Where agentic AI is doing real work for understaffed investigations units, and what to insist on before you buy. 

Ask a working investigator how many open cases they're carrying and the answer is usually somewhere around twenty. The rhythm is familiar to anyone who's done the job: work a case until you hit a wall (a subpoena in flight, lab results pending, a witness who won't call back), then shelve it and pick up the next one. The walls aren't the problem. The context-switching tax is. Every time an investigator picks a case back up, they re-read their notes and try to remember where they left off.  

Meanwhile, digital evidence volumes keep climbing, and the pattern-level insights that clear cases stay locked inside data silos no one has time to stitch together. The problem isn't access. It involves reconciling a suspect's alias across systems, walking their associates out two degrees, and cross-referencing communications. And cases go cold not because investigators missed something, but because they ran out of hours.

AI isn't going to hire anyone. But a particular kind of AI, often called "agentic" or "multi-agent," can do real work on behalf of the investigators you already have, precisely in the gaps where cases go cold waiting for attention.

What "multi-agent" actually means

Rather than one general-purpose chatbot, a multi-agent system is a team of specialized AI components working under an orchestrator. Picture your most capable analyst who can delegate to an army of equally capable analysts: one sub-agent pulls records, another monitors open sources, another reads through case files. The orchestrator synthesizes what comes back into a structured, cited briefing. Humans stay in the loop at every consequential step. Done right, the output doesn't look like a magic answer. It looks like the work product of a careful analyst: sourced, graded for confidence, and ready for an investigator to verify and act on.

Here are three places this pattern is starting to earn its keep in law enforcement.

1. Connecting the dots across siloed systems

Most of what an investigator needs is scattered across the agency's own systems: RMS, CAD, jail bookings, field interviews, CCTV and other video, regional intel platforms, plus a subject's open-source footprint. Increasingly, the dominant category of evidence isn't physical. Warrant-based digital evidence (call detail records, tower dumps, phone extractions, etc.) has overtaken traditional forensics in importance, and it arrives with enough noise that the problem isn't finding a needle in a haystack. It's finding that needle while buried underneath the haystack.

The bottleneck isn't access. It's time. Reconciling an alias across five systems, cross-referencing communications across a dozen devices, walking a suspect's associates out two or three degrees is hours of work per subject, which is why pattern-level insight so often arrives after the arrest rather than before.

A multi-agent setup handles this in parallel. Specialized sub-agents query each source simultaneously. An orchestrator reconciles entities (same person, different spellings; same vehicle, different plates), and a human reviews a consolidated picture with every claim cited back to its source. The use cases that benefit most:

• Gang and organized-crime intelligence, where relationships matter more than any single incident
• Prolific-offender tracking, where patterns across low-level contacts often predict escalation
• Serial and pattern crimes, where the signal is in the aggregation

The investigator's job shifts from assembly to judgment, where it belongs.

2. Cases that keep moving. Knowledge that doesn't walk out the door.

Agencies are losing senior officers to retirement faster than they can replace them. A brief handoff, a few hours with a case file, a conversation over coffee, can't transfer what a veteran detective carries in their head: the mental library of what-to-do-when, built over twenty years, almost none of it written down.

This is the quieter crisis. And some would argue it matters more than throughput.

A system trained not just on investigative practice but on continued interaction with your own senior investigators converts that expertise into something a junior investigator can actually use: a partner that surfaces the next logical step, flags what the veteran would have noticed, and explains why. The rookie doesn't become an expert overnight. But they operate closer to one from day one.

The productivity argument for AI gets made constantly. The workforce continuity argument doesn't get made nearly enough. For agencies watching their most experienced people walk out the door, tools that help junior investigators operate closer to their senior colleagues aren't just a productivity gain. They're a hedge against a workforce crisis.

3. Building the case file, not just finding leads

The volume of material attached to any given case has exploded: body-camera transcripts, interview recordings, warrant returns, phone extractions, and corroborating open-source material. Sub-agents can extract entities and events from each source type, reconcile timelines, surface contradictions between statements, and produce a structured summary with every claim cited and confidence-graded. A human investigator reviews and owns the output.

Strong applications:

• First-pass case-file summaries to orient a supervisor or a new investigator picking up a case
• Cold-case revisit, where running new tools and open-source corroboration against old files regularly surfaces new connections
• Prosecutable-package preparation, where the discipline of citation and confidence grading pays dividends under discovery

The goal isn't speed for its own sake. It's moving investigator time from reading to interviewing.

The Non-Negotiables

Any chief evaluating tools in this space should insist on four things:

1. Every claim cited. If the system can't tell you where a fact came from, it shouldn't be in the brief.
2. Confidence grading at the claim level. Investigators already think in degrees of certainty. The tool should too.
3. Human in the loop on every consequential decision. Agents assemble; humans decide.
4. An audit trail that survives discovery. If the workflow can't be defended in court, it doesn't belong in an investigation.

At Rilian, we've spent years building multi-agent systems for national-security work, where these four properties aren't optional. The mission is different. The standards are the same.

Mike Joy is a retired NYPD Captain and leads the Product team at Rilian.