1. Service Delivery and User Experience
From fragmented portals to self-composing public services and
personal concierges.
How It (Doesn’t) Work Today
Users of public services, citizens and businesses alike, navigate fragmented portals, websites and forms, triggering digital services with varying levels of automation. For most public services, the user experience (UX) tends to be transactional and brittle. Standard cases follow rigid workflows, while anything outside the norm often results in manual handling and bureaucratic back-and-forth.
Some countries’ public authorities have made significant investments into improving government UX over the past decade (such as design systems, modern user interfaces, user testing, product management, proactive services). The results of all this have been mixed. While there have been some great successes, such as the Gov.UK design system, there have also been failures and stalled efforts.
Recently, designing user-centric life events is becoming part of the standard playbook of digital government. This approach organises public services around key moments in people’s lives, such as having a child, starting a job, or retiring, rather than around institutional structures. The goal is to deliver seamless, end-to-end experiences that reduce complexity for users.
While effective for capturing broad, predictable stages of life, the ‘life events’ model misses the granular, messy, and often urgent needs that define real user journeys. For instance, a ‘birth of a child’ life event may help parents register the birth and apply for leave or benefits. But it fails to adapt to more complex scenarios, such as a single non-citizen mom giving birth in a different jurisdiction, while facing housing insecurity and urgent healthcare needs. These edge cases typically fall outside the scope of predefined service bundles, forcing citizens to navigate disconnected systems on their own.
Generally, the current manual approaches to service design do not scale. Furthermore, good design requires significant upfront investment, particularly when coordination is needed across multiple layers of government and external partners. Labour is the main cost driver here. Even in small and digitally advanced countries like Estonia, revamping services for major life events, such as the birth of a child or business compliance, has cost millions of euros to design, build, and launch. These costs limit how broadly such efforts can be repeated worldwide.
A Vision for Agentic Service Delivery and UX
Digital-era UX limits users to interactions and service definitions defined in advance by humans, bottlenecked by the cost and friction of the (re)design process. Generative AI, by contrast, can compose novel interactions and services addressed to the specific needs and limitations of each user.
In contrast to the current, static ‘life events’ model, agentic AI unlocks the potential to go beyond this one-size-fits-all framework by dynamically composing services tailored to the individual’s situation, preferences, and constraints. The widespread use of agents, whether directly by users or on their behalf by public authorities, will enable endlessly customised services at zero marginal cost.
In this new paradigm, agentic public services will be characterised by:
Personalisation and multi-modality: As a first step, service interactions will become more tailored to individual preferences. User interactions can be customised in real time to any language and offered consistently across different modalities, channels and devices (e.g. browser, app, text message, phone call, video avatar, AR/VR headset, brain-computer interface). Tone and emotion can be adjusted to circumstances and cognitive styles (warm and effusive conversation for one person, austere recitations of rules and processes for another). Multimodality will support universal coverage (e.g. monitoring detects a user struggling partway through an online benefit application and proactively offers a support call).
Proactivity: Proactive government services today usually rely on fixed rules or triggers, such as sending reminders when a deadline is approaching. By contrast, agentic services will be far more responsive and personalised. For instance, an AI agent could identify individuals who would benefit from a new training programme and contact them directly. Over time, the system could learn which types of messages work best for different people and adapt its approach accordingly.
Ideally, AI-driven public services should help close social and economic gaps. Unlike human systems, AI Agents can apply inclusive rules more consistently, taking into account things like different levels of digital skills, disabilities, cultural backgrounds, or income. With the right safeguards, they can also be designed to detect and correct unfair patterns as they go.
Agents can also unlock a form of public sector cross-selling. In contrast to the private sector, which routinely leverages touchpoints to suggest relevant actions, services, or upgrades, governments typically treat interactions as single-purpose. But agents, operating with contextual awareness and public intent, could prompt citizens toward beneficial next steps: notifying them of available programs they are likely eligible for, inviting participation in local decision-making, or surfacing rights and responsibilities they may not be aware of. A request for a permit could also prompt an offer to update business registry information; a tax filing could be paired with a participatory budgeting invitation. Done well, this kind of value-aligned upselling could enhance inclusion, transparency, and engagement, without becoming intrusive.
Self-composition: Instead of hand-building 10 or 30 standardised life events, public administrations will be able to compose tailored services for each citizen and business. Every life path is unique: a home birth followed by an emergency hospitalisation in a rural municipality is very different from an unmarried refugee giving birth in a city hospital; similarly, the insolvency of a startup founded by a serial founder differs profoundly from a multi-generational family firm going out of business during a recession.
On top of defined rules (i.e. laws and regulations) and resources (i.e. budgets, support thresholds, and reimbursement amounts), agents will be able to compose public, and also private, microservices into seamless, end-to-end offerings. Citizens simply state intent, and agents resolve full service flows dynamically. For example: “I need help after my house was damaged in a storm” could trigger an agent dynamically assembling the relevant services, including filing insurance claims, applying for relief funds, scheduling inspections, by connecting banks, insurers, construction firms and government agencies in one unified process.
Key Questions
Is this a continuation or a reset? Does it make sense to continue along the familiar linear path of UX, refining life events one workflow at a time? Or does agentic AI require a radical reset, starting from scratch to avoid perpetuating legacy assumptions?
Who will build the agents, and who will govern them? Multiple models will likely coexist. Governments could develop their own trusted agents on behalf of citizens. Alternatively, users may rely on personal agents built by large platforms such as OpenAI, Google and Apple, or they may run their own. Should governments treat these agents as endpoints, like browsers, or build for direct orchestration?
What happens when only some citizens can afford the best agents, trained on privileged data and optimised through constant feedback? Agentic systems raise critical questions about equity. If only some citizens can afford high-quality agents, will access to public services become unequal? What obligations do governments have to ensure fair access, and what technical or institutional tools could guarantee this? Should governments provide open APIs, agent testing environments, or even baseline public agents to prevent capability capture by private actors?
How do we avoid going too fast or over-promising? What causes early AI deployments to falter, and how can governments avoid similar pitfalls in critical public functions?
What developments will help agentic UI and services comply with high legal standards for public services? What can we learn from risk aversion and guardrails (in the form of e.g. filters, templates, or silos) that have limited the usefulness of user-facing public sector Large Language Model (LLM) deployments so far?