2. Internal Workflows
From manual casework to outcome-driven agents with humans on the loop.
How It (Doesn’t) Work Today
Governments are infamous for their process inefficiencies. Public servants are often their own worst critics: 94 per cent of UK civil servants reported process inefficiencies in a recent survey. Digitisation has often replicated broken, siloed paper processes in digital form, resulting in little more than the ‘PDFication’ of bureaucracy, without meaningful automation, time savings or user benefit.
Many internal workflows are resistant to traditional, deterministic business process automation (BPA). They often require human judgment, discretion, or contextual reasoning, e.g. eligibility decisions, exception handling, or prioritisation. Others are constrained by incomplete or inconsistent data, with critical inputs scattered across legacy registries, PDF attachments, or unstructured case notes. These limitations make end-to-end automation difficult and often impractical when using conventional tools.
A Vision for Agentic Internal Workflows
Generative AI is already being used to augment the internal tasks of bureaucracy, from drafting and analysis to reporting, document processing, and coding. Benefits are even greater when AI assistants can pull on in-house proprietary data. As in any enterprise, this will allow civil servants to work more efficiently and productively across a wide range of responsibilities.
But AI agents’ potential goes beyond individual productivity. AI agents can help fix broken internal workflows and government back office processes, working around inefficiencies without the need to achieve perfect data integration or process standardisation. Their capabilities include:
Working with imperfect and unstructured data: Governments often hold vast amounts of critical information in non-standardised formats, such as scanned PDFs, legacy databases with inconsistent schemas, handwritten notes, or free-text case files. AI agents can be trained to interpret and reason over unstructured, heterogeneous data, reducing the need for large-scale pre-processing or data cleansing.
Automating for outcomes: Agentic government will come into its own as teams of AI agents become more capable of delivering on specific, measurable objectives, enabling them to contribute to achieving high-level outcomes. These agents will orchestrate complex workflows, intelligently breaking down high-level goals into smaller, executable tasks. They will autonomously interact with diverse systems, data sources, and APIs, coordinating across organisational silos. Crucially, they will monitor progress against defined metrics and adapt their approach to stay aligned with the target outcome.
Navigating complexity and intentions: Public sector workflows often involve a degree of discretion, interpretation of complex regulations, or handling numerous edge cases; these are scenarios where rigid, rule-based automation typically fails. Outcome-focused agents offer a new model. Rather than following fixed processes, these agents can be designed to understand policy intent, navigate procedural variations, and apply nuanced judgment to a wider range of situations, with only the most complex or novel cases escalated to human experts. This adaptability is crucial in domains where achieving universal process standardisation is unrealistic.
For instance, agents designed to automate for outcomes while navigating complexity and policy intent might be tasked to:
Approve 90 percent of construction permits within ten days while rigorously enforcing zoning and climate standards. An agent assigned this task would orchestrate the full workflow, from document ingestion and rules-checking to issuing the draft permit, all while adjusting its approach on the fly to meet performance metrics.
Issue new business licences in real time, contributing to a 75 per cent reduction in new business setup times. Here, the agent would coordinate API calls, verify documentation, reconcile data, and send notifications.
Process routine benefit claims with over 97 percent accuracy and generate automated, plain-language explanations for decisions.The agent would handle end-to-end adjudication, flag outliers, and track its clarity and accuracy metrics over time.
Instead of manually pushing cases, civil servants shift to supervisory roles, managing exceptions, overseeing high-value decisions, and defining or refining the outcome metrics that guide automated agents handling routine eligibility checks, approvals, and adjudications.
This shift would also bring important secondary benefits: In areas historically susceptible to corruption or rent seeking, fully logged, agentic decisions aligned with fair and transparent outcomes reduce opportunities for discretionary abuse. At the same time, workflows driven by outcome-based agents can continuously learn and optimise for efficiency, compliance, and fairness based on their performance against the target metrics.
Key Questions
Which workflow earns the ‘first-pilot’ slot, and why? How do we balance high pain points, data readiness, and political capital when choosing the inaugural end-to-end agent deployment?
How should the safety net work at scale? What confidence thresholds, break-glass rules, and escalation paths keep humans responsibly in charge without stalling efficiency gains or inflating costs?
What kind of changes to management practices are needed for the shift to an AI-native operating model? Leaders will need to orchestrate hybrid human-AI teams and focus on their combined effectiveness.
What are the canonical Key Performance Indicators (KPIs) and quantifiable goals for an agentic government? What are the right metrics to measure success in agentic government — in terms of domain-specific outcomes and operational performance, such as time-to-service, cost per transaction, public trust in digital services, and compliance lag (the time between a law’s passage and its full implementation)?
At what point do we consider the agent’s output to be a legitimate public act? In many bureaucracies, proxies (e.g., interns, staffers, accountants) already complete official work that is formally signed by a responsible civil servant. But agentic workflows blur this distinction. What mechanisms will governments need to ensure delegated legitimacy, such as agentic power-of-attorney models, and how should they define when, how, and under what conditions an AI-generated action is considered authoritative and binding?
How to avoid unnecessary overheads? At what point does human-in-the-loop become not a safeguard but an unnecessary overhead, and a costly reflex that undermines agentic efficiency without adding real value? Are there domains where insisting on human review should itself be justified?
Should the ‘right to a human’ be absolute? Or should it come with trade-offs, such as longer adjudication times or a threshold cost, to avoid overuse that could erode the viability of automated systems?