Public sector organizations are beginning to move AI beyond pilots and into operational environments, with early deployments focusing on frontline support, citizen engagement, and workforce productivity.
Speaking during a panel at the opening of Salesforce’s new London AI Center, representatives from policing and the National Trust discussed how agentic AI is being applied in public-facing and public-benefit services, while also acknowledging the need for human oversight and governance.
The discussion comes as Salesforce positions its Agentforce platform as a way for organizations to deploy AI agents across customer, employee and citizen service workflows. However, the public sector use cases highlighted during the panel suggest that adoption is not simply a matter of automating processes. Leaders need to carefully define where AI can support staff with improved access to information while maintaining consistent service standards.
AI as a Support Tool for Frontline Policing
One of the clearest examples came from Claire Hammond, Temp Detective Chief Superintendent at the National Centre for Violence Against Women and Girls and Public Protection (NCVPP), who discussed the development of “Ask Em”, an AI agent designed to support frontline officers dealing with violence against women and girls and public protection cases.
Hammond said the challenge for policing was how to make specialist knowledge available to officers at the point of need, particularly when they may be working alone or outside regular hours.
“What you find in policing is at 3 o’clock in the morning when you’re sat in a lonely room… you’re faced with a problem and there’s no one to ask,” she said.
Ask Em is intended to act as a source of guidance, giving officers access to the right information and language when dealing with victims. “Ask Em has become that person,” Hammond said.
“We’re putting in all of the tools that they need, the frontline officers, but in an empathetic language, so that they get the right language back for how to speak to a victim and what they should do.”
For Hammond, the purpose is to improve consistency across forces and situations.
“Victims can expect a level of service for all points, and no matter what time of day it is, or which station you go to, you get the same answer.”
That consistency is a central challenge for public services, where outcomes can depend on the experience, confidence, or availability of individual staff members. AI agents, when carefully constrained and governed, may help reduce those gaps by making approved guidance easier to access.
Human Oversight Remains Central
Hammond was clear that policing remains cautious about where AI should and should not be used, particularly where decisions may be scrutinized in court or have safeguarding implications.
“In policing we’re culturally quite risk-averse. We very much like a human to make that final decision.”
That is especially important in public protection, where decisions can involve risk and vulnerability.
“We work on something called the National Decision Making Model, so that process will never be replaced in the sense of there’s always going to be a human… at the center of that.”
Hammond said the assurance model for Ask Em depends on controlling the information the system can draw from, adding that the system can help gather and compare information, but should not take over decisions that require human judgment.
“We’re not going to get it to make risk decisions. And we’re not going to get it to make safeguarding decisions, because people need to see the wider picture.”
That distinction is likely to become increasingly important as public sector organizations consider agentic AI. The strongest early use cases may be where it supports workers with information, consistency and process navigation.
The National Trust Explores AI for Personalization and Public Engagement
Jon Townsend, Chief Information Officer for the National Trust, also discussed how AI could support large-scale engagement with members, visitors and the wider public.
Although the National Trust is not a government body, it operates in a public-benefit context, managing heritage, environmental, and cultural assets.
“The challenge that we have is, to reach all of those people and audiences… that involves us needing to think about how do we deliver that sort of hyper-personalized messaging for people at scale,” Townsend said.
Agentic AI is expected to play a role in that engagement, particularly where people want to interact with the organization in a more conversational way.
“When people want to engage with us… they can do that using agentic AI in a way that it feels like they’re engaging with a person,” Townsend said.
However, there is a more complex issue for organizations that publish large volumes of information online: how AI systems interpret and present that information to users.
“The big challenge that I see for us is not about what we as an organization think about AI. It’s actually, what does AI think about us?”
Organizations need to understand how AI models present their content to the public, particularly as users increasingly receive answers directly from AI systems rather than clicking through to websites.
That raises the question, “how are the different AI models presenting our content to people in an accurate way, so that they can consume it without necessarily having to click through to our website?”
For public sector and public-benefit organizations, that also has implications for the accuracy and control over information. As AI interfaces become a more common route to services and knowledge, organizations may need to manage how their information is represented by third-party models and agents as well as their own digital channels.
Grounding AI in Organizational Knowledge
Townsend also discussed the importance of grounding AI in accurate, up-to-date organizational knowledge and linked this to the idea of a “human at the helm” approach, where human teams set the rules and context within which AI operates.
Townsend pointed to Salesforce’s Agentforce as enabling users to separate their choice of AI model from their customer data, while still using organizational context to shape responses and maintain a trust layer.
“That contextualizing of large language models and grounding it within your organization, that’s the secret sauce for me.”
This issue is particularly sensitive for public-sector organizations, many of which have use cases that involve personal data, vulnerable users, or high-trust interactions. AI adoption will depend on data protection and the ability to explain how outputs are generated as much as capability.
Townsend warned that privacy decisions should be treated carefully. “Privacy rights are easy to give away and very, very hard to get back.”
Salesforce Calls for AI Adoption to Reimagine Public Sector Processes
Paul O’Sullivan, UKI Chief Technology Officer & Head of UK AI Center at Salesforce argued that the value of AI depends on whether it reaches the people who need it in their day-to-day work.
O’Sullivan pointed to examples such as Heathrow’s AI agent and Ask Em as evidence that AI agents are becoming more context-aware and useful in real-world service environments and emphasized that successful AI adoption depends on process redesign, rather than applying new tools to old workflows.
“Reimagine your processes. Don’t just apply AI to the same stuff you’ve always done.”
That point is particularly relevant in the public sector, where processes are often shaped by legacy systems, regulatory requirements and organizational silos. AI may expose inefficiencies, but it will not automatically resolve them unless services are redesigned around users and staff.
Despite the optimism around AI’s potential, O’Sullivan also warned against over-reliance on AI outputs without scrutiny: “Critical thinking is absolutely fundamental.”
There is a risk of “cognitive offloading,” where users accept AI-generated answers without understanding how they were produced.
If AI is used to support decisions, draft responses or guide frontline workers, organizations will need to ensure that staff can challenge outputs, understand their limitations and explain decisions when required.
A Cautious but Practical Path Forward
The panel suggested that public sector AI adoption is entering a more practical phase, with use cases emerging around information access, service consistency, workforce support and citizen engagement.
But the discussion also made clear that adoption will need to be carefully governed. The most credible use cases are likely to be those where AI supports human workers rather than replaces them; where systems are grounded in approved organizational knowledge; and where clear boundaries are set around risk, safeguarding, privacy, and decision-making.
As Townsend put it:
“These technologies offer a fantastic opportunity to deliver more public benefits, and I think that’s the key… people and AI working together to deliver more for other people.”
That framing has emerged as the rallying call for adopting AI in the public sector, using it to help stretched services provide more consistent and accessible support while keeping humans accountable for the decisions that matter most.
