The AI Theatre at DigiGov Expo opened with an engaging session that cut straight to the heart of what many public sector organisations are grappling with: how to move beyond AI experimentation and towards meaningful implementation. Nadav Mordechai, Product and Strategy Director at Elsewhen, delivered a practical and refreshingly jargon-free exploration of Agentic AI and its potential to transform government operations.
Mordechai began with a revealing show of hands that painted a clear picture of where the public sector stands on its AI journey. Whilst virtually everyone in the packed theatre had used ChatGPT or similar generative AI tools, the numbers dropped sharply when asked about Agentic AI specifically. Fewer still had organisations that had experimented with Agentic AI proof-of-concepts, and by the time he asked about multiple systems running in production, only a handful of hands remained aloft. "That's a good starting point," Mordechai noted, acknowledging what many in the room were experiencing: a gap between awareness and action, between experimentation and implementation.
Rather than diving into technical complexity, he established a working definition that would ground the conversation: AI systems composed of agents that behave and interact autonomously to achieve their objectives. The key distinction, he emphasised, is autonomy with purpose, these aren't just tools that respond to queries, but systems with missions to accomplish. He drew a crucial distinction between the personal AI assistants that have dominated headlines, ChatGPT, Gemini, and the like, and Agentic AI systems. "This is the point where we move from employee efficiency to organisational-level automation," he explained. It's a shift from helping one person at a desk work faster to transforming complex processes across entire departments.
For government organisations, Mordechai argued, Agentic AI represents an opportunity to become simultaneously more human, more responsive, and more efficient. This wasn't empty rhetoric but a framework with specific applications: better interactions with the public through understanding sentiment and needs; more responsive policy-making through accessible data insights; and greater efficiency by automating repeatable procedures. The timing, he suggested, is propitious. The UK public sector has been laying the groundwork with AI guidance frameworks, growing skills ecosystems, and increasing familiarity with generative AI tools. Recent research indicates that 90% of public sector organisations are planning to explore or implement Agentic AI within the next two to three years.
The session's strength lay in its concrete example: procurement. Mordechai highlighted that nearly half of UK public contracts face an average 2.5-month delay in awarding. It's not a broken process, he noted, just a slow one, involving creating scoring metrics, reviewing responses, conducting initial sifts, performing deep reviews, scoring individually, comparing results, and seeking consensus. For a team of two or three people, this easily consumes a quarter's worth of work.
He walked through an Agentic architecture designed to address this challenge, breaking down the system into three types of agents. Orchestration agents coordinate the system, ensuring work is dispatched to the right agents at the right time. Worker agents perform specific tasks, one builds the evaluation framework, other scores responses, a third creates summaries. Supervisor agents check outputs for bias, consistency, and quality, a crucial safeguard given concerns about AI hallucinations and biases. Crucially, the system maintains a "human in the loop" at every stage. "There is a person, there is a human that is always able to interact with that process," Mordechai stressed, before demonstrating a live tool that, whilst not "particularly pretty," showed how the weeks-long procurement evaluation process could be dramatically compressed.
The demonstration revealed a straightforward workflow: documents are uploaded, evaluation criteria can be amended by humans, and the system begins processing whilst explaining its actions. The result is a comparative report with star ratings and commentary justifying the scores. "It may not be perfect," Mordechai acknowledged, "but the reason we do that is to speed up that process massively."
With about a third of the session remaining, Mordechai shifted from theory to practice, addressing what he identified as a key challenge: anxiety about getting started and uncertainty about what makes a good Agentic AI candidate. He offered five principles that Elsewhen has developed through its work with clients. Repetitive work provides ideal candidates, tasks that happen frequently and follow consistent structures. "Agents thrive on consistency," he noted. "If it's boringly predictable, it can probably be agentic."
Time-consuming processes often hide in plain sight. Work that doesn't feel broken but simply takes time, dragging across weeks or departments, represents significant opportunity. Clear data is non-negotiable. "No data, no agent," Mordechai stated plainly. "Agents don't guess, they act on real data." Organisations need at least solid foundational data for agents to work with. Tedious tasks, form-filling, scheduling, logging, reconciliation, create hidden drag and undermine morale. "Agents don't replace people," he emphasised, "they replace and remove the friction." Finally, low-risk opportunities make the best starting points. Not everything should be automated, but tasks with lower stakes, no sensitive data, and manageable consequences for errors provide ideal pilot candidates.
In a generous gesture, Mordechai pointed attendees to a freely available workshop framework, a Figma file that organisations can duplicate and use internally. The tool guides teams through evaluating opportunities across the five principles, with continuums for assessing factors like risk, tediousness, and data quality. It also includes space for capturing thoughts throughout the day's sessions. "We would absolutely love the opportunity to discuss those thoughts with you," he added, inviting attendees to visit Elsewhen's booth.
The Q&A session revealed practical concerns about implementation. When asked whether organisations should choose their AI model before designing their agent, Mordechai's advice was pragmatic, start with what's most accessible and build model-agnostic systems that can swap models as the technology evolves. For Elsewhen's procurement tool, they used Gemini, but other tools have used OpenAI or even both simultaneously, with the system deciding which model handles which task. "The space is at that point where you constantly want to evaluate the models you use," he noted.
When pressed on whether humans would eventually be removed from the loop, Mordechai was measured in his response. Involvement might change, and confidence in systems will grow, but wholesale removal seems unlikely. He drew parallels to financial services, where loan underwriting moved from relationship-based decisions to automated systems for smaller tickets whilst retaining human oversight for larger, riskier decisions. "I think there's quite a few pilots in that story," he observed.
As the opening session concluded, it left the audience with both inspiration and practical direction. The message was clear: Agentic AI isn't science fiction or a distant future, it's available now for organisations willing to start small, choose the right use cases, and build with humans firmly in the loop. For a room that began with few hands raised about Agentic AI experience, many left with a clearer sense of where to begin.
Liuba Pignataro
