Summary
- The Home Office is backing a £1.4 million AI call-routing system for the police 101 service.
- Around four million of 20 million annual 101 calls concern organisations other than the police.
- Callers can reject the system’s interpretation, uncertain cases go to police, and human handlers remain available.
Britain’s police are applying artificial intelligence to a problem that long predates the current technology boom: millions of people use the non-emergency 101 number each year for issues that another public service should handle, leaving operators to spend time identifying the problem before sending the caller elsewhere.
The Home Office and National Police Chiefs’ Council have developed an AI call-routing system, with Vodafone Business as delivery partner, that matches the reason somebody gives for calling 101 against likely categories and redirects clear non-policing issues before the caller enters the main police queue.
The government is investing £1.4 million in the project and says the 101 service receives about 20 million calls each year, of which roughly four million concern organisations such as local councils, NHS 111, or the electricity-network 105 service rather than the police. Ministers estimate that the system could eventually save policing as much as £8.5 million annually.
The deployment is narrower than allowing an AI model to decide how police should respond to an incident. Its immediate task is classification at the entrance to the service, where predictable demand can potentially be redirected while uncertain or more complicated calls continue towards human operators.
Triage is narrower than replacing the operator
At present, many misdirected calls are identified only after an operator has spoken to the caller. Automating part of that initial sorting could reduce time spent handling contacts that should never have entered the police queue.
The safeguards are therefore central to the design. The Home Office says callers are told how the system has understood the reason for their contact and must accept or reject that interpretation. Broader categories trigger additional safeguarding questions, while cases where the system remains uncertain are passed to the police.
Callers can also speak to a human handler if they want to do so. That distinction prevents the routing model from becoming a compulsory automated gatekeeper for every person trying to contact the non-emergency police service.
At the reported volumes, modest reductions in unnecessary handling could release substantial staff time. Yet the size of the saving will depend on caller behaviour as well as technical accuracy because people frequently use 101 precisely when they are unsure which public body is responsible.
Speech systems have to work across callers
The Home Office says the system has been tested to ensure that performance is not affected by factors including accents and phrasing. Pilots across 15 police forces have already redirected examples such as damage-only road incidents, fly-tipping, and noise complaints to other agencies.
Language systems nevertheless operate in an environment where errors carry different consequences from conventional commercial customer service. Redirecting a routine council enquiry incorrectly may create inconvenience, whereas diverting a caller whose description contains an understated safeguarding or criminal concern could delay access to police assistance.
Fallback behaviour is therefore as important as average classification accuracy. The additional questions and direct route into police handling where doubt remains are intended to constrain that risk.
Performance after deployment will still need to be examined through operational evidence, including how often calls are routed correctly, how frequently people override the system, whether redirected callers return to 101, and whether any categories of caller experience disproportionately poor outcomes.
The programme reaches beyond one phone queue
The call-routing system forms part of £16.5 million of government investment intended to modernise contact between police and the public. Other elements include AI transcription for 999 and 101 calls and technology intended to link crime reports to identify patterns in demand.
Those projects share a common attempt to reduce the administrative work required to convert human communication into structured information. Calls have to be understood, recorded, categorised, and connected with other information before officers or staff can act, creating tasks where language technology can remove manual steps without necessarily making the final policing decision.
Each new automated layer also creates infrastructure that forces have to govern. Models need monitoring after updates, call data can contain highly sensitive personal information, and the system may interact with people who are distressed, confused, or vulnerable.
Efficiency consequently cannot be judged only by average call-handling time. Police contact services exist partly to identify risks that may not be obvious when somebody first explains a problem, so automation that removes workload has to preserve the ability to recognise an unusual case among millions of routine contacts.
If the system reliably redirects obvious non-policing enquiries, it could become one of the more practical applications of AI in public administration. The strongest evidence will not be a claim about model sophistication but shorter queues, fewer unnecessary operator interactions, preserved safeguarding, and proof that the projected savings represent genuinely removed workload rather than demand shifted elsewhere.
