AI-driven Blockchain Intelligence Must Be Built for Accountable Decisions
By Daniel Schink, Head of Marketing at Recoveris
A blockchain investigation may begin with an alert, but its practical value depends on whether the findings can support an asset-freeze request, regulatory filing or court proceedings. In our work at Recoveris, tracing a transaction is rarely the slowest part. The harder task is turning that trace into a documented case that gives the relevant institution or authority a clear basis to act. The evidence needed to do that is often scattered across block explorers, analytics platforms, databases, spreadsheets and messaging channels. AI can shorten that path, but speed has little value if the result cannot withstand scrutiny.
This is an operational problem. FATF's July 2026 update describes the industrialisation of virtual-asset fraud by organised crime groups, alongside gaps in supervision, enforcement and international cooperation. Banks and crypto businesses face more alerts and cross-chain activity, with too few investigators to interpret them. A new dashboard leaves the fragmentation untouched. Our view is that teams need one workflow that records how each conclusion was reached, so the case can withstand regulatory and legal scrutiny.
What AI needs to support real investigations
Useful AI systems assemble and preserve the case as they work. They can collect cross-chain transaction data, group related addresses, compare behaviour with known financial-crime patterns, connect blockchain transactions with bank records, customer information and other evidence outside the blockchain and draft a chronology for review. Reducing manual copying gives specialists more time to test hypotheses and decide what the evidence supports.
FIS's Financial Crimes AI Agent points in this direction. Announced in May 2026, it is designed to assemble evidence across bank systems, evaluate activity against known patterns and surface high-risk cases for review. FIS says the system keeps data within FIS-controlled infrastructure and records how its conclusions were reached. Although these are expected rather than independently proven results, the design illustrates an important point: AI must be connected to reliable data, secure systems and clear human oversight.
At Recoveris, this problem shaped the AI Investigator we are developing. It uses two separate models. The first analyses transactions using patterns learned from verified investigations. The second turns that assessment into language an investigator can understand. Keeping those roles separate means the language model explains the decision without making it. Investigators can then see how the result was produced, question it and repeat the analysis.
The same principle applies to the wider investigation workflow. An explainable model is not enough if its findings are later copied into spreadsheets or reports without the supporting evidence. At Recoveris, we address this through Case Workspace and the Tracker browser extension.
Tracker connects to multiple blockchain analytics tools at once, allowing investigators to compare wallet information without moving between platforms and save their findings directly to the case record. Case Workspace keeps the evidence, decisions and reporting together. This helps keep conclusions connected to the supporting data and human judgement.
A defensible result needs a visible evidence trail
We define explainability as a visible path from a wallet label or risk score back to the transactions, sources and reasoning that produced it. The case file should show where the money moved, when it moved, how the person or business was connected to it, and whether there are other reasonable explanations. Without that chain, faster pattern recognition cannot support a freeze, customer restriction or legal action.
The Financial Stability Institute's 2025 paper reaches the same conclusion at the governance level. It warns that opaque models weaken accountability and independent review, while AI-generated explanations can be inaccurate or unstable.
Systems should record which sources, data and model versions were used, how certain the result is, and what decisions the investigator made. High-impact findings should also be checked independently. When providers disagree, the disagreement should remain visible instead of disappearing into a single score.
Human control must shape the investigation
Human review belongs at every point where judgement changes the course of a case. Investigators need to see what information the AI used, reject an incorrect link between a wallet and a person or organisation, change how the money is being traced, and record why they made those decisions. Routine operations, such as applying an approved rule, importing transaction data or drafting a standard report section, can be automated. Exploratory leads remain provisional until a specialist verifies them.
That distinction should determine permissions. AI can surface an unconfirmed lead; any customer restriction, freeze request or legal attribution must remain a human decision. Access permissions should limit who can take consequential actions, while clear approval rules keep those decisions with authorised investigators and record who decided what.
Responding at the speed of financial crime
Modern financial crime moves quickly across platforms and jurisdictions, so the response must move with it. In one high-volume crypto-exchange deployment of our Aftercare Protocol, our team handled around 100 cases monthly, typically responding within 12 hours, and achieved a 27% freeze rate. These results came from collecting case information quickly, preserving the evidence and connecting investigators with the exchanges, asset issuers and authorities able to act before the money moved again.
This is the standard AI-driven blockchain intelligence should meet. It should help investigators reach a defensible decision and act sooner while preserving the evidence trail, human judgement and accountability required for regulatory or legal scrutiny.
Sources
Financial Action Task Force, Seventh Targeted Update on Implementation of the FATF Standards on Virtual Assets and VASPs, 16 July 2026. https://www.fatf-gafi.org/en/publications/Fatfrecommendations/targeted-updated-virtualassets-vasps-2026.html
FIS, FIS Brings Agentic AI to Banking with Anthropic Starting with Financial Crimes, 4 May 2026. https://www.investor.fisglobal.com/news-releases/news-release-details/fis-brings-agentic-ai-banking-anthropic-starting-financial
Financial Stability Institute, Managing explanations: how regulators can address AI explainability, 8 September 2025. https://www.bis.org/publications/fsi-paper-24-managing-explanations-how-regulators-can-address-ai-explainability
Recoveris, Why the model behind our AI investigator is a white box, Aug 19, 2026. https://recoveris.io/why-the-model-behind-our-ai-investigator-is-a-white-box/
