There is real money in selling AI to police. Axon, the company that makes TASER devices and body cameras, reported $669 million in quarterly revenue in mid-2025, up 33 percent from the year before. Its AI subscription plan, which bundles tools like automated report-writing into a flat annual fee, saw subscriptions jump 140 percent in a single year. The company has $14.4 billion in future contracted bookings. That is not a startup bet; that is an established business.
The pitch to police departments is simple: officers spend roughly half their working time on paperwork rather than actual policing. An AI tool that listens to body camera audio and writes the first draft of a report can claw back hours per shift. For a department already stretched thin on budget and staff, that sounds appealing. And the sales are working: Axon's AI product revenue grew 700 percent year over year according to company earnings disclosures.
The problem is that police reports are not the same as corporate meeting notes. A police report can put someone in prison. It can be used in court for decades. When a human officer writes that report, defense attorneys can question them on their choices: why they included certain details, what they left out, what mood they were in. That kind of scrutiny is impossible to apply to a machine.
Axon's Draft One was originally designed so that once an officer submitted a report, no original AI-generated draft was saved. The company updated that in late 2024, after California passed a law requiring departments to keep the original AI draft on file so that judges and defense attorneys could see which parts were written by the machine and which by the officer. Most US states still have no such requirement.
The harder problem runs deeper than paperwork. The same companies selling AI report-writing tools also sell the data-collection tools that feed larger AI systems used to help decide where to send officers, how to allocate resources, and which people or areas to watch more closely. This is not science fiction: these systems are live in departments across the country right now. Axon's Fusus product, for instance, connects feeds from cameras, license plate readers, and other sensors into a single dashboard.
This pattern has failed before. In the 2010s, a system called PredPol used crime statistics to tell departments where to patrol next. The algorithm had no way to understand that higher reported crime rates in poorer neighborhoods were partly the result of those neighborhoods already being over-policed. It just read the data and sent more officers to the same places, in a self-reinforcing loop. Researchers who tested the algorithm found it directed police almost entirely toward lower-income, minority areas, even when public health data showed drug use spread evenly across cities. PredPol quietly rebranded and eventually collapsed, but the lesson has not yet been fully applied to the new wave of tools.
The regulatory situation is fragmented. There is no comprehensive US federal law governing AI in policing. In 2024, 45 US states introduced AI-related legislation, but only around 10 enacted anything. In December 2025, a Trump executive order aimed to limit states' ability to regulate AI, though it specifically exempted laws governing how state governments themselves buy and use AI, meaning state-level police oversight rules still stand. California's new disclosure law is a start. Most places are still operating with no guardrails at all.
The same issue is not limited to the US. In the UK, one police force was found to have made a decision based on information that Microsoft's Copilot had fabricated, and that was not caught before action was taken. The pattern is consistent globally: AI tools are being sold into high-stakes environments faster than anyone has thought through what to do when they go wrong.
For business operators and professionals, the relevance here is not abstract. Any business that interacts with law enforcement, either as a victim reporting a crime, as a party in a legal dispute, or as an employer whose staff are involved in incidents, may find AI-generated documentation on the other side of that interaction. Understanding that these reports exist, that they can contain errors, and that the rules around disclosing their AI origins vary by location is genuinely useful knowledge.