At Dreamforce 2026, Thames Valley Police and Hampshire and Isle of Wight Constabulary shared how they are using an AI agent to support public-facing, non-emergency contact.
The agent, Bobbi, is designed to help people self-serve on common enquiries, reduce pressure on emergency and non-emergency call handlers, and identify vulnerability or high-harm situations that should be escalated to a person.
That final point is the important one.
For the forces, the deployment is not simply about reducing contacts. It is about ensuring the right contacts are automated and the right people reach a human operator when they need help.
Why Thames Valley Police Deployed Bobbi
Tom Kempster, Director of Digital for Thames Valley Police and Hampshire and Isle of Wight Constabulary, said the forces wanted to make an AI agent available to the public across both organizations.
There were two aims.
First, people could access answers to frequently asked questions they would otherwise bring to the police. That, in turn, could free call handlers so people who genuinely needed police support could reach the service more quickly.
Second, the agent could identify vulnerability and high-harm situations, escalating them to a human where needed.
“We wanted to ensure that people were able to self-serve and get answers to frequently asked questions they would normally come to the police for.”
The forces could have taken a more conventional internal-first approach, beginning with IT or HR and applying the lessons to public contact later. Instead, the contact-centre leadership had the confidence to go directly to the public.
The Goal Is Not Just Automation. It Is the Right Escalation.
Bobbi had been live since November of the previous year, with just over six months of data at the time of the interview.
The forces said around 75–76% of contacts coming into Bobbi were subject to automation. The remaining 25% were escalated to operators on the 24/7 digital desk.
But the team’s central measure is not the automation percentage in isolation.
“Crucially, the right 25% are escalated.”
The force says its knowledge-base articles and topic structure are intended to ensure that lower-risk demand can be handled through the agent, while the contacts most in need of human support reach an operator.
“The right 25% is dealt with by a person, and the right 75% is subject to that automation.”
That distinction is central to any organization using AI in a sensitive service environment. The operational question is not only whether the technology can handle demand. It is whether it can recognize when it should not.
How Bobbi Identifies Vulnerability
Behind Bobbi are 1,900 knowledge-base articles covering topics ranging from parking and abandoned vehicles to serious assault and domestic abuse.
The team has assigned topics to those articles. When a topic is identified during a conversation, it can trigger different instructions, guardrails, and actions.
For example, when a domestic-abuse topic is identified, the agent is instructed to move toward help and support using a more supportive tone. If the language used suggests the person may be a child—using words such as “mommy” or “daddy”—the forces have instructed the agent to adapt its language toward a reading age of ten.
“Applying different guidance and instructions for the agent to follow depending on the subject that it’s dealing with. That’s how we did it.”
The forces said this is where the nuance of large language models can be valuable. But the standard they set is not simply a technically correct answer.
They want the interaction to preserve the victim’s agency and provide an appropriate experience for someone who may be contacting the police in difficult circumstances.
