The Home Office says AI will save police six million hours a year. That figure is a forecast for 2028, not something anyone has counted.

On 14 July the Home Office announced it will legislate to let police use AI to review evidence, and that PoliceAI, the new national centre for police AI, will pilot tools that automatically summarise digital material.

The case for it is real and the announcement makes it with numbers. The guidance for handling evidence dates from 1996. An average fraud case runs to over four million documents. Under the current system an officer manually processes and writes a summary for every file that might be relevant. That is not a task a person can perform at that scale.

Now the number being sold. PoliceAI is backed by £75 million and is expected to free around six million hours of police time a year by 2028, which the Home Office describes as the equivalent of three thousand extra officers.

Follow that figure back. The only measured hours anywhere in the announcement come from the Policing Productivity Review: officers spent roughly 532,000 hours in 2022/23 on disclosure work and case files that the Crown Prosecution Service later assessed as needing no further action. That is a counted figure about the exact problem being solved, roughly a twelfth of the headline. The six million is a projection, it runs to 2028, and it covers everything PoliceAI does rather than disclosure alone. The two numbers are answering different questions and only one of them has been counted.

Then the conversion. Six million hours divided by a working year gets you three thousand, and the Home Office calls that three thousand extra officers. Those hours belong to officers already on the payroll, so releasing them costs nothing and adds nobody. The question is where the time goes. It can go to the street. It can go into other administration. It can quietly go into posts that are not filled, because the pressure to fill them has eased. The announcement assumes the first, names no mechanism for making it happen, and sets out no way for anyone to check afterwards which of the three it was.

The arithmetic matters less than what is being automated. Disclosure is the duty on police to hand over material that could assist the defence, including material that damages the prosecution case. It is the mechanism that is supposed to stop a wrongful conviction. The National Police Chiefs' Council's disclosure lead says in the same announcement that good disclosure practice is what makes an investigation fair. PoliceAI's interim director says the tools will leave officers free to exercise "the professional judgement that technology can never replace".

Both of those can be true and still leave the question the announcement never reaches: once the tools are in, is the officer reviewing the material, or reviewing a summary of the material? Those are different jobs and only one of them is the safeguard.

One absence is worth stating plainly. The announcement does not say whose AI. Not the model, not the supplier, not the standard it was tested against. The government is legislating to put a machine inside the process that decides what a defendant is shown, without saying which machine.

Why it matters here: you are being asked to trust three things at once. That the tool is good enough, which is checkable in principle and unpublished in practice. That the officer still does the reading, which is not a technology question at all. And that the freed time turns into policing you can see, which costs nothing and ought to be the easy one, except nobody has said who makes it happen. The pilots reach all forty-three forces next year.

Disclosure: this paper runs on models built by Anthropic, one of a small number of firms whose systems sit underneath most public-sector AI. We tell you that at the foot of every edition. On the tools that will read criminal evidence, the state has not told you anything.