The danger of removing humans from AI-powered investigations

AI can analyse vast datasets in minutes, but corporate investigations are rarely resolved on data alone. Ana Pereu, associate director, disputes and investigations at S-RM, argues that as organisations embrace increasingly sophisticated investigative tools, human judgement remains essential to providing context, accountability and defensible conclusions.

Artificial intelligence is reshaping corporate investigations at pace. Tasks that once required weeks of manual review, from analysing thousands of documents to tracing financial flows across jurisdictions, can now be completed in a fraction of the time.

For organisations handling large-scale investigations, the efficiency gains are undeniable. But efficiency is not the same as insight. As AI becomes more deeply embedded in investigative workflows, understanding that distinction is becoming increasingly important.

The limits of pattern recognition

AI excels at scale, identifying anomalies in financial data, surfacing relevant documents within large datasets, and flagging potential risk indicators across multiple sources simultaneously. What it cannot do is determine what those findings mean in context.

A machine learning model reviewing internal communications may process language too literally, missing industry-specific euphemisms, deliberate misdirection, or the coded language that experienced investigators immediately recognise. In a bribery investigation, the difference between a legitimate commercial arrangement and a disguised facilitation payment often comes down to context, not content. References to “marketing support” or “business development expenses” can appear entirely routine to an algorithm while signalling something very different to an investigator familiar with the sector, jurisdiction and individuals involved.

The same principle applies in forensic accounting. Automated tools can quickly identify unusual transactions and links between entities, but determining whether those transactions represent fraud, error or legitimate business activity still requires professional judgement, sector knowledge and an understanding of how organisations actually operate.

The misinformation problem

The volume of publicly available information has grown enormously, but so has the volume of information that is unreliable, politically motivated, or generated by AI itself. Open-source intelligence platforms are powerful and, when combined with advanced analytics, can reveal networks, affiliations and patterns that were previously invisible. But OSINT is inherently ‘noisy’.

Distinguishing credible information from background noise, verifying sources and understanding why information exists in the public domain in the first place all require human judgement. Human intelligence sources remain equally important. Local expertise, trusted relationships and on-the-ground knowledge provide context that no database can offer. In many complex investigations, the most important information never appears online at all.

Automation bias is a real risk

One of the less discussed consequences of increased AI adoption is automation bias – the tendency to treat machine-generated outputs as more reliable or complete than they actually are.

In an investigative context, this is a serious problem because an output that appears comprehensive may simply reflect the boundaries of the dataset from which it was drawn. A clean result can create a false sense of certainty. The absence of evidence in a dataset does not necessarily mean the absence of misconduct.

Experienced investigators approach findings with scepticism regardless of how they were generated, asking what is missing as much as what is present. They understand that some of the most significant findings in an investigation may sit outside the available data entirely. That instinct cannot be built into a model, and it is often what separates a thorough investigation from one that merely appears thorough.

Corporate investigations are also fundamentally about people. Interviews remain a cornerstone of investigative work, relying on rapport, empathy and the ability to assess both what is said and what is left unsaid. Inconsistencies in tone, hesitation and non-verbal cues all inform an investigator’s assessment in ways that no automated process can replicate.

Regulatory expectations are catching up

There is a growing regulatory dimension to this that organisations cannot afford to ignore. For organisations conducting investigations, enhanced due diligence or compliance reviews, the challenge is no longer whether to use AI, but how to use it responsibly.

Guidance published by the Civil Justice Council in October 2025 made clear that AI used in legal and investigative contexts must operate within a framework of professional oversight and accountability, with clear governance structures in place if public confidence is to be maintained. GDPR’s Article 22 already places constraints on automated decision-making where significant conclusions are drawn about individuals.

For organisations conducting enhanced due diligence, the implications are straightforward. Regulators are interested not only in whether checks were carried out, but in how conclusions were reached and who was accountable for them. Without clear internal review processes, AI can scale inaccuracies as well as efficiencies, and an investigative process that cannot demonstrate a clear chain of human judgement and oversight is increasingly difficult to defend.

Human expertise remains essential

Used well, AI allows investigators to focus their expertise where it matters most, moving quickly through lower-priority material and concentrating on areas that require deeper analysis. Technology and human expertise are most effective in their combination, with AI handling scale and consistency while humans provide context, judgement and accountability.

AI will continue to transform corporate investigations, improving speed, consistency and scalability. But the organisations that derive the greatest value from these tools will not be those that automate the most. They will be those that combine technology with experienced investigators capable of challenging assumptions, interpreting context and exercising sound judgement.

In an environment where regulatory scrutiny is increasing and misinformation is becoming harder to detect, human expertise remains the difference between gathering information and understanding what it actually means.