Government may need to rethink its search area in the hunt for AI savings

Automating citizen-facing interactions delivers fewer improvements than departments expect. Hannah Bolton and James Ainsley, experts in government and public sector service operations at Baringa, explain how the efficiencies behind the Spending Review 2025 targets will likely come from AI applied to back-office work instead of customer-facing chatbots

Credit GettyImages

By Baringa

https://www.linkedin.com/company/baringa/

01 Jul 2026

Partner Content

Stretched government departments are under pressure to do more with less. The promise from AI companies – rapid and low-cost efficiency gains – makes automating citizen-facing contact a tempting starting point. Yet our latest research suggests that the benefits of AI are rarely realised in this way. Only 14% of the people we surveyed prefer AI to human interactions, and just 0.07% want fully automated services.^[1]

Using AI to automate citizen interactions tends to deliver fewer benefits than expected. More often, the gains come from improving the back-office work that citizens never deal with directly. When you listen to what people want from public services, it’s easy to understand why. For now at least, AI works best in a supporting role.

People prefer interacting with other people

In our research, we surveyed more than 1,000 customers and analysed over 37,000 online reviews of AI-enabled services to understand how consumers experience AI. Our findings reveal a consistent picture. Two in five people (40%) find AI in services frustrating, a figure that climbs to 50% among the over-45s.^[2] Only 8% would describe their experience of interacting with AI as ‘caring’.^[3]

These findings suggest a distrust of AI services, but consumers recognise opportunities, too. Nearly three-quarters of them want AI to help service staff do more.^[4] Yet the single most requested change, at the top of 51% of consumers’ AI wish lists, is the ability for AI to escalate an issue to a human intelligently.^[5]

Half of consumers think that AI should be used to enhance and augment employees' work.^[6] Some 82% it's essential or important to speak to a human to deal with a customer service issue.^[7]

Unsurprisingly, people are least comfortable with AI in the sensitive situations that public services handle daily. Close to two-thirds (65%) are uneasy about using AI for emotional or urgent matters,^[8] and nearly three-quarters say it has no place in distressing situations such as bereavement.^[9]

The Government Digital Service's (GDS) own experiments give us an indication of why. GOV.UK Chat, the flagship generative AI trial, reached 90% accuracy this year,^[10] which GDS says is in line with industry benchmarks.^[11] As strong as that is, it still leaves close to one in ten answers falling short. Furthermore, GOV.UK Chat cannot access anyone's personal account information (some users told GDS they wanted an adviser who could).^[12]

Looking across these findings, the overarching message is that departments should start with routine, lower-stakes AI use cases, and ensure they’re working well before progressing to more challenging tasks.

Curious what 1,000+ customers and 37,000 reviews reveal about how people really experience AI? Read our report

Where can AI drive efficiency improvements effectively?

To identify the highest-value uses of AI in public services, departments need a clear view of the end-to-end user journey, the operational flow and the workforce. Where does demand enter the system? Where does work stall, and where is effort duplicated? And where do citizens have to make contact again because their issue wasn’t resolved the first time?

Repeat contact is the costly factor. A significant proportion of customer calls stem from failure demand: people chasing progress or calling back because the first attempt didn’t resolve their issue. Each repeat call adds cost to the service and frustrates citizens without adding value. In cases like these, deploying customer-facing chatbots usually results in the wrong things being done more quickly, rather than actually resolving citizen queries.

However, AI can be effective in addressing these issues from the back office. It can triage requests, flag missing evidence and route work to the right team. Case summaries, document processing and root-cause analysis of recurring complaints all help fix issues rather than re-handle them. Plus, none of this replaces the interactions that citizens value most.

There are early public-sector examples to point to. The government's Consult tool, built by the Incubator for AI, has run across 26 live consultations and saved thousands of admin hours and over £500,000.^[13] Fraud and error, meanwhile, cost the taxpayer between £55bn and £81bn every year^[14], and departments are already turning to machine learning to tackle this issue. Both sit far away from contact-centre chatbots.

These examples and the research findings reveal a pattern that tallies with our own experience: use cases are only successful if they improve the economics of service delivery. They tend to disappoint if they optimise fragments while leaving the underlying operating model untouched – for example, deploying a standalone chatbot.

At Baringa, we use various techniques to determine whether an AI use case will be effective. If a capability doesn’t improve the end-to-end journey for citizens or staff, or if it generates activity without cutting costs or improving outcomes, then these are red flags. By contrast, if it reduces failure demand, backlog, or casework, it is far more likely to deliver savings and improvements. The best use cases for deployment are those where teams can articulate the current baseline, the intended improvements across operational and service-led metrics, and the safeguards that need to be put in place before building anything.

Want help pinpointing where AI will genuinely improve outcomes in your service? Talk to our government team about your use cases

Meeting the efficiency targets

The pressure on departments is considerable. The Spending Review challenges them to find at least 5% in savings and efficiencies by 2028–29, while cutting departmental admin budgets by at least 16% in real terms by 2029–30.[15] Departmental efficiency plans published with the Spending Review identify almost £14bn of annual efficiency gains by 2028–29.[16] These ambitious targets assume that AI will deliver savings. To meet them, departments must select the right use cases, then carve out a clear path from pilots to realised value.

However, scaling use cases isn’t necessarily straightforward. Pilots often succeed because a small, motivated team compensates for inconsistent data quality, a lack of an agreed-upon way to check AI outputs, and no clear accountability when the tool gets things wrong. These informal fixes almost never survive a full rollout.

To avoid a false start, departments should begin with where AI genuinely improves outcomes and how to unlock that value faster. This is the key to cutting through the AI hype and delivering lasting results that benefit citizens and the public sector.