Can AI Plan Your Investigation?

AI can produce an investigation plan remarkably quickly. Give it a short incident description and it can suggest witnesses, documents, photographs, equipment records, procedures and other information that might be relevant. Ask it to organise those sources using PEEPO and, with the right instructions, it can produce something that looks very much like the beginning of an investigation plan.

That capability is useful. It is also where investigators need to be careful.

In this series, we have been exploring a simple position: investigation is fundamentally a human thinking process. 

AI can support that thinking, sometimes very effectively, but there is a significant difference between using AI as an investigation tool and allowing AI to become the investigator.

Planning exposes that difference particularly well because an investigation plan is not simply a list of things to collect. It reflects what the investigation team currently understands about the incident, what remains unknown, what evidence may help resolve those unknowns and what needs to happen next.

There is another problem we need to add to that discussion: garbage in, garbage out.

Most people understand GIGO in the simple sense. Give an AI system poor information about an incident and you should expect a poor answer. But in investigations the problem goes deeper than that. The quality of the examples, reports, templates and investigation practices we provide to AI can also establish the benchmark from which it works.

If the benchmark is poor, AI can become very efficient at reproducing poor investigation practice.

The investigation plan is already part of the investigation

Within the Safety Wise investigation process, planning sits between the immediate response and detailed data gathering. The team needs to understand the Terms of Reference, receive the available incident briefing, establish resources and develop an Investigation Action Plan before moving into detailed evidence collection. PEEPO then provides a structured way to plan data gathering across People, Environment, Equipment, Procedures and Organisation.

That order matters because planning is already an investigative activity.

The team is deciding what it needs to understand. It is identifying potential sources of evidence and considering where information may be lost or altered if it is not preserved quickly. It is also making decisions about expertise, priorities and the boundaries established by the Terms of Reference.

PEEPO Mark1 supports this by helping investigators brainstorm potential data sources before detailed gathering begins. The purpose is to consider the system of work broadly rather than immediately concentrating on the people closest to the incident.

This seems like an ideal application for AI, and in many respects it is.

Imagine an unexpected release of stored energy during maintenance. The investigation team could provide the basic incident information, the Terms of Reference and its preliminary understanding of the task, then ask AI to suggest potential sources of information under each PEEPO category.

It might suggest maintenance history, isolation documentation, equipment drawings, interviews with the people performing the task, control-system logs, permit documentation, training records, engineering modifications, shift handovers or previous similar events.

Some suggestions may be useful. Others may have no relevance at all.

The important point is that the AI has generated possibilities. It has not determined what evidence the investigation requires.

That decision remains with the investigator.

The first GIGO problem: what did you tell it happened?

Consider an incident notification that reads:

"Operator reversed into parked light vehicle after failing to check mirrors."

Ask an AI system to develop an investigation plan from that description and you have already given it a theory of the incident.

The operator failed to check the mirrors.

The resulting plan may quite reasonably include questions about competency, reversing procedures, training, situational awareness and whether the operator checked before moving.

But what has actually been established?

Perhaps very little.

A more neutral description might simply state that mobile plant and a stationary light vehicle came into contact while the mobile plant was reversing in a maintenance area.

That description leaves considerably more investigative space.

Now there are questions about the parking arrangement, traffic separation, visibility, mobile plant design, reversing requirements, work area configuration, supervision, communication, physical controls and the way the task was normally undertaken.

The difference between those two investigation plans may begin with a few words in the initial information supplied to AI.

This is GIGO at the incident level.

If an assumption is presented to AI as a fact, the system may reason from that assumption. Worse, because the response can be structured and professionally written, the original assumption becomes harder to see.

The investigation now looks disciplined while potentially heading in the wrong direction.

That is not unique to AI. Human investigators can anchor on an early description as well. AI changes the speed and scale at which that initial framing can be turned into a detailed plan.

The second GIGO problem is more concerning

There is another version of GIGO that organisations need to think about before embedding AI into investigation systems.

What are you teaching the AI that a good investigation looks like?

Suppose an organisation has ten years of investigation reports. There may be hundreds or thousands of them. That looks like an incredibly valuable body of information for an AI system.

And it could be.

But quantity and quality are not the same thing.

Imagine those reports repeatedly contain findings such as:

"Operator failed to follow procedure."

"Employee demonstrated poor situational awareness."

"Worker did not conduct an adequate risk assessment."

"Supervisor failed to ensure compliance."

Recommendations then routinely include:

"Retrain all personnel."

"Remind employees of procedural requirements."

"Conduct toolbox talk."

"Employee to receive counselling."

If those reports are supplied to AI as examples of completed investigations, what has the AI actually been shown?

It has been shown that this is how the organisation investigates.

If 70 per cent of the reports stop close to the individual and 70 per cent recommend retraining, repetition itself creates a pattern. The system can reproduce that pattern extremely well.

It may even make the language cleaner.

That does not make the investigation better.

Can AI tell a bad investigation from a good one?

This needs a careful answer.

Sometimes.

If you provide AI with a defined quality standard, methodology or assessment criteria, it can compare a report against those requirements and identify potential weaknesses. For example, it could be asked whether findings appear to be supported by evidence, whether recommendations are linked to identified contributing factors, or whether the analysis stops at an Individual/Team Action without examining the surrounding Task/Environmental Conditions, Absent/Failed Defences and Organisational Factors.

That could be a very useful quality-assurance application.

But without that reference point, AI does not possess some independent organisational memory that says, "This investigation looks professional, but it is methodologically weak."

It works from patterns, instructions and the context available to it.

This creates an important distinction between examples and standards.

A collection of previous investigation reports tells AI what has been done.

It does not necessarily tell AI what should be done.

That distinction matters enormously.

Our ICAM approach is specifically intended to look beyond the errors and violations of individuals and examine the contributing factors surrounding an occurrence. The objective is to identify systemic deficiencies and strengthen defences against recurrence.

So if historical reports repeatedly stop at human error, they should not automatically become the benchmark for an AI-supported investigation system simply because they are the organisation's previous reports.

Before using historical investigation material to guide AI, somebody needs to ask whether those investigations were actually good.

That is a human governance decision.

Garbage can look very professional

This may be one of the biggest risks AI creates for investigations.

Poor investigation reasoning has traditionally left clues.

The report might be thin. The logic may jump around. Findings might not connect properly to evidence. Recommendations may feel generic. Experienced reviewers can often sense that something has gone wrong.

AI can remove some of those clues.

It can turn weak reasoning into excellent prose.

An assumption can become a polished paragraph. A shallow finding can be expressed confidently. A generic recommendation can be wrapped in the language of risk reduction and organisational improvement.

The presentation improves while the underlying investigation does not.

Our report-writing guidance requires investigation reports to present facts rather than opinions, remain accurate, clear and concise, and explain the reasoning used to reach conclusions. It also requires facts to have supporting evidence for validation.

AI can help with clarity and structure. It cannot manufacture the evidence that makes a statement a fact.

That boundary is critical.

A plausible answer is not an investigation finding

AI systems are designed to generate useful responses from the information and context available to them. In an investigation, however, "plausible" is a dangerous standard.

Investigators are not trying to produce the most likely sounding explanation.

They are trying to establish what happened and identify the contributing factors supported by evidence.

Our approach to data gathering emphasises gathering relevant facts and validating information through supporting evidence. Effective investigators are expected to maintain neutrality, use information validated as fact and refrain from premature conclusions.

This creates an interesting tension with AI.

Ask:

"What could explain why this worker bypassed the interlock?"

AI can probably provide numerous credible possibilities. Production pressure, poor design, previous equipment problems, normalised work practices, inadequate supervision, time pressure or a poorly understood procedure could all sound reasonable.

But "could have" is not "did".

Those suggestions may be useful as prompts for thinking about what evidence might exist. They should not quietly migrate into the PEEPO as established information, then into the timeline as fact, and eventually into the ICAM analysis as a contributing factor.

Once that happens, the investigation has begun analysing its own assumptions.

AI can make confirmation bias easier

Consider what happens if an investigation team already suspects inadequate training.

The team asks AI:

"What evidence should we collect to determine whether inadequate training contributed to this incident?"

AI responds with a comprehensive list: competency records, refresher training, assessment results, training materials, supervisor interviews, worker interviews and records of previous incidents.

The team gathers all of it.

What did nobody ask?

"What evidence should we gather to understand the system of work surrounding this incident without assuming training was a contributing factor?"

The first prompt seeks evidence around an existing theory. The second opens the investigation.

Again, this is not a new problem created by AI. 

Confirmation bias has always existed in investigations. The concern is that AI can make biased inquiry much more efficient.

The answer is not to avoid AI. It is to use a structured investigation methodology that provides the guardrails.

PEEPO gives AI boundaries

This is where PEEPO becomes particularly useful in AI-supported investigation planning.

Rather than asking AI to tell us what caused the incident, we can ask it to help us consider where relevant information might exist across the system of work.

People.

Environment.

Equipment.

Procedures.

Organisation.

The question changes from:

"What caused this?"

to:

"What information might help us understand what happened?"

That is a much safer role for AI.

An investigation team could develop its own PEEPO Mark1 first, then use AI as a challenge:

"Review this PEEPO Mark1 and identify potential categories of evidence we may have overlooked. Do not identify causes or findings."

That sequencing matters.

The investigators think first.

AI challenges second.

The team then evaluates the suggestions.

This avoids a situation where the AI-generated list becomes the starting frame for the entire investigation.

More data does not fix GIGO

There is a natural response to concerns about poor AI outputs: give the system more information.

Sometimes that helps.

Sometimes it simply creates more garbage.

An organisation may have thousands of documents available to an AI system, but if those documents include outdated procedures, superseded risk assessments, poor investigation reports, unverified incident summaries and inconsistent terminology, the system now has a larger information problem.

The same applies to an individual investigation.

Giving AI 2,000 pages of material does not mean it understands which information is reliable, current, relevant or contradictory unless the system and the investigator have a way of establishing those distinctions.

Evidence has context.

A procedure may have been current at the time of the incident but replaced afterwards. A witness statement may conflict with equipment data. A training record may show attendance but say little about actual competence. A risk assessment may exist in the system while workers describe a completely different method of completing the task.

An investigator has to work through those differences.

AI can help locate, compare and organise information.

The investigator determines what it means.

The investigation plan should keep changing

A good investigation plan is not finished at the planning meeting.

The team begins gathering evidence and discovers something unexpected. That creates another question, which identifies another source of evidence, which may require another interview or inspection.

The investigation develops.

A witness may mention that the equipment was modified three months earlier. Now engineering change records become relevant. Those records may identify a design decision. That decision may lead to another person who needs to be interviewed. The new evidence may then alter the timeline or change what the team thought it understood about a control.

That is why investigation planning is a thinking process rather than a checklist exercise.

AI can be extremely useful in managing that evolving plan. It could track open evidence requests, identify unanswered questions, compare new information with the existing timeline or flag areas of PEEPO that have received little attention.

But the decision that a new piece of information changes the investigation direction belongs with the investigation team.

The benchmark matters more than the technology

Before organisations ask how AI should be integrated into investigation planning, there is another question worth asking:

What standard are we asking it to reproduce?

If the organisation has a disciplined investigation process, trained investigators, clear evidence standards, quality PEEPO development, defensible findings and recommendations linked to contributing factors, AI has something useful to support.

If the existing process routinely accepts assumptions, stops at human error and produces repetitive low-level recommendations, AI may simply help the organisation do poor investigations faster.

This is why investigation capability becomes more important as AI capability improves.

A competent investigator can look at an AI-generated suggestion and ask whether it is relevant, what assumption sits behind it, what evidence would support it and whether it fits within the Terms of Reference.

Someone without that investigation knowledge may simply see a detailed and confident answer.

The quality difference sits with the human.

Use AI as the second set of eyes

There is a strong case for using AI during investigation planning.

It can review a draft PEEPO for gaps. It can help identify possible sources of corroborating evidence. It can organise actions from planning meetings, compare the plan against the Terms of Reference, highlight potentially time-sensitive evidence and help teams manage large numbers of outstanding actions.

It may also have a valuable role in investigation quality assurance, provided it is assessing work against a defined and trusted investigation standard rather than merely comparing the current report with previous reports.

That last distinction may prove particularly important.

Don't ask AI:

"Does this report look like our previous investigation reports?"

Ask it:

"Does this report meet our investigation standard?"

Those are completely different questions.

The first rewards conformity.

The second creates the opportunity for quality control.

So, can AI plan your investigation?

Yes, but that answer needs a qualification.

AI can help an investigation team develop, test and manage an investigation plan. It can generate possibilities much faster than most teams and can provide a useful challenge when investigators want to check whether they have considered the system broadly enough.

What it should not do is establish the direction of the investigation from untested assumptions, decide what evidence means, or determine the standard of a good investigation from a collection of previous reports that nobody has first checked for quality.

Garbage in, garbage out remains relevant. In investigation work, however, the "garbage" may be more difficult to recognise than a simple factual error.

It might be ten years of reports that consistently stop at human error.

It might be a finding template that encourages investigators to identify one cause.

It might be examples where every recommendation involves retraining.

It might be an initial incident description that presents an assumption as established fact.

Feed those patterns into AI often enough and we should not be surprised when they come back.

The opportunity is much better than that.

Give AI a sound investigation methodology. Give it clear boundaries. Give it evidence that has been properly identified and controlled. Give it a quality standard against which outputs can be challenged.

Then use its speed and processing capability to support the investigator.

Because AI may be able to generate the plan.

The investigator still needs to know whether it is a good one.