Every five minutes, a UK police call handler picks up a 101 call that should have gone to someone else entirely. Multiply that across the 700,000+ non-emergency calls a single force handles in a year, and you get a control room buried under questions an AI could answer in seconds.
Non-emergency lines are the front door of policing, and right now that door is jammed. Call handlers spend entire shifts repeating the same answers to the same questions, while citizens sit on hold for information that's already sitting in a knowledge base somewhere. There's a better way: let an AI agent handle the repetitive share of contacts, and free human officers for the calls that actually need judgement.
Thames Valley Police, Hampshire & Isle of Wight Constabulary, and Humberside Police built exactly that. In its first week live, their AI agent — named Bobbi — was used by more than 600 citizens and resolved 82% of inbound queries without escalating to a human officer. Here's how three regional forces built and deployed it in weeks, not months.
What Bobbi Actually Does (And Why It's Not Just Another Chatbot)
Bobbi is an AI agent built on Salesforce's Agentforce 360 Platform and deployed as a chat widget — a blue speech bubble — on each force's website. It's grounded entirely in "closed-source" content: the same internal FAQ and policy knowledge base that Call Handlers and Digital Desk Operators already reference, not public web data scraped from elsewhere. That distinction matters for accuracy and accountability in a policing context, where a wrong answer has real consequences.
Bobbi answers non-emergency questions directly, walks citizens through existing self-service forms (for example, reporting lost property), and is hard-blocked from two things by design: it cannot be used to report a crime, and it is not a substitute for the 999 emergency line. Either request gets redirected at the first message, before the conversation goes any further.
If you run a 101 contact centre, a council helpdesk, or any public-facing service drowning in repetitive FAQ-style contacts — housing repair requests, council tax queries, library and leisure-service questions — this is the playbook. It isn't built for life-and-death triage; it's built for the dozens of "how do I…" and "what happens if…" questions a knowledge base already answers, just more slowly than an AI agent can.
How It Works in Practice — No Hidden Steps
Because this is a real-world deployment rather than a tutorial, there's nothing held back here — this is the full mechanism, end to end:
- Citizen opens the chat widget — the blue speech-bubble icon on the force's homepage. No app download, no account, no login.
- Intent triage — Bobbi reads the question and checks it against two paths: in-scope (covered by the closed-source knowledge base) or hard-routed (crime report, emergency).
- Resolve, guide, or hand off — in-scope questions get answered directly or routed through an existing self-service form; anything Bobbi can't handle, or any citizen who asks for "a person," is handed live to a Digital Desk Operator within the same chat thread, with full context carried over.
That's the entire system. There's no multi-month software build behind it — three forces went from concept to a public-facing AI agent in a matter of weeks, largely because Agentforce 360 sits on top of knowledge content the forces already owned rather than requiring a new platform to be built from scratch.
The Results: 82% Resolved, No Human Required
Salesforce's own December 2025 case study put hard numbers on the rollout: 82% of inbound queries resolved with zero escalation to a human officer, across more than 600 conversations in the first week alone. Some force-level reporting puts the resolution rate for purely non-emergency questions as high as 90%.
Other public-sector deployments tell a similar story at different scale. Williamsburg, Virginia's resident chatbot resolves 79% of questions on first contact, and Raleigh, North Carolina's city chatbots now handle 90% of calls into administrative departments. That consistency across three different government bodies on two continents suggests this isn't a one-off — it's a repeatable pattern wherever contact volume is high and the answers are already documented.
The gap between adopters and laggards is starting to show. An IDC survey commissioned by Salesforce found that 82% of government organisations have already adopted AI agents in some form, and most of their leaders now believe they're ahead of the private sector on agentic AI — a reversal of the usual public-sector technology lag. Forces still routing every 101 call through a human are competing for the same stretched headcount as everyone else, while early movers free up officers for the calls that actually need a badge.
The Tech Stack Behind Bobbi
This isn't a multi-vendor build — Bobbi runs on a single platform layered on top of content the forces already owned.
| Tool | Role in This Deployment | Public Pricing? | Paid From |
|---|---|---|---|
| Salesforce Agentforce 360 Platform | Builds and hosts the conversational agent, runs intent triage, executes live handoff to a human operator | Partially — general rate card published | $500 per 100,000 Flex Credits (~$0.10/action) or $2/conversation flat |
| Existing force knowledge base | Grounds every answer Bobbi gives — the same closed-source content Call Handlers reference internally | N/A — internal asset, not purchased | — |
| Force website chat widget | Entry point for citizens — an embedded widget, not a new app | Included in deployment | — |
⚠️ [REQUIERE VERIFICACIÓN] The pricing above reflects Salesforce's general published Agentforce rate card. Police forces typically negotiate public-sector or government-cloud terms that aren't published, so treat these figures as a ceiling — not necessarily what Thames Valley, Hampshire, or Humberside actually pay.
Who Should Use This — And Who Shouldn't
This model fits any public-facing service line where contact volume is high, the answers already live in a documented FAQ or policy, and only a minority of contacts genuinely need human judgement or empathy. Council tax helplines, housing repair requests, and other force control rooms are obvious candidates — as is any 101-style non-emergency line drowning in calls that, per West Yorkshire Police's own data, shouldn't have reached them in the first place (the force logs more than 100,000 such misdirected calls a year, one roughly every five minutes).
It is not a fit for anything safety-critical or emergency-facing. Bobbi itself is explicitly blocked from crime reporting and 999-equivalent requests, and any deployment modelled on it needs to preserve that same hard boundary — the goal is deflecting routine contacts, never replacing emergency response.
Frequently Asked Questions
Is Bobbi a real police officer?
No. Bobbi is an AI chat agent, not a sworn officer. The three forces are explicit that it cannot be used to report a crime or replace the 999 emergency line — it only handles non-emergency, FAQ-style questions, and hands off to a human Digital Desk Operator whenever it can't help.
Which forces use Bobbi, and could other forces deploy something similar?
Thames Valley Police, Hampshire & Isle of Wight Constabulary, and Humberside Police built and launched Bobbi together. Because it runs on Salesforce's Agentforce 360 Platform rather than bespoke software, any organisation already using Salesforce — another force, or a local council — could deploy a comparable agent on its own knowledge base.
How long does a deployment like this actually take?
Reported as "just weeks" from build to public launch — fast for a public-sector technology rollout, mainly because the underlying knowledge content already existed. The work was wiring an AI agent to retrieve and present it accurately, not authoring new policy from scratch.
Want more real-world AI deployments like this one?
Sityos AI publishes a new automation tutorial or case study every week — covering everything from enterprise AI agents to the tools small teams can deploy this month. Explore the full archive at sityos.com.
Three regional forces just answered a national question: agentic AI in public service doesn't require a multi-year transformation programme. The technology is already sitting on the shelf — the only variable left is which control room wires it up next.