Research
The science
behind the intelligence.
Every result Cleo produces is traceable, measurable, and auditable. We decompose, evaluate, and publish our work openly.
Latest Paper
A Multi-Agent Pipeline Approach to Regulatory Intelligence
Cleo Labs, Neuilly-sur-Seine · Forthcoming, 2026
We formalize regulatory intelligence as a distinct computational problem and present a production system that, given a company domain, identifies applicable regulations across 19 regions and 16 sectors, discovers developments in 8 languages, scores findings on a risk matrix, and generates impact assessments, orchestrating 30+ specialized LLM calls and hundreds of web searches per run.
19
Geographic regions
16
Sector-specific domains
8
Languages monitored
30+
Specialized analyses per run
420–1,200
Search calls per run
0.81
F1 score (full pipeline)
The five-stage pipeline
Regulatory intelligence decomposed into five sequential tasks, each addressing a distinct capability gap. Click any stage to inspect its inputs, outputs, and prompt template.
Stage 01 · Profile & Map
§3.1 · companion paperInput
Company domain (e.g. loccitane.com)
Output
Structured entity profile · 19 region map · 16 sector tags
TASK: extract structured entity profile from {domain} OUTPUT_SCHEMA: sectors: [string] // align to NACE rev. 2 regions: [string] // ISO-3166 alpha-2 products: [string] // identified from sitemap size_class: enum["startup","sme","midmarket","enterprise"]
Beyond keyword search
Regulatory intelligence requires four capabilities that keyword search cannot provide.
→"Which regulations apply to L’Occitane cosmetics sold in Germany?"
Keyword searchidle
LLM only · GPT-4idle
MARIA pipelineidle
01 · sourcing · 660 sources
02 · filtering · cosmetic × DE
03 · scoring · 8 dimensions
04 · validation · cross-check sources
05 · synthesis · structured brief
Filling the market gap
Existing tools cluster along the diagonal. Proactive, company-specific regulatory intelligence remains an empty quadrant.
Company-specific
Generic
Reactive
Proactive
GRC Platforms
Change Mgmt
Legal AI
Compliance Automation
Cleo
Search infrastructure
A single pipeline run decomposes into hundreds of targeted searches across jurisdictions and languages.
S_total = (20 + ε) + (37 × 4) + 5 · N_reg
where N_reg is the number of regulations passing the relevance filter, ε ∈ [0, 28] captures agentic follow-up searches
Streaming architecture
Batch (traditional)
Score
A1
A2
A3
A4
Enrich
A1
A2
A3
A4
Assess
A1
A2
A3
A4
T = Σ max(tij)
Streaming (Cleo)
A1
A2
A3
A4
T ≈ max(Σ tij)
The White Paper
Dive deep into the methodology, architecture, and evaluation behind Cleo's regulatory intelligence pipeline.
Latest Paper
Forthcoming, 2026
A Multi-Agent Pipeline Approach to Regulatory Intelligence
We formalize regulatory intelligence as a distinct computational problem and present a production system that, given a company domain, identifies applicable regulations across 19 regions and 16 sectors, discovers developments in 8 languages, scores findings on a risk matrix, and generates impact assessments, orchestrating 30+ specialized LLM calls and hundreds of web searches per run.
Download White PaperTrusted by international brands
Interested in our
research?
We're building in the open. Get in touch to discuss research collaborations.
