, and the system fails closed.

Governed architecture for Investment Committees

Problem
Letting a language model write scores and recommendations into house data erodes trust: approximations compound, and synthetic fields poison the record. Ask a generative model the same question ten times and you often get ten answers, which is incompatible with risk process. An Investment Committee cannot act on intelligence it cannot audit, or on analysis built from history that was not knowable at the time.
Built
A path from house sources and vendor feeds to a contestable committee pack, then to allocation proposals under human authority. Scores, merge and reallocation sit in deterministic engines; language models draft and flag, with no authority over scores or capital. A second rail compiles research hypotheses to a closed vocabulary and runs them through pre-declared gates.
Validation
The scoring constitution has been stressed on a mandate shape, with outcomes recorded as process evidence.
Read the paper →Reference architecture in practice →

Working paper: legal

Citation-Grounded Case-Law Research for Jurists

Trust as a matter of architecture: every citation is checked against the exact source passage, and the system fails closed.

Citation-grounded case-law research for jurists

Problem
In legal research the difficulty is not fluency but verifiability. Even retrieval-based tools return the right document with the wrong passage, or a correct-looking quotation that does not support the claim, and Belgian courts have sanctioned submissions built on invented case law. Before a court, an unsupported citation is more damaging than no answer. Confidentiality is a further constraint: the client’s facts are covered by professional secrecy, and a general-purpose assistant sends them to third-party services with no guarantee that this secrecy is kept.
Built
Every displayed sentence is tied to a source passage kept at exact character offsets. A verifier removes any sentence whose quotation cannot be found at its cited location or does not support it; when nothing survives, the system says so. The case file is processed in memory and retrieval runs locally; only the language-model call leaves the machine.
Validation
Every displayed sentence passes two checks before it is shown: the quotation is found in full at its cited offsets, and the passage supports the claim.
Read the paper →Reference architecture in practice →

Contact

For a first conversation about where AI can create value in your organisation, and how to deploy it in a way your teams and regulators can trust.

Luc Lehette

Principal Consultant

Engineering at a Fortune Global 500 electronics group, then Head of AI & Data Engineering at a scale-up named 2022 and 2024 overall winner of Deloitte’s Technology Fast 50 Belgium and ranked no. 18 on Fast Company’s 2024 World’s 50 Most Innovative Companies.