How evidence-based investigation actually works — Deepheem

Most people assume AI investigation means asking a chatbot a question and hoping the answer is accurate. Deepheem works very differently — every finding is grounded in retrieved source text, not generated from training data alone. Here's the exact process behind every investigation you run.

Step 1: Ask — define the brief

Every investigation starts with a brief — your question or topic. The more specific you are, the better. "Tell me about Tesla" produces a weak investigation. "Investigate Tesla's financial health and debt position in the European market over the last 18 months" produces a focused, evidence-grounded report.

The brief is routed to the correct investigation module — Compliance & AML or Legal Investigation. Each module uses different source priorities and evidence standards appropriate to that professional context.

Step 2: Search — live source retrieval

Before any findings are written, Deepheem retrieves sources from the open web, primary records, regulatory databases, and specialist sources. For compliance investigations, this includes FCA enforcement records, JMLSG guidance, MLR 2017, and FATF recommendations. For legal investigations, this includes BAILII and legislation.gov.uk. Multiple sources are retrieved per investigation, with actual content pulled — not just headlines.

This is the critical difference from a standard AI response. The findings are written from retrieved text, not from training memory. If a source doesn't exist in the retrieved set, it cannot contribute to a finding.

Step 3: Weigh — classify and assess

Every retrieved source is classified by authority type — so you always know what kind of source you're relying on before you act on it:

Judicial Authority — court judgments and case law
Official Legislation — statutes and statutory instruments
Regulatory Authority — FCA, CMA, Sentencing Council, and equivalents
Government Source — GOV.UK, ONS, Home Office datasets
Academic Source — peer-reviewed research
Authoritative Source — established independent publishers
Secondary Commentary — analysis, press releases, trade publications

Each finding is then assessed against the retrieved text and assigned one of four evidence statuses:

Supported — retrieved source text directly confirms this claim
Partially Supported — a source confirms part of the claim but not the specific conduct or outcome asserted
Unsupported — no retrieved source addresses this claim
Contradicted — at least one retrieved source directly conflicts with this claim

Step 4: Verdict — evidence-grounded conclusion

The verdict is written using only Supported and Partially Supported findings. Unsupported and Contradicted findings are excluded automatically. If evidence is incomplete, the verdict explicitly acknowledges the gaps rather than inferring a confident conclusion beyond what the retrieved sources confirm.

For legal investigations, UK citations are also verified directly against BAILII — confirmed citations receive a VERIFIED badge, unconfirmed receive LIKELY or UNVERIFIED so you know exactly what to check manually.

Why this matters

Traditional research takes hours or days. Deepheem does it quickly — and because every finding is grounded in retrieved source text and transparently labelled, you know exactly what the evidence confirms and what it does not, before you act on it.

The goal was never to replace human judgement. It was to give you the evidence base in minutes so your judgement can focus on the parts that actually require it.