Identify the role AI played
The phrase “AI-assisted investigation” can describe very different activities: extracting dates from documents, proposing a lead, grouping records or drafting a narrative. Ask which task the system performed and which decisions remained with an investigator.
A useful description identifies the inputs, the output and the human validation step. It should make clear whether the model generated new language, ranked possible connections or applied a defined classification method. Those activities create different questions for the reviewer.
Used well, automation can organise larger transaction sets, surface patterns and prepare material for review, giving investigators more time for interpretation and strategy. In a fictional workflow, a system flags a possible service connection; an investigator checks the source records and alternatives before recommending a targeted enquiry. The suggestion becomes useful through that validation.
Read the full explanation: Recoveris’s current AI-assisted investigation workflow.
Treat convincing presentation as a separate question from authenticity
The FBI has described the use of generative AI in fraudulent text, images, audio and video. For a crypto matter, this means a persuasive message or recording may need provenance checks before it is relied upon as evidence of identity.
Keep the original material and its context, including the channel through which it arrived. Record what the client observed without assuming that every inconsistency proves AI use. Specialist authenticity analysis may be needed; a visual impression or an automated detector alone should not be treated as a conclusive finding.
Reference: FBI: generative AI and financial fraud
Verify the claims that could affect the case
NIST’s Generative AI Profile identifies risks including confabulation and automation bias. A fluent explanation can therefore be a starting point for review while still containing unsupported facts or references.
For a report draft, trace important dates, addresses, amounts and citations back to retained sources. For a suggested link, ask what independently observable evidence supports it. A model-generated confidence value should not be treated as a calibrated probability of guilt or ownership.
When a correction changes the conclusion, retain a record of the change and the reason. This helps distinguish the investigator’s final analysis from intermediate machine output.
Reference: NIST: Generative AI Risk Management Profile
Document use without exposing the case unnecessarily
For a reviewable record, capture the tool and version where available, the task, relevant configuration, source inputs, review actions and limitations. The purpose is to explain how the final result was reached, not to preserve sensitive data in every possible location.
Before using external AI services with case material, establish the permitted data handling, retention and access arrangements. A public chatbot should not become an unplanned repository for privileged or confidential records. Counsel and the organisation’s responsible teams should determine which controls apply.
The final report should state conclusions the responsible expert can explain and support. An impressive demonstration of software capability does not answer that evidential question.
Apply the distinction to an example
Optional knowledge check
An AI tool gives a wallet link a “98% confidence” label. What should you do?
Read the explanation without answering
Ask what the score means and independently check the supporting evidence. The definition, validation and underlying evidence matter more than the apparent precision.
Review the points you want to take forward
Your review checklist
0 of 4 steps reviewed. This is a learning checklist, not a safety or recovery score.
Answers stay in this tab and reset on reload. The PDF is a concise reference with a blank checklist.
Explore a related guide
References and further reading
Examples are fictional. Checklists are editorial learning aids and do not assess an individual case.

