Will AI Speed Up the Way We Work — or Change It?

Public sector AI is already improving productivity. Across government, health, education and local services, teams are using it to draft, summarise, search, classify and respond faster than before.

But there is a risk in how this progress is being applied.

The output is faster. But the work has not really changed.

That is the strategic question now facing public sector leaders: will AI simply accelerate existing workflows, or will it change how work happens in the first place? 

• The productivity trap
  
• The next phase is changing the shape of work
  
• The citizen is not one person
• Start with the service problem
• Internal services matter too
• Data matters, but knowledge is often the first unlock
• Why pilots really stall
• Scale patterns, not pilots
• Measuring what matters
  
• The leadership questions
  
• What comes next
  

The productivity trap

The first wave of AI adoption has understandably focused on individual productivity. And the results are real: people are drafting faster, summarising meetings, retrieving information and structuring ideas in ways that would have taken much longer before.

But there is a structural problem that individual productivity tools do not solve — and in some cases, do not even touch.

In most organisations, people have quietly become the integration layer between systems. Look at how knowledge work actually happens. A task begins in one system. Someone checks another for context. Information gets copied into a document. A message goes to a colleague. A ticket gets updated. A system of record gets amended manually. None of these steps are particularly difficult on their own. But together they create an enormous amount of invisible work: the coordination, the chasing, the copying, the context-gathering. Work about work.

This is the job nobody describes in job descriptions. But it consumes a significant share of the working day.

The public sector version is familiar. A caseworker navigates multiple systems to build a picture that should already exist in one place. A contact centre advisor repeats information a citizen has already provided. A finance team spends time chasing approvals that could have been automated. A policy team manually compiles updates from sources that could be monitored automatically. These are not technology failures. They are operating model failures — and they have accumulated over years of software systems that were never designed to talk to each other.

When AI arrives into this environment, it often helps people perform the middleware role faster. It can summarise the context someone has gathered across four systems. It can draft the message that updates a colleague. It can speed up the output at the end of a fragmented process. But the fragmentation itself remains.

That is the trap. If AI is deployed inside broken workflows, it makes broken workflows faster. The contact centre advisor still absorbs demand that should never have arrived. The caseworker still bridges systems that should be connected. The citizen still chases an update that should have been proactive. The organisation gets busier... but it does not get better.

The more interesting question is not how to help people perform the integration work faster. It is how to remove the need for that integration work in the first place.

That requires a different kind of ambition. Not AI as a better tool for individuals. AI as a way of changing the shape of the work itself.

The next phase is changing the shape of work

The greater opportunity comes when organisations use AI to change the workflow itself.

That means looking beyond individual tasks and asking where AI agents could reduce avoidable demand, improve access, remove unnecessary hand-offs, support decisions, trigger next steps and make services easier to navigate.

This is the more meaningful version of agentic AI.

Not “where can we put a chatbot?”

Not “how do we give everyone another assistant?”

But “where is the work itself creating friction, and how could an agent help redesign the interaction?”

Agentic AI should not be understood as just another interface. It is better thought of as a capability: AI that can use approved knowledge, follow rules, work within guardrails, take defined actions and escalate to humans when needed.

That might mean helping a citizen understand what evidence is missing from an application. It might mean guiding someone to the right reporting route. It might mean helping an advisor respond with the right policy and context. Or it might mean routing a complex case to a human with a clear summary of what has already happened.

But people do not experience “agentics” in the abstract. They experience a surface.

That surface might be Slack, email, WhatsApp, SMS, a website, a portal, a contact centre console, a case management system or a mobile app — wherever the interaction already happens.

The agent is the engine. The surface is the experience.

That distinction matters because the surface is not a cosmetic choice. It is a service design choice.

A citizen does not care which model is being used. They care whether the interaction is clear, trusted, accessible and useful. A member of staff does not care about the architecture in the abstract. They care whether the system helps them resolve the issue safely, accurately and with less friction.

The citizen is not one person

We should also be careful with the word "citizen". It can make very different experiences sound the same.

A digitally confident small business owner checking a regulatory query, a carer managing a relative’s health appointment, a victim of crime seeking reassurance, and a vulnerable claimant navigating a complex benefits decision do not have the same needs, risks or trust thresholds.

This is not a marginal point. It is central to whether AI-enabled services succeed or fail.

If surfaces are designed for the median user, they will systematically exclude those with lower digital confidence, higher vulnerability or more complex needs. That is where trust breaks down, and where demand returns to more expensive channels.

Surface design is therefore not just about convenience. It is about accessibility, confidence, urgency, safety and human judgement.

A good public service experience does not force everyone down the same route. It helps people use the right route for their situation, and it makes escalation clear when the stakes are higher.

Start with the service problem

The best examples do not start with the technology. They start with a service problem.

A useful example is the work with Thames Valley Police and Hampshire & Isle of Wight Constabulary. The story did not start with an AI agent. It started with a very human issue: victims of crime needed better visibility, and contact centres were absorbing avoidable demand from people asking for updates.

The scale of the problem was clear. The forces were receiving over 400,000 non-emergency calls a year, nearly half of them from people seeking case updates. The first redesign focused on transparency and communication — giving victims a secure way to track cases, receive updates and message officers.

That change mattered. By making updates visible and self-service, call volumes reduced by 10%, with £1.4m in cost avoidance from self-service updates and a 97% reduction in cost per interaction.

Agent Bobbi then became the next logical step: a digital front door for non-emergency policing, grounded in approved knowledge, designed with guardrails, and built to escalate high-risk cases to human operators.

Bobbi now handles roughly 200 non-emergency conversations a day, with many resolved without human intervention. But the most important lesson is not simply efficiency. It is accessibility.

In areas such as violence against women and girls, a discreet digital surface can help people ask questions, validate their experience and take a first step towards support in a way that may feel safer than picking up the phone. Bobbi carries a 4.6 out of 5 customer satisfaction score, and on average two VAWG cases are identified and routed for human intervention daily.

That changes who engages, how they engage, and when human help can be brought in.

Notice what this example is actually about. The first unlock was not data transformation for its own sake. It was making knowledge, status and next steps visible at the moment people needed them. Victims were not chasing because data was missing — they were chasing because they could not access the knowledge of what was happening, what would happen next, and what they should do. The self-service update system made that knowledge visible. Agent Bobbi made policy and process knowledge accessible in a discreet, trusted way. The measurable impact — on demand, on access, on VAWG identification — came from knowledge availability.

This is not just "an AI agent answering questions". It is service redesign, then agentic capability, then safer escalation — with measurable impact on demand and access.

Internal services matter too

The same principle applies inside public sector organisations.

NHS Shared Business Services is a useful example because it is not primarily about citizen-facing AI. It is about changing how work happens in shared services that keep the NHS running.

Invoice queries, finance processes and procurement support may not sound like the front line. But when they are slow, fragmented or hard to navigate, the pressure is felt across the system. Staff chase. Suppliers wait. Queries are repeated. Time is lost. Frontline resources are indirectly affected.

NHS SBS processes £395bn of NHS funding each year, so small improvements to the operating model can matter at national scale. Through its digital help centre, 84% of NHS staff queries are now initiated through the platform. Average handling times have fallen by 20%, most queries are resolved within 24 hours, and the average time to raise a query has fallen from 12 minutes to three.

The important shift was not simply introducing AI. It was redesigning how NHS staff and suppliers raise, track and resolve queries — combining self-service, automation and human expertise so that routine work can move faster and more complex issues can reach the right people with better context.

And again, the first unlock was not data transformation for its own sake. Staff were not blocked by missing data — they were blocked by not knowing where to go, what to ask, and how the process worked. The digital help centre made operational knowledge navigable. That is what reduced query time from 12 minutes to three.

This is the internal version of the same strategic fork. AI can help an individual respond faster, or it can change how the work moves through the organisation.

Both matter. But they are not the same.

Data matters, but knowledge is often the first unlock

There is rightly a lot of focus on data foundations. Public sector organisations need to understand their data, permissions, security, auditability and governance if they are going to scale AI responsibly.

But there is a risk that the conversation jumps too quickly to data nirvana.

Many valuable AI opportunities begin with making organisational knowledge usable. Public services hold vast amounts of knowledge — from policy and guidance to eligibility rules, process steps, service standards, previous decisions and institutional memory.

Citizens and staff are often trying to extract that knowledge at the moment they need it. They want to know what they are eligible for, what evidence is needed, what happens next, whether an application has been received, who owns a case, or what to do if circumstances change.

When people cannot access that knowledge easily, demand increases. Citizens call. They email. They chase. They repeat information. Staff search manually. Cases slow down. Confidence falls.

In many services, failure demand is not created because citizens want something new. It is created because they cannot get the knowledge, status or reassurance they need from the service they are already using.

Data helps determine status, eligibility, personalisation and action. Knowledge helps people understand what is true, what applies and what to do next.

Both examples in this piece illustrate the point. In neither case was data transformation the first unlock. Both Thames Valley Police and NHS Shared Business Services started by making knowledge visible, accessible and usable at the moment people needed it. The measurable results followed from that.

Public sector AI needs both.

Why pilots really stall

There is nothing wrong with pilots. Public sector organisations should test, learn and manage risk carefully.

The issue is not that pilots exist. The issue is that many pilots are not designed with a route to service change.

Sometimes they are demo-led rather than service-led. Sometimes they rely on a champion who moves on. Sometimes the knowledge source is not production-ready. Sometimes procurement, security or governance slows the route to live. Sometimes legacy systems cannot connect. Sometimes nobody owns the operating model change. Sometimes success is measured by adoption rather than service impact.

And sometimes the hardest part is human.

Staff concerns about job security, accountability and professional judgement are not abstract. They shape behaviour. If AI is introduced without clarity on roles, decision boundaries and accountability, staff will either resist it or work around it.

Leaders need to address this directly. That means involving staff in redesigning workflows, being explicit about where human judgement remains essential, setting clear guardrails for AI use, and aligning performance measures so that staff are rewarded for using new processes rather than reverting to old ones.

Changing the shape of work is not just a technical challenge. It is a workforce and trust challenge.

If staff are not involved in redesigning the work, AI risks being experienced as something done to them rather than with them. That is not a recipe for transformation.

Scale patterns, not pilots

The organisations that scale agentic AI successfully will not simply collect pilots. They will identify patterns.

A citizen asking the same status question repeatedly, a staff member searching across multiple knowledge sources, a form returned because evidence is missing, a contact centre receiving demand that could have been prevented, or a service with a clear next action but poor visibility — these are not isolated problems. They are patterns.

The specific policy area may change, but the shape of the work is often familiar.

That is where AI can move from experimentation to transformation. Not by scaling one pilot everywhere, but by identifying the reusable service pattern: the knowledge source, the data requirement, the workflow, the decision boundary, the escalation route, the governance model and the surface through which people will engage.

This is also where leaders can have a more useful conversation about value. Not “how many pilots do we have?” but “which patterns are we proving, and where else could they apply?”

Measuring what matters

If leaders want AI agents to move beyond pilots, they need to measure more than adoption.

Weak measures include the number of pilots launched, tools deployed, prompts used or hours self-reported as saved. These can be useful indicators, but they do not prove service transformation.

Stronger measures connect AI to operational and citizen outcomes. Has repeat contact reduced? Are incomplete submissions falling? Are cases being triaged faster? Are manual hand-offs being removed? Is staff capacity improving? Are citizens reaching the right outcome sooner? Are complex or sensitive cases being escalated to humans more appropriately?

The aim is not simply to show that AI works. It is to show that the service works differently because of it.

The leadership questions

Three questions matter more than the rest.

• First: are we using AI to speed up tasks, or to change the workflow? If the answer is only the former, the impact will be limited.
• Second: where are citizens and staff currently chasing knowledge, status or reassurance? These are the points where demand can be reduced and experience improved.
• Third: what outcome would prove that the service has genuinely changed? Without a clear answer, pilots will continue without delivering transformation.

These questions move the conversation away from isolated tools and towards service design.

What comes next

The next phase of public sector AI will not be defined by the number of pilots in flight. It will be defined by whether organisations can turn AI capability into measurable service change.

That means valuing individual productivity, but not stopping there. It means improving internal knowledge access, but also reducing citizen-facing friction. It means investing in data, but recognising the immediate value of usable knowledge. It means choosing surfaces deliberately, because experience design shapes trust, adoption and behaviour.

Most importantly, it means deciding which work should no longer exist at all.

Because the real test of AI in the public sector is not how much faster we can do the work we already have.

It is whether leaders are prepared to redesign the work that should not exist in the first place.

That is the strategic fork.

If you'd like to explore these ideas in more detail, we've recently published the second edition of the AI Agents Handbook: Redesigning Public Services for the Age of AI. It expands on many of the themes in this article, including how to identify high-impact opportunities for AI agents, avoid common pilot traps, and build a practical path from experimentation to public service transformation. You can download your copy here.