Finding the information is becoming the easier part

A compliance analyst using AI for due diligence receives a dossier more quickly than before. It contains news coverage, corporate records and a summary of possible concerns. The material has been assembled efficiently, but the analyst still has to decide whether the named person is the right person, whether a report is credible and whether an old incident changes the decision at hand.

That is a different kind of work from collecting the information. It concerns the relationship between evidence and a particular judgement. As AI makes research easier, the distinction becomes more visible because the remaining effort sits precisely where the organisation needs a defensible interpretation.

S-RM’s report, From Screening to Decision-Making: The Next Phase of AI in Compliance, published on 5 October 2026, captures that tension. In a survey of 225 professionals, 81% reported a significant or moderate reduction in manual research time. Yet 58% said AI-enabled due-diligence solutions struggled to assess context, credibility or relevance sufficiently to replicate human-led research. These are respondents’ assessments, not independent measurements of the tools’ performance.

A trust gap is also a task boundary

The survey was commissioned in August 2026 and covered responsibilities spanning compliance and ethics, procurement, supplier management and third-party or vendor risk across the Americas, EMEA and APAC. S-RM is a corporate intelligence and cyber security consultancy. The evidence describes perceptions in those workflows; it does not directly measure marketing compliance or prove that the same proportions apply to campaign review.

The useful implication is narrower. Information gathering and interpretation can improve at different rates. A tool that removes repetitive research may still depend on experts to assess a finding’s significance. That dependence need not mean the investment has failed. It means the operating model should distinguish the work the system has accelerated from the judgement it has not established.

An illustrative marketing case makes the distinction tangible. An assistant retrieves a study supporting a product benefit, but the study concerns a different formulation or population. It has found relevant subject matter without establishing relevance to the claim. Alternatively, it finds a previous approval using similar language, but that approval depended on a disclosure missing from the new communication. The material is useful precisely because a reviewer can inspect why it does or does not apply.

An explanation that hides those conditions makes the analyst’s remaining work harder. The time saved in retrieval can be lost reconstructing the provenance behind a summary. A better handover preserves the source and states the connection being proposed, leaving the reviewer able to challenge it.

Human judgement contains work that can be reused

It would be easy to conclude that context is inherently human and leave the problem there. That is too broad. Expert judgement includes both novel interpretation and established reasoning that an organisation could make more available.

A reviewer may repeatedly distinguish between evidence for an ingredient and evidence for a finished product. A due-diligence specialist may repeatedly check identity before treating adverse coverage as relevant. Those procedures do not determine the answer in every case, but they can improve the sequence of work before the expert reaches the difficult question. Reuse can concern the method rather than the verdict.

This is the argument developed in what an AI agent should retain from expert work. The useful inheritance includes what to check and which conditions change the conclusion. An archive of final decisions is a weaker substitute because it encourages resemblance to stand in for applicability.

To capture that reasoning, a review record needs more than an accepted or rejected status. It should explain the decisive fact and the scope of the conclusion. If an analyst rejects a source because it concerns the wrong entity, the reusable lesson is the identity check, rather than a permanent judgement that the source is unreliable. If a claim is accepted only with a particular qualification, that condition belongs with the approval.

Such records take effort to create, and not every case deserves a lengthy account. The business can focus on decisions that recur or expose an important boundary. A concise statement of why a precedent does not apply may be more valuable than a detailed transcript of the entire discussion.

Escalation should begin where the reasoning changes

The more precise question is which conditions make a case materially different from the settled procedure. A new source type, disputed identity, conflicting evidence or an unusual customer context may justify specialist attention. Routine retrieval and comparison can continue to benefit from automation while those differences are surfaced early.

This creates a more useful role for the human than checking every output with equal intensity. The reviewer receives the reason the case has become uncertain and the evidence needed to assess it. The system should be able to explain which established procedure it followed and where the present case exceeded it.

The principle connects with meaningful human oversight. A person cannot exercise control simply by being present. They need time, evidence and authority to alter the result. If AI produces a much larger volume of findings without prioritising uncertainty, the organisation may speed up research while increasing the burden on the people whose judgement remains essential.

Performance measures should reflect that trade-off. Faster dossier completion is useful, but the team should also understand how much rework the handover creates and whether important ambiguities are being identified. A process that delivers more material to a reviewer is not necessarily a process that helps them reach a better conclusion. The measures should follow the decision the organisation needs to make.

Expertise can move without disappearing

The S-RM findings do not establish a permanent ceiling on AI capability. They describe reported experience with current tools in a defined set of responsibilities. Nor do they show that every human judgement is reliable. Experts also need sound evidence, consistent procedures and opportunities to challenge one another.

The commercial opportunity is to make expertise travel further through the workflow. An organisation can preserve settled reasoning, use it to structure routine work and direct specialists towards cases where the conditions have changed. That is a more specific ambition than either replacing judgement or insisting that all judgement must remain manual.

Research speed matters because it releases attention. What the organisation does with that attention will determine the value: whether experts spend it rebuilding context that the tool discarded, or examining the cases where interpretation can change the outcome. The next frontier is a better relationship between the finding and the decision it is supposed to support.