Resources
Research in regulatory intelligence.
How reliable is the data? Cleo's architecture is described in a research paper published in February 2026: 93 to 97% precision on binding texts, evaluated on 7,676 regulations.
Research
The research paper
February 2026 · 20 pages
A multi-agent pipeline approach to regulatory intelligence
Original title: MARIA (Multi-Agent Regulatory Intelligence Architecture)
Naomie Halioua, Alexandre Bloch, Anaëlle Guez · Cleo
Key figures, per pipeline run
- 19geographic regions
- 16sector domains
- 8languages tracked
- 30+specialised analyses
- 370 to 1,200web searches for a typical company
Research
The chain in five stages
Beyond keyword search
| Key question | Keyword | LLM | Pipeline | |
|---|---|---|---|---|
Entity understanding |
What does this company do? | No | Partly | Yes |
Legal knowledge |
Which frameworks exist and interact? | No | Partly | Yes |
Temporal awareness |
What has changed this week? | No | No | Yes |
Domain reasoning |
How does this affect this company? | No | No | Yes |
Research
Where Cleo sits
Research
Search infrastructure
where Nreg is the number of regulations passing the relevance filter, ε ∈ [0, 28] captures the agentic follow-up searches, and C the number of regulatory blocks retained from a set of 37 (19 geographic, 2 cross-cutting, 16 sectoral), with C ≤ 37.
Streaming architecture
Each stage waits for every article to have been processed before the next one starts.
Each article moves through scoring, enrichment and assessment as soon as it is ready.
30 minutes with the team
See what Cleo finds on your own products
One product range and two markets are enough.
