{"resourceId":"tvp-hiowc-bobbi-ai-agent-public-contact-20261005","versions":[{"version":"external-351e1a2e14964a8e03d5cf7d941026fcaaa729229eab14a999617552fc805bc9","resource":{"id":"tvp-hiowc-bobbi-ai-agent-public-contact-20261005","title":"Thames Valley and Hampshire police AI agent Bobbi automates about 75% of non-emergency contacts while escalating vulnerability cases to humans","organization":"Thames Valley Police and Hampshire & Isle of Wight Constabulary (reported by CX Today; vendor case study by Salesforce)","sector":"Police public contact / non-emergency services","geography":"Thames Valley and Hampshire, England, United Kingdom","publishedAt":"October 5, 2026","publicationDate":"2026-10-05","eventDate":null,"sourceName":"CX Today","sourceLabel":"CX Today interview: How Thames Valley Police is using AI to flag vulnerability","sourceUrl":"https://cxtoday.com/how-thames-valley-police-is-using-ai-to-flag-vulnerability","evidenceClass":"independent-reporting","outcomeClass":"emerging","topics":["developers-agents","knowledge-work","governance-procurement","accessibility-workforce","operating-model"],"finding":"CX Today reported on Oct. 5, 2026, from a Dreamforce 2026 interview, that Thames Valley Police and Hampshire & Isle of Wight Constabulary's public-facing AI agent Bobbi has been live since the previous November, with just over six months of data at interview time. The forces said about 75–76% of contacts are subject to automation and the remaining 25% are escalated to operators on a 24/7 digital desk. Topic-tagged knowledge articles trigger different instructions and guardrails (for example a supportive tone for domestic abuse and simpler language when a child appears to be writing). The agent was stood up in four days and taken live within ten weeks; the team tested about 3,000 conversations before deployment and continues to monitor and amend it, treating it 'like an employee' with clear accountability. The forces report fewer contact-centre contacts, reduced waits and higher satisfaction, and cite a cuckooing case that led to four arrests. Salesforce's vendor case study claims 200+ daily conversations, 75% resolved independently, a 97% lower cost per interaction, a 4.6 CSAT score, and about two potential violence-against-women-and-girls cases routed daily. Next steps under consideration include having the agent take actions on users' behalf and an internal staff agent.","sledRelevance":"Policing public-contact lane: a generative AI agent handling non-emergency police enquiries with explicit human escalation for vulnerability. Cross-tagged emergency-services because it diverts 101 demand from police call handlers, though it is not a 999 dispatch tool.","evidence":"Trade-press interview with force digital leaders (operator claims) plus a visibly attributed Salesforce customer story (vendor claims). No independent evaluation, missed-vulnerability error rate or escalation accuracy is published.","architectureImplications":"Retrieval-grounded agent over force-approved knowledge articles, with topic-triggered instructions, case creation and live handoff passing transcript, risk indicators and summary to operators.","governanceImplications":"Accountability is framed as ownership of the agent's knowledge and behaviour, with pre-launch testing and ongoing monitoring; planned moves toward agents taking actions will need new authorization and audit controls.","securityPrivacyImplications":"The handoff passes contact details, IP address and transcripts to operators; vulnerability detection on chat text raises retention and use questions for sensitive disclosures that the sources do not address.","caveats":"Figures are operator and vendor claims without independent verification; CX Today cites 1,900 knowledge-base articles while Salesforce cites 91, an unexplained discrepancy; no false-negative rate for missed vulnerable contacts; ROI and staff-hours figures are projections; UK context limits direct U.S. transferability.","streamIds":["public-safety","emergency-services"],"roles":{"sales":"Sales framing: Bobbi is a reference case for AI agents in police non-emergency contact, but its value rests on routing the right 25% to humans. Discovery should cover current non-emergency volume and wait times, how vulnerability is defined, and who owns knowledge content. Offer a bounded pilot with escalation-accuracy testing; do not cite the 97% cost reduction or ROI projections as expected outcomes, since they are vendor claims.","engineering":"Engineering framing: this is retrieval-augmented generation over agency-approved articles with topic-based guardrails and a live handoff to operators. Engineers must design vulnerability detection with measured recall, structured handoff payloads and logging of every escalation decision. Prerequisites are a curated knowledge base and a staffed digital desk. Proof of value should test missed-vulnerability rates on labeled conversations before launch, mirroring the forces' roughly 3,000-conversation pre-launch test.","delivery":"Delivery framing: run the agent like a staff member, with a named owner, content governance, routine quality review and an escalation desk staffed around the clock. Acceptance should rest on escalation accuracy, wait-time change and complaint trends, reported publicly. Risks include missed vulnerability, over-automation as agents begin taking actions, and the knowledge base drifting out of date. Any move to action-taking should require a separate approval."},"roleDetails":{"sales":{"customerProblem":"Police non-emergency lines face long waits and high volumes, while vulnerable callers must still reach a human quickly.","stakeholders":["Force contact-centre leadership","Digital and data leads","Public protection and safeguarding units","Victim services","Police and crime commissioner","Data protection officer"],"discoveryQuestions":["What share of non-emergency contacts are status requests or FAQs today?","How is vulnerability defined and labeled for testing?","Who owns and approves the knowledge content?","Is a 24/7 human desk available for escalations?"],"valueHypothesis":"Bounded: an agent may absorb routine enquiries and shorten waits if escalation accuracy is demonstrated on local data.","potentialEngagement":"Non-emergency AI agent pilot with a labeled escalation test set, monitoring dashboard and governance runbook.","unsupportedClaims":["97% cost reduction or 900% ROI as expected local results","Reliable detection of all vulnerable contacts","Transferability to U.S. non-emergency lines without testing","Crime-reduction effects"]},"engineering":{"fit":"Agencies with web or chat non-emergency channels and a curated knowledge base that can staff a live escalation desk.","architectureIntegration":"Public chat to retrieval over approved articles to topic-triggered guardrails to case creation and live operator handoff with transcript, risk flags and summary.","prerequisites":["Curated and tagged knowledge articles","Labeled vulnerability test conversations","24/7 operator desk","Conversation and escalation logging"],"constraints":["Not an emergency (999/911) channel","UK data protection rules on sensitive disclosures","Knowledge base must be maintained","Planned action-taking needs new authorization"],"security":"Minimize personal data captured, protect transcripts containing abuse disclosures, and log every automated answer and escalation decision.","proofOfValue":["Escalation recall on labeled vulnerable conversations","Wait-time change against baseline","Share of automated answers rated correct in quality review","Complaint and re-contact rates"]},"delivery":{"work":"Stand up content governance, a labeled test set, monitoring reports and a staffed escalation desk before launch.","dependencies":["Knowledge content owners","Contact-centre staffing","Vendor platform configuration","Safeguarding team input on topics"],"ownership":"Force contact management and digital leadership, with safeguarding leads owning vulnerability rules.","skillsAdoption":"Operators need training on reading handoff summaries and risk flags; content owners need skills in writing agent-ready articles.","governanceCheckpoints":["Pre-launch conversation test sign-off","Monthly quality review","Safeguarding review of missed escalations","Approval before any action-taking capability"],"acceptanceCriteria":["Proposed: escalation recall above an agreed threshold on labeled tests","Proposed: published monthly automation and escalation rates","Proposed: every knowledge article reviewed within a set period","Proposed: measured wait-time reduction versus baseline"],"risks":["Missed vulnerability in automated chats","Over-reliance on vendor metrics","Knowledge drift","Scope creep into action-taking agents"]}},"retrievedAt":"2026-10-06T02:52:34Z","enrichedAt":"2026-10-06T03:01:43Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Offers a lower-barrier text channel (including adaptive reading level) and shifts operators toward higher-harm conversations; workforce effects beyond reduced contacts are not measured.","procurementImplications":"Agencies should require vendors to report escalation accuracy and allow agency control of knowledge content, rather than buying on vendor ROI claims.","operatingModelImplications":"Contact handling becomes agent-first with human escalation, and the agent is managed like staff through monitoring, coaching and content updates.","sourceVerification":{"openedUrl":"https://cxtoday.com/how-thames-valley-police-is-using-ai-to-flag-vulnerability","referenceExcerpt":"The team tested around 3,000 conversations before deployment.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}