AI in investigations: speed is only useful if it sharpens decisions

A recent announcement from Relativity caught my attention. As part of an update to its Academic Program, it will now train future lawyers on how to use tools like aiR for Review in practice.

At first glance, that sounds like a training initiative. In reality, it’s a signal.

AI‑assisted review is moving from “interesting capability” to expected baseline skill. The next generation of lawyers will not be learning whether to use AI in document review, they will be learning how to use it well.

That shift matters, especially in investigations.

From speed to decision‑making

Most discussions about AI in review focus on speed. Faster review, lower cost, fewer documents for humans.

That’s true, but it misses the point.

In investigations, speed only matters if it improves decision‑making at the right moments:

• What do we escalate?
• Who do we speak to first?
• Are we looking at isolated behaviour or something systemic?
• Is there enough here to take a position?

Generative AI tools, powered by large language models, are increasingly being positioned as a first‑pass layer to surface key material earlier in the lifecycle. 

That changes the shape of an investigation.

Instead of spending weeks getting to “something to work with”, teams can move straight into testing hypotheses and building a narrative.

The real shift: faster iteration

The more interesting takeaway is not speed, it is iteration. This latest push from Relativity is not just about capability, it is about how these tools are used, refined, and validated in practice. 

In reality, that looks like:

• Start with a working theory
• Express it clearly
• Test it quickly across the data
• Refine based on what comes back
• Repeat

That loop is familiar. It is how good investigations already work. The difference is scale. Where that loop used to take days or weeks, it can now run in hours. Multiple theories can be tested in parallel. Contradictions show up earlier. Gaps become visible sooner.

For investigations, that opens up a different approach:

• Explore competing narratives rather than locking into one too early
• Surface risk signals before they become issues
• Direct human effort to judgement calls, not volume handling

The risk: speed without control

There is a catch. If AI makes it easier to run analyses, it also makes it easier to run them badly. As access increases, the risk is not misuse in a technical sense, it is unstructured use:

• Changing criteria without tracking why
• Treating early outputs as answers rather than indicators
• Using results to support a theory, instead of challenging it

In an investigations context, this matters. Poorly controlled iteration can lead to:

• Bias appearing faster, not slower
• Inconsistent outputs across workstreams
• Difficulty explaining how conclusions were reached if challenged later

Speed amplifies whatever process sits underneath it. Good or bad.

What this means in practice

If this kind of review becomes standard practice, then the way investigations are run needs to move with it. A few practical points to consider:

• Who defines the initial theory?
  And how clearly is it expressed before testing begins?
• How are iterations captured?
  Could you explain what changed between version one and version three, and why?
• What gives you confidence to act?
  Is there a clear threshold, or does it rely on judgement?
• Where does human skill add the most value?
  At the framing stage, the interpretation stage, or at key inflection points?

This is not about replacing people. If anything, it places more weight on judgement, clarity of thinking, and the ability to frame and test hypotheses properly. The technical barrier may be lowering, but the quality bar is rising.

Final thought

Training the next generation to use tools like aiR for Review is a clear sign of direction. The question is not whether this technology will be used in investigations. It already is. The real question is whether it is used to simply move faster, or to make better decisions sooner. Because in investigations, that is what actually changes outcomes.

As these tools become part of standard legal training and day‑to‑day workflows, the focus will shift from access to application, how they are used in live matters, how outputs are tested, and how decisions are supported. It raises practical questions around approach, governance, and consistency which are starting to come up more frequently in live investigations and disputes. For those dealing with these challenges in real scenarios, there is clear value in stepping back and sense‑checking how this is being embedded into existing workflows, and where it can genuinely improve outcomes without introducing new risks.