The Future of Police Investigation: Verify AI

Policing has come a long way on evidence capture and storage.

Now a different challenge emerges: help humans find what matters inside an ever-growing flood of evidence.

Digital evidence now shows up in roughly 90% of criminal cases. In one Colorado DA office, annual video volume jumped from 24,000 hours in 2022 to 41,000 hours in 2025. A routine vehicular homicide there can now generate up to 90 hours of body-worn and dashcam footage.

That is the backdrop for the next era of policing.

Not a shortage of evidence.

A crazy amount of it.

And once you see that clearly, the future becomes obvious:

Investigations are shifting from humans manually analyzing evidence to humans verifying AI findings.

The investigator of the future will not start by watching everything. They will start by asking:

What should I look at first? What is most likely connected? What did we miss?

That is where I think evidence playback on Axon Evidence (evidence.com) is going.

1. AI triage: not “summarize this case,” but “show me what to watch first”

Today, multi-cam incidents are still brutally manual.

Five officers respond to one scene. All five record overlapping slices of the same event. The investigator scrubs through each one, guesses which angle matters most, and mentally stitches together a timeline.

Consider the math: 90 hours of footage from a single vehicular homicide. Even at 2x playback speed, that is 45 hours of scrubbing — to find the 2-3 minutes that actually matter in court. Investigators I talk to describe spending entire shifts just orienting on where to look. That ratio is the problem.

A smarter system would let the investigator ask:

• Which recordings cover the 90 seconds before the foot pursuit started?
• Show me the clearest angles of the first use-of-force moment
• Which videos capture the suspect’s right hand before the weapon appears?

The first query is buildable today with timestamp correlation and transcript search. The last one requires visual understanding that is getting close but is not production-ready yet. The point is that the direction is clear even if the full capability arrives in stages.

This is not a chatbot gimmick. It is an investigative triage layer for high-volume incidents — multicam use-of-force reviews, officer-involved shootings, prosecutor prep for a known incident window.

The natural starting point is not “AI understands all your evidence.” That would mean running expensive vision models across every frame of every upload, most of which is routine patrol, transport, static scenes, or dead air. Bad unit economics, weak ROI.

The smarter version indexes cheaply everywhere — metadata, timestamps, transcripts, camera-officer linkage, overlap detection — and reasons deeply only when the case earns it. A serious incident, an explicit investigator query, a prosecutor preparing for trial. That is when the system spends the compute.

The goal is to cut time-to-first-relevant-clip from hours to minutes.

That sounds simple, but it is a big change in product philosophy. Most evidence systems still assume the user already knows where the answer lives. In real investigations, they often do not.

2. The evidence graph: from files and folders to entities and relationships

This is the more important leap.

Most digital evidence products still think like storage systems. File here. Clip there. Metadata over there. Search box on top.

That worldview is running out of road.

Investigators do not think in files. They think in people, places, vehicles, weapons, events, claims, and contradictions.

So the system should too.

Imagine an evidence graph that maps entities — suspects, officers, victims, witnesses, vehicles, addresses, phones — and the relationships between them.

Now the search experience changes completely:

• Show me every evidence item connected to this vehicle and this address within 48 hours
• Which witnesses appear in incidents involving this same suspect?
• What evidence clusters around this firearm across cases?

That is a shift from finding files to understanding relationships.

Some of these connections already exist in isolation — case linkage, officer assignments, CAD dispatch records, RMS entries. The gap is that they live in separate systems and nobody has unified them into a queryable graph. The evidence graph is not inventing new data. It is making existing data talk to each other.

And that matters because investigations are not about discovering isolated facts. They are about connecting facts into something coherent enough to act on and rigorous enough to defend.

The hard part is accuracy. Entity resolution across messy police data — misspellings, aliases, partial DOBs — is genuinely difficult. In public safety, a false link is not a bad search result. It is a civil rights problem. So the system has to be honest about confidence levels and keep a human in the loop on every connection it surfaces.

Here is the thing that makes the evidence graph more than a research project: a surprising amount of investigative triage does not require deep visual understanding of video at all. Metadata, transcripts, timestamps, entity linkage, and CAD/RMS joins already capture most of the signal an investigator needs to orient on a case. Vision models matter for the hard queries — “show me the clearest angle of the suspect’s hand” — but graph plus transcript plus synchronization does the heavy lifting long before expensive frame-level inference is required.

A platform that already holds the evidence, the metadata, and the chain of custody is in the best position to build this. The graph is not a separate product. It is the next layer on top of an evidence platform that agencies already trust.

3. The quiet agent in the background

The most interesting AI in public safety may end up being the least flashy.

Chatbots that answer policy questions are useful. One-click video summaries save real time. But the next breakthrough may be something quieter — an agent that keeps working in the background, connecting dots across newly arriving evidence and notifying investigators when something worth reviewing emerges.

Here is what that looks like concretely: A new video gets uploaded to Case #4471. The transcript mentions “Marcus.” The system already knows Marcus Williams is a person of interest in Case #3892 from last month. The investigator on #3892 gets a notification: “A new video in an unrelated case mentions a name matching your POI. Review?”

That is not science fiction. It is transcript search plus entity matching plus a notification layer. The ingredients already exist. What does not exist yet is a system that runs that check continuously, across cases, without being asked.

Humans are bad at re-scanning the whole graph every time new material arrives. Machines are good at exactly that.

The risk here is obvious: alert fatigue. If the system notifies too often, investigators will ignore it entirely. So the bar for interruption has to be high, tunable, and transparent about why it fired. The worst version of this is a notification feed nobody reads. The best version is the investigator who gets a tap on the shoulder at exactly the right moment.

The best AI systems in investigations will not just answer questions. They will know when to interrupt with a better one.

4. A system that can escalate before the report is filed

One story from California has stayed with me.

In a KQED investigation, a police officer encountered a missing 14-year-old girl in a car with a 23-year-old man, did not write a report, did not investigate, and did not arrest the man. The story later describes another devastating failure involving the girl’s younger sister.

That story is horrifying on its own. It is also a product lesson.

Too many systems still assume the report is the start of the real investigative workflow. But sometimes the report is late. Sometimes it is thin. Sometimes it never comes.

What if the system did not wait?

The most feasible version starts with transcripts. If the transcript of a newly uploaded video contains keywords matching an active missing persons case — a name, a case number, a location — the system flags it before a report is written.

This video likely contains information tied to an at-risk person. Review now.

That is buildable today with transcript search and entity matching against existing case data. Harder capabilities — face matching against victim databases, cross-referencing prior abuse records — arrive later and require deeper integrations. But the transcript-based trigger alone could have changed the outcome in stories like the one above.

This is the most legally sensitive idea in this entire post. If the system flags something and nobody acts, the platform owner may share liability. If it fails to flag something, same problem. Any product team building this has to design the escalation model alongside legal and compliance from day one, not bolt it on after.

But that is not a reason to avoid it. It is a reason to build it carefully.

Not “store first, investigate later.”

But “listen early, escalate early.”

5. Most “AI for public safety” is still aiming too low

A lot of AI in this category still feels like workflow garnish.

Transcribe the audio. Summarize the clip. Generate the report. Add a chat box. Add a voice command.

Some of that is useful. None of it is the main event.

Even newer public safety AI rollouts from large vendors tend to emphasize analytics, voice controls, AI chat for chart generation, and automated reporting. That may improve workflow around the edges. It does not solve the harder problem of evidentiary reasoning across a messy case.

Here is the difference: today, a detective reviewing a gang-related shooting gets a transcript and a summary. That helps. But it does not tell her that the same vehicle in this footage appeared in a drive-by three weeks ago, that a witness in this video gave a contradictory statement in that earlier case, or that the firearm matches a description from an unsolved robbery. That is the difference between summarizing evidence and reasoning about it.

The future will not belong to the product with the prettiest summary.

It will belong to the product that helps investigators get to the right doubt faster.

A weak system helps you consume more evidence. A good system helps you ignore more of it. A great system helps you notice what is missing.

That is the standard I care about.

6. The job of the human gets smaller and more important

None of this means the machine solves the case.

It means the machine does more of the first-pass cognitive labor, while the human spends more time on what humans are actually good at: skepticism, contextual judgment, ethical restraint, and deciding what is strong enough to stand up in court.

That is why I do not think the future investigator looks less important.

I think they look more leveraged.

Less time retrieving. Less time scrubbing. Less time guessing where to start.

More time verifying. More time challenging the machine. More time answering the only question that really matters:

What actually happened?

The last decade in public safety was about capturing everything.

The next decade will be about knowing what deserves attention first.

The starting point is narrow: AI-assisted triage for multicam critical incidents. The destination is an evidence platform that reasons. The gap between those two is where the most important product work in public safety will happen over the next five years.

---