Governing in the Age of AI: Reimagining the UK Department for Work and Pensions

The welfare system run by the Department for Work and Pensions (DWP) is meant to be a safety net, but today it is more like a sticky spider’s web. Instead of receiving support when they need it and help to move forward, citizens find themselves trapped in a system that is slow, inefficient and does not tackle the root causes of need.

With more people reliant on the DWP for longer, the welfare bill is large and growing under the pressures of an ageing population and the ongoing health-care and cost-of-living crises.

Tackling these challenges with pace and purpose is key to creating long-term security, opportunity and prosperity. Today, the government has at its disposal a new set of tools, enabled by advances in artificial intelligence, that can deliver change in the DWP.

The recent election campaign focused on political debates of the past: tweaking benefit-eligibility thresholds, vague promises to cut costs and the issue of whether or not to raise taxes. However, these are the wrong questions for the new government to focus on when considering how to get to grips with the high costs and ineffectiveness of the DWP.

The new government must use this opportunity to show what it means to govern well in the age of AI.^[_] Almost everywhere, AI can help us reimagine the state. Even the largest, most complex departments such as the DWP can be transformed so they can reduce need, provide timely, high-quality support and drive economic growth in the long term. This is what makes today the most exciting and creative time to be in government.

The day-to-day tasks of the DWP – those it shares with other departments, such as paperwork or policy design, and those unique to it, such as the operation of job centres – can all be performed faster, better and at less cost using AI.

Using the Tony Blair Institute for Global Change’s (TBI) unique database of almost 20,000 tasks ranked according to the potential impact of AI on them, we find that the DWP workforce could free up as much as 40 per cent of its time using AI tools. This is equivalent to a productivity gain of close to £1 billion a year.

At the moment, the DWP’s operating model is bureaucratic and reliant on complicated forms and paperwork. It is labour-intensive, which means it struggles to respond to citizens’ changing needs, and beset by backlogs and delays. Fraud and error cost the department close to £9 billion a year.

But a different model is possible.

Harnessing AI tools represents a once-in-a-generation opportunity to free up much-needed time and resources within the public sector. At the DWP, this would help it achieve its objectives: helping people move into work and progress in their careers, helping them plan financially for their futures (with a safety net for those who need it in the present) and providing high-quality services, all while ensuring value for money for the taxpayer.

By embracing AI, the DWP would be equipped to deliver three ambitious signature policies to create security, opportunity and prosperity:

1. Within a year, reduce backlogs for every type of benefit to zero to give every citizen the support they need when they need it. This can be achieved through better prioritisation and triage as well as improvements to application and eligibility-assessment processes. Delivering this policy would show that change is possible, allow support to reach citizens before their circumstances worsen, and free up time and space for deeper reform.

2. Reimagine job centres by introducing a digital employment assistant for every claimant so they can find the right job or training to progress in their career and gain financial independence. This can be achieved by personalising job and training advice, introducing virtual career clinics and matching opportunities more effectively to claimants’ circumstances and aspirations. Delivering this policy would transform the life chances of people trapped in low-wage, low-skill jobs, prepare citizens for an evolving labour market and unlock more opportunity for everyone.

3. Turn the DWP into an “AI exemplar” department that spurs cross-governmental collaboration to drive economic growth and reduce the long-term cost of benefits. This can be achieved through faster identification of error and fraud, the use of AI-enabled policy-development and monitoring tools to improve design and coordination, and new ways of sharing information and tools across government. Delivering this policy would serve as a clear demonstration that embracing AI is the key to modernising even the largest and most complex departments, setting the pace for others in government, and help design and deliver programmes that tackle the root causes of need.

A well-functioning DWP would create more security, opportunity and prosperity for every citizen. The only way to achieve this is for the new government to embrace the opportunity of AI and show in practice what it means to govern well in the age of AI.

The Department for Work and Pensions (DWP) is the second-largest government department by headcount (89,866 employees,^[_] second only to the Ministry of Justice)^[_] and the largest by expenditure (more than £240 billion in FY2022-23).^[_] It employs 16 per cent of the civil service and operates a massive estate, amounting to 20 per cent of the government’s civil estate, at a cost of £500 million a year, with more than 600 job centres employing 16,500 work coaches.^[_]

The DWP serves more than 20 million people, a little over half of whom receive the state pension (12.7 million people). The remainder are in receipt of some form of working-age benefit (9.3 million people) or under the age of 18 (660,000 people).^[_] These groups include some of society’s most vulnerable citizens – those who rely on the DWP to protect them from destitution and help them move towards a better life.

The department’s responsibilities^[_] are to help people move into work and progress within the workforce, help people plan financially for their future (with a safety net for those who need it in the present) and provide high-quality services, all while ensuring value-for-money for the taxpayer.

But instead of getting help when they need it or access to new opportunities, citizens get caught in the system. Rather than a safety net, the DWP has become a sticky spider’s web.

Many working-age claimants are not supported to find meaningful work or to make timely progress in their careers. Some 1.1 million universal-credit (UC) claimants are in employment but do not earn enough to become financially independent; another 1.4 million are able to work but remain unemployed.^[_] Of this total of 2.5 million people who are unemployed or low earners, 60 per cent (or 1.5 million) have been in receipt of UC for more than a year and 250,000 people for five years or more.^[_] Less than 1 per cent of those on the higher rate of incapacity benefit move off the benefit each month.^[_]

As the number of claimants exceeds the DWP’s capacity to review applications, backlogs develop^[_] and the most vulnerable people, including pensioners in poverty or people with disabilities, end up waiting months for support to arrive. For example, the median time from application to decision for personal independence payments (PIP), a benefit for people with certain disabilities, is 15 weeks.

The combined pressure of the large number of claimants and significant backlogs affects the quality of service. In job centres, known as Jobcentre Plus, where work coaches meet benefit claimants to monitor their progress and advise on next steps, appointments can last as little as ten minutes, limiting opportunities for meaningful support. When it comes to pension provision, people’s ability to plan for their futures can be derailed by communication failures from the DWP.^[_]

Decision-making under pressure is often poor: more than half the decisions to refuse an application for the PIP benefit that are appealed end up overturned by a tribunal panel drawing a different conclusion from the same evidence.^[_]

The DWP can also make costly errors. In 2021, the National Audit Office concluded that the lack of connected computer systems contributed to underpayments in state pensions amounting to over £1 billion.^[_] Across all benefits, in FY2023-24, the DWP overpaid a total of £9.7 billion (including £7.4 billion due to fraud, £1.6 billion due to claimant error and £0.8 billion due to official error), recovering just £1.1 billion of that sum for a net loss of £8.6 billion.^[_] Another £1.1 billion was underpaid, complicating financial planning for both citizens and the department.

The result is that citizens, the economy and the nation’s finances suffer. Insufficient support prevents citizens from making positive choices about their lives at critical junctures. Economic inactivity and wasted potential hamper growth and productivity. And government finances are strained, with spending on working-age health-related benefits as a proportion of GDP set to double in real-terms between FY2018-19 and FY2028-29.^[_]

But a different approach is possible. Our paper, Governing in the Age of AI: A New Model to Transform the State, written in collaboration with UK-based applied AI company Faculty, sets out the opportunity presented by advances in artificial intelligence to reimagine how government engages with citizens, operates and makes decisions.

It maps out the path from the current operating model of the state – bureaucratic, labour-intensive and reliant on one-size-fits-all policies that cannot respond to changing circumstances – to a new model that uses AI to transform departments’ interactions with citizens, day-to-day functions and policy-development processes.

As one of the largest and most complex government departments, interacting with millions of people on a regular basis, the DWP serves as a perfect case study for the impact this new model could have on the state’s ability to deliver for its citizens.

The DWP Operating Model and AI-Enabled Productivity Gains

The DWP’s vast remit leads to a complex operating model that encompasses how it engages with citizens through job centres and call centres, how it structures its back-office operations and how it makes decisions in Whitehall. This involves a wide range of tasks across each of these areas, as this paper now explores.

Citizen-Engagement Flows

DWP-run job centres, known as Jobcentre Plus, are the main point of face-to-face contact for UC claimants searching for work. There are around 650 job centres in the UK; these cost £500 million a year to run^[_] and are staffed by around 16,500 work coaches. These work coaches meet claimants weekly or fortnightly to monitor both their progress and their compliance with the conditions attached to their benefit payments. They are also meant to provide advice on employment matters, discuss career goals, signpost training opportunities, review CVs or applications, make decisions on sanctions and maintain claimants’ records. In practice, each coach is responsible for upwards of 100 claimants^[_] and each appointment may last as little as ten minutes.

For other benefits, such as child-maintenance support, PIP, carers allowance or pensions, the main point of contact is a DWP call centre. Customer-service agents manage upwards of 34 million calls a year,^[_] responding to queries about eligibility and entitlement, helping claimants complete applications, collecting relevant information, brokering difficult conversations, making referrals and coordinating with other agencies as needed to resolve complex cases.

Operational Flows

The DWP back-office functions are responsible for the assessment and processing of claims. Case managers assess applications against eligibility criteria, cross-referencing submitted information with data held on government systems, updating claimants’ details, registering decisions and triggering the provision of payments.

Counter-fraud staff are responsible for compliance, debt management and fraud investigations; they manage or conduct fraud investigations in person or via telephone, maintain records and capture relevant evidence, redact and annotate sensitive material, request relevant evidence from designated information providers, develop and prepare prosecution cases, and represent the DWP in court proceedings.

Decision-Making Flows

Finally, policy officials and other staff, such as communications, data, digital and technology professionals, direct and help design and implement welfare policies. They identify gaps, challenges and opportunities for improvement within the DWP’s areas of responsibility, collect and analyse data on service uptake, availability and the quality of welfare policies, provide expert advice and recommendations to senior officials, ministers and other decision-makers on policy matters, and support the complex systems on which the DWP relies.

The DWP’s digital team is already implementing a range of AI projects^[_] to improve departmental efficiency and productivity. The recent National Audit Office review of the use of AI in government^[_] highlighted the department’s creation of an AI steering board and an assurance group to scrutinise and maintain a register of AI projects. Existing projects include the use of robotic process automation to speed up repetitive tasks such as creating correspondence bundles,^[_] investment in machine-learning tools^[_] to surface potential cases of fraud or error for case workers to review and, more recently, a set of experiments with generative AI tools.^[_]

The Size of the Prize

How AI Enables a New Operating Model of Government

TBI and Faculty’s paper, Governing in the Age of AI: A New Model to Transform the State, showed that across these three groups of processes – citizen-engagement flows, operational flows and decision-making flows – a large proportion of tasks could be improved or automated using AI technologies that exist and are in use today, amounting to a potential productivity gain of £40 billion a year across Britain’s public sector. Next, we briefly describe how AI could be applied to each of these processes to deliver positive change and the technology’s potential impact on the DWP more specifically.

The current model for citizen-engagement flows is characterised by significant costs, friction and delays, driven by a reliance on lengthy, complicated forms and analogue processes, poor data sharing and insufficient capacity to provide meaningful, targeted support. Engaging with government is a poor, often alienating experience for citizens. This is as true of the DWP as it is of other departments. AI could help to:

• Streamline citizens’ access to the information they need at the right time, in the right format for them: for example, provide clear summaries about eligibility rules for benefits, personalised to citizens’ circumstances, and answer questions about them.

• Automate significant aspects of transactional services by safely linking data and making it easier for citizens to receive the support they are entitled to, including financial transactions: for example, pre-approve citizens for benefits where the department already knows they are eligible, so they do not need to fill in forms.

• Combat unfairness in access by expanding the range of available channels so that the same quality of service is available in person, over the phone or online: for example, provide the same up-to-date information to coaches whether they are meeting a claimant face-to-face in a job centre or supporting them remotely.

Harnessing AI in this way would allow departments to shift to a proactive pre-approval model for same-day delivery of most services, enabled by a Digital Public Assistant^[_] tool that intelligently suggests services to citizens, simplifies payments and provides accurate, simple and up-to-date information.

The current model for operational flows is inefficient, reactive and hampered by low productivity. This leads to growing backlogs, lengthy waiting times and difficulties responding to changes in demand or the wider operating environment. AI could help to:

• Accurately forecast patterns in demand, allowing operational managers to optimise their workforce and align the supply of services: for example, ensuring sufficient numbers of health professionals are available to carry out PIP assessments.

• Speed up prioritisation and triage, accelerating and facilitating the rote tasks that cost staff time so that they can focus their attention on the high-value task of scrutinising complex or high-impact cases: for example, prioritising the most vulnerable claimants for faster decisions.

• Upgrade investigations and analysis by analysing streams of real-time data to produce actionable insights at an otherwise impossible pace and scale: for example, identifying fraudulent or erroneous claims in progress and nudging citizens away from proceeding with these claims.

Staff would be guided and supported in their day-to-day tasks by tools integrated into a Multidisciplinary AI Support Team platform^[_] that acts not just a co-pilot, picking up parts of tasks, but as a co-worker completing its own tasks alongside them, freeing up their time for more meaningful work.

The current model for decision-making flows is limited in terms of its ability to act quickly on large volumes of data and information about the operating environment, past or present performance and citizens’ needs. As a result, decisions lack accuracy and can lead to costly errors, while unintended consequences go uncorrected. AI could help to:

• Model policy systems based on frontline data from across the country in present time, forecasting likely developments and simulating the impact of potential decisions: for example, rapidly generate dozens of scenarios for proposed changes to eligibility thresholds and demonstrate how each would impact employment rates, poverty levels and demand for other services such as the National Health Service (NHS).

• Accelerate research and support tasks such as consultation-response analysis, composing briefing notes or evidence reviews and tailoring the output to user needs: for example, summarise expert input and international evidence on the efficacy of different labour-market interventions, highlighting key themes and data points.

• Understand public opinion in detail with rapid, real-time aggregation, analysis and interpretation: for example, analyse citizens’ satisfaction with different DWP services across population groups, enabling the department to respond rapidly and improve provision.

The development and implementation of policies could be supported by AI-enabled National Policy Twin platforms^[_] that bring together information from across a wide range of sources to act as a single source of truth for policy planning, service monitoring and policy iteration.

The Potential Impact of AI Across the DWP

Embracing AI technologies in the public sector represents a once-in-a-generation opportunity and the only viable route to modernise a crumbling public realm. In the context of the DWP, harnessing AI tools would deliver significant productivity gains that would enable transformational change across its services.

The tasks performed by the DWP described earlier in this report are carried out by almost 90,000 employees, each of whom belongs to a grade within the civil service. For example, work coaches, who provide employment support, fall within the grade of executive officer. Many back-office roles are performed by administrative officers. Many policy analysts would come under the grade of higher or senior executive officer.

The department publishes^[_] the headcount and mean salary for each grade, allowing us to quantify the value of the tasks performed by each role. We analysed 900 tasks for 28 groups of roles across six grades, matching them to a bespoke TBI tool to evaluate the impact of AI on each task and occupation, and quantify the overall productivity gain.

Our analysis shows that embracing AI within the DWP would save employees 40 per cent of their time, delivering productivity gains of almost £1 billion a year. The findings of this analysis are summarised in Figure 4, with a methodology note and examples in the Annex.

This level of productivity gain would have a profound effect on the cost, quality and volume of services the DWP provides to citizens, and allow it to adopt and implement three signature policies:

1. Within a year, reduce backlogs for every type of benefit to zero to give every citizen the support they need when they need it.

2. Reimagine job centres by introducing a digital employment assistant for every claimant so they can find the right job or training to progress in their career and gain financial independence.

3. Turn the DWP into an “AI exemplar” department that spurs cross-governmental collaboration to drive economic growth and reduce the long-term cost of benefits.

Embracing AI is the only pathway enabling the DWP to deliver the kind of change it needs. Next, this paper sets out how the DWP could achieve these three signature policies.

Cut Backlogs to Zero to Give People Support When They Need It

The amount of time it takes for a citizen in need to receive support from the DWP is the first and most visible signal of whether the department is doing its job. While people wait, their health conditions deteriorate, their finances suffer, pushing them deeper into need, and their skills become less relevant to a fast-evolving labour market, making it harder to find good roles. Addressing these backlogs prevents the accumulation of greater need and demonstrates to claimants that the DWP is there to help. It also creates the breathing space needed for deeper reform to the department.

As a first step, the DWP should set itself an ambitious goal of using AI to reduce all existing backlogs to zero and ensuring all new claims are processed within target timelines within one year, so that every citizen gets the support they need when they need it. Going forward, it should maintain that record and reduce the target timelines for every benefit to under two weeks by the end of this parliament.

The DWP faces mounting backlogs across many of its services. The Covid-19 pandemic saw a marked increase in the number of working-age claimants, from 7.2 million in November 2019 to 9.89 million in November 2020.^[_] This spike in demand has had a knock-on effect across the department. Addressing this problem will not be possible without rethinking its operating model to put AI at its heart.

To scale operations under its current model, the DWP needs to hire more people or shift them between services. Increasing the headcount is not only expensive but also time-consuming. The right candidates must be found and trained. This introduces a time lag; as Figure 5 shows, increasing headcount to address a spike in demand can take up to a year. In the meantime, work piles up.

The main disability benefit, PIP, now has a backlog of over 330,000 new claims^[_] waiting to be processed as well as 450,000 cases awaiting a review decision. Another service, Access to Work, which provides tools and bursaries to remove barriers to employment, has seen the waiting list grow by 12,000 since January.^[_]

Waiting times have also ballooned, leaving some of the most vulnerable people without the support they need. The median PIP clearance time is 15 weeks from registration to decision,^[_] and 22 per cent of new pension-credit claimants wait more than 10 weeks for a decision.^[_]

By the DWP’s own admission, it routinely exceeds its own processing-time targets, despite the target for some benefits, such as PIP, being as long as 75 working days. Planned timescales for processing new claims were met for 75 per cent or more of the caseload for only four out of eight main DWP benefits. In one case, fewer than 5 per cent of applications were processed on time. Attempts to address delays in some services, such as UC, have led to productivity dropping for others, in a constant game of catch-up.

As our previous paper described, AI can dramatically speed up the processing and prioritisation of applications. It does this by pre-checking digital applications for completeness, triaging cases by calculating complexity and sending them to the right person, and automating rote tasks such as identifying and extracting key information (including from paper forms or physical documents). For example, automating triage to guide investigators to the most relevant evidence in each case has helped the National Crime Agency speed up the processing of batches of thousands of referrals by 90 per cent – cutting the time it takes from months to days.

The DWP is already deploying AI to speed up prioritisation, using tools to scan incoming paper letters to identify the most vulnerable people^[_] so that they can be matched to a caseworker more quickly. It should set itself an ambitious mission to reduce backlogs to zero, introducing AI across the benefit-application process to make it easier for claimants, prioritise the most vulnerable and assist with decision-making. Steps to achieve this mission could include:

• Using AI to rapidly digitise the remaining paper and PDF forms, turning them into online versions that use applicants’ answers to determine the subsequent fields for them to fill in, removing a major source of applicant confusion.

• Checking claims for completeness prior to submission, highlighting missing information with actionable suggestions on addressing evidence gaps.

• Moving to a pre-approval model for benefits, assessing eligibility prior to application based on known information and approving applications automatically where eligibility criteria are met.

• Reaching out to citizens automatically, using their preferred communication channel, when the department becomes aware of changes in circumstances (such as redundancy or a relevant diagnosis) that make an individual eligible for a benefit. Citizens would be allowed to opt out of this type of communication.

• Introducing AI systems to triage casework for more complex cases (such as PIP claims), prioritising the most vulnerable individuals and ensuring their claims are fast-tracked for consideration. This would include providing caseworkers with clear summaries of the facts in each case and the relevant guidance to save time cross-referencing complicated rules.

• Trialling and rating the outputs of “earned autonomy” AI systems^[_] as they shadow human decision-makers to provide a “second opinion” on cases, as well as building up a library of good practice in complex cases to support future staff training and decision-making.

• Monitoring relevant changes to the wider environment, such as local economic conditions or demand for particular types of health services, to assess likely spikes in caseloads and pre-allocate resources to deal with them proactively.

As analysis in our previous paper demonstrated, improving the throughput of new claims for PIP through better prioritisation alone would clear the backlog for that benefit in three to ten months, depending on the scale of productivity improvements. If the same changes were applied across all DWP services, this would free up time and resources to improve the provision of services, and in particular the struggling system of support for unemployed or low-income UC claimants.

Reimagine Jobcentre Plus to Build Long-Term Financial Independence

Supporting citizens into stable, fulfilling work is a major part of the DWP’s mission. To thrive in a changing labour market, embark on and progress in fulfilling careers, and achieve financial independence, people need high-quality employment support that is accessible, addresses their needs and minimises job-search friction. This kind of support would help drive economic growth, create opportunity and ultimately reduce the level of need across society.

With backlogs gone, the DWP would have the capacity to embrace a second signature policy designed to deliver this: reimagining Jobcentre Plus. This would involve creating new channels of highly personalised support, with face-to-face appointments available for those who need or prefer them. An omnichannel approach would allow a shift away from monitoring as the core of the system in favour of providing high-quality advice and support for jobseekers, helping them find work or training that sets them on the path towards better-paid, fulfilling roles and financial independence.

The current UC system for claimants in the “searching for work” or “working – with conditions” groups is built around weekly or fortnightly face-to-face meetings between a work coach and the claimant at a Jobcentre Plus location. The intention behind this approach is correct: in principle, this is an opportunity for an individual to maintain regular contact with the DWP, receive personal advice and obtain referrals to relevant services. In a pre-AI world, it made sense for this to happen in person.

In practice, however, job centres suffer from a lack of capacity, accessibility and quality.

The average caseload per work coach exceeds 100 claimants^[_] in “intensive work search”, leaving very little time to engage with each one individually – often as little as 10 minutes per meeting. In that time, the work coach is expected to review the claimant’s job-search efforts, assess their vulnerability, identify barriers and agree a job-search progression plan (ideally linked to local opportunities).

Instead, work coaches are effectively reduced to a UC-monitoring and compliance service. This focus on compliance with eligibility and job-search requirements also harms their relationship with claimants, making it more difficult to provide targeted advice that cuts through. The Joseph Rowntree Foundation has estimated that the DWP spends a total of £350 million^[_] a year monitoring claimants, diverting 13 million hours a year of work coaches’ time away from their intended role of creating opportunities for claimants to find jobs and progress in them.

The reliance on face-to-face meetings not only effectively limits the number of claimants the DWP is able to support but also creates inequalities of access. Job centres are not evenly distributed across the country and their locations do not reflect the current geography of need. When citizens’ needs change, as they did during Covid-19, there are no mechanisms in place to provide support in the “cold spots” short of opening new job centres at significant expense before shutting them down again as needs evolve.^[_]

This leads to a mismatch of supply and demand for services. As Figure 7 shows, in some local authority areas, there are no registered job centres, while other job centres have a significantly larger customer base.

When job centres are not easily accessible to claimants and attending requires lengthy and costly travel, this imposes a significant burden, particularly on those with limited means or mobility issues. At the same time, job centres currently focus on support for UC claimants and are largely unable to support the 3.9 million people^[_] who are not in paid work, do not claim benefits and are not students. The result is that the UK has the least well-used public-employment service^[_] in Europe, with only 18 per cent of jobseekers making use of the DWP’s services, compared with the European average of 54 per cent.

AI can help free up much-needed capacity within the system while also improving the quality of support and expanding access to services.

In the UK, the DWP is also trialling AI tools, including "a-cubed",^[_] a generative AI system that helps work coaches quickly navigate relevant guidance to improve the quality of advice they provide. The department should go further and reimagine Jobcentre Plus by introducing a digital employment assistant for every claimant and a network of virtual professional-growth hubs, with the intention of phasing out the majority of brick-and-mortar job centres. This could include:

• Moving away from a face-to-face-by-default approach, allowing claimants to set their own preferences for how to access support, with app-based or call-based channels available alongside traditional job centres. Citizens should be able to choose how to access support, with a parity of experience across different channels, rather than be obliged to visit Jobcentre Plus premises no matter how inconveniently located.

• Providing each claimant with a digital-assistant tool that supports them in moving into employment and progressing within the workforce. This should include their CV and could be linked to a digital-learner ID^[_] to build a complete picture of the claimant’s skills and areas of strength. Access to the tool should be expanded to include those not on UC but wishing to access career support.

• On first contact, carrying out a virtual assessment (with the option of in-person support) to generate a highly personalised version of the claimant commitment^[_] (a plan to look for work and increase earnings that current UC recipients must agree to) and ensure this is regularly updated in line with any changes to the individual’s circumstances.

• Using AI to analyse the citizen’s skills and experience, and suggest roles that would be a good fit now or, with training, in the future. To best match the claimant’s aspirations, the tool could take into account their preferences in terms of location and local labour-market conditions, such as in-demand jobs or jobs at risk of automation.

• Using AI tools to intelligently suggest activities that improve the claimant’s long-term employment prospects, such as local opportunities to upskill or entry-level roles in promising industries that match their experience and interests.

• Using AI to assist claimants in the job-search process, helping to tailor their CVs and covering letters to vacancies, identifying gaps relative to the job description, conducting mock interviews and providing tips on best job-search practices, such as attending networking events. These types of tools are already being rolled out by job-search platforms such as LinkedIn.^[_] With user consent, the assistant can keep a record of all or some of the job-search activities, reducing the need for, and burden of, manual monitoring.

• For those who are actively engaged in the AI-assisted job-search process, removing the requirement for mandatory meetings to free up further capacity within the system to support people who have been unemployed or underemployed over the long term. Evidence from the DWP shows that three-quarters of customers believe they could access government services using the internet without help.^[_]

• Setting up a virtual career clinic where claimants can book a video call with an advisor, following the example of virtual fracture clinics in the NHS^[_] which reduced the reliance on in-person appointments to assess and manage orthopaedic injuries. The advisor should have access to the relevant information from the claimant’s digital assistant as well as a database of locally available support, but need not be located in the same building as the claimant.

• Enabling citizens with more complex needs or those who prefer in-person support to access longer face-to-face appointments with coaches in existing premises. These coaches should be trained specifically to deal with more complex cases and given AI tools such as a-cubed to help them deliver high-quality help.

• Using demand modelling based on local economic data, as well as trends in claimant numbers, to address the lack of Jobcentre Plus premises in locations where there may be growing need. For any areas of shortage identified in this way, the DWP should develop a model of “pop-up” hubs, co-located with local authorities, anchor businesses such as banks or post offices or in vacant government-owned high-street premises (at a discounted rate until a permanent tenant moves in).

Initially, this reimagined system can be set up in parallel with existing Jobcentre Plus infrastructure. TBI analysis suggests that AI tools can free up over 40 per cent of work coaches’ typical day-to-day activities, amounting to almost 14 million hours a year across the department (see Annex). This time can be used to provide the parallel virtual advice to current claimants, as well as to a broader set of clients in need of career support.

It should be treated as an iterative, experimental programme, with short feedback loops to continually adjust and improve the model, with a view to replacing the current job centre-based approach by the end of this parliament.

Become an AI Exemplar to Drive Collaboration, Deliver Growth and Reduce Costs

The DWP is one of the largest and most complex departments. It is a visible presence in the lives of millions of people, including the most vulnerable, and its ability to do its job well reflects on the government’s ability to deliver for citizens. Fixing the DWP would send a strong signal about the role of AI in modernising the public services – if it can be done here, it can be done elsewhere.

By reducing backlogs to zero and improving the quality of its support, the DWP would create capacity for officials to work strategically with other departments to accelerate economic growth while reducing the costs of the welfare system. In particular, the DWP should become an exemplar for the use of AI in government to improve productivity, reduce fraud and error, and enable better collaboration across government.

At present, the pressure of backlogs and a demoralised, overburdened workforce limits the DWP’s capacity for effective decision-making. One area where this has an impact is the quality of decisions about individual applications. Increasingly, initial decisions made by DWP staff are reversed on appeal when the tribunal disagrees with the caseworker’s assessment.

For example, in 2022, six in ten PIP decisions were overturned not because of new oral or written evidence but based on the same facts as in the original application, a substantial increase compared to the previous decade. This both undermines trust in the system and carries a financial cost; the DWP spent £47.3 million on staff costs for PIP mandatory reconsiderations and tribunal appeals in 2023.^[_]

In some cases, good decision-making is hampered by legacy infrastructure within the department. As discussed earlier, a 2021 report from the National Audit Office found that historical underpayment of the state pension amounting to an estimated total of more than £1 billion resulted from a lack of effective automation and interaction between the department’s data systems, and a lack of specialised guidance for staff assessing eligibility.

Fraud and error in general constitute a wider challenge for the department.^[_] The annual cost is close to £10 billion, of which only about £1 billion is recovered. This includes overpayments due to fraud (amounting to £7.4 billion), claimant error (£1.6 billion) and official error (£0.8 billion). For some benefits, these overpayments represent a significant proportion of the total spend – for UC, fraud-related overpayments account for 12.4 per cent of the total expenditure.

These types of inefficiencies limit the DWP’s ability to design, test and scale good policy. The result – as in many other departments – is a proliferation of programmes and interventions, announced and often forgotten without much impact. Other more significant policy changes can create perverse incentives that lead to poor outcomes,^[_] such as trapping people in low-wage, part-time work.

The department can struggle to quickly source data that could help it learn from past mistakes^[_] or deliver large projects, such as the troubled initial implementation of UC.^[_] Out of six DWP-led major projects monitored by the Infrastructure and Projects Authority, five are amber and one is red.

Changes to assumptions or objectives take weeks, if not months, to translate to policy options and processes for consulting key stakeholders are equally laborious, risking public mistrust. The DWP does not always adapt well to new evidence^[_] after initial plans are made, leaving it to react belatedly, if at all, to poor outcomes, rather than address problems as they emerge.

The root cause of these problems is poor use of data and inefficient information flows within the department that narrow possibilities for change^[_] and prevent accurate and agile policymaking. Instead, ministers rely on manual and/or delayed data collection, with slow processes and reactive tendencies.

AI can be used to improve the quality of, and access to, data and transform the end-to-end policymaking process. Natural-language processing is being used to provide insights on structured and unstructured data. AI can help transform legacy infrastructure by helping to rewrite old code^[_] into modern programming languages, making it more usable and efficient.

AI can also improve policymaking. It can support the rapid implementation of complex projects by improving access to actionable insights on the current state, modelling policy options and risk, and understanding public opinion.^[_] It can also help design delivery mechanisms and embed agility into decision-making by shortening feedback loops, while ensuring fairness of service provision (for instance, identifying bias and fraud).

Complex computational twins and scenario modelling can be used to test assumptions about the outcomes of proposed policies and even run real-time experiments. For example, during the Covid-19 pandemic, the NHS used AI to test the impact of encouraging more patients to access support via the 111 remote service instead of attending Accident and Emergency departments. Anomaly detection and other algorithmic approaches can be used to identify instances of fraud while digital-ID systems can help improve performance tracking and service targeting, as well as provide an additional layer of verification.^[_]

The DWP is already doing some of this. It has replaced old code used in processing several benefits, with a 60 per cent improvement in efficiency and a net present value of £150 million.^[_] It has also invested £70 million to develop machine-learning tools to identify fraud.^[_] It should go further and aim to become an AI exemplar department that uses AI tools to reduce inefficiencies such as fraud and error, and uses data insights to drive collaboration across government. This could include:

• Developing a “policy twin” tool to support better policymaking. This would use data about the DWP’s users, wider economic conditions and past and present service performance, as well as evidence from the UK and examples from overseas regarding existing rules and regulations to accelerate the modelling of different policy scenarios.

• As recommended in our previous paper, aiming to adopt some of the working practices of leading UK firms that carry out similar types of transactional work – for example, fintech successes such as Monzo or Starling. These organisations prioritise good user-experience design, engage closely with their user communities, frequently introduce and rapidly iterate new features for beta-testing and look to reimagine traditional processes such as banking outside the constraints of legacy estates.

• Continuing to develop systems to reduce fraud and error, with a focus on real-time analysis, and using AI tools to provide more equitable access to services, for example by analysing biases in existing decision-making processes or by using anomaly detection to prevent gaming of backlog targets.

• Developing and, with a specific mandate from Number 10 and the Cabinet Office, sharing re-usable AI components and tools that can be both used by the department for its own purposes and easily adopted, as proven technology, by information-technology teams in other departments.

• Tracking the effectiveness of programmes and policy initiatives, focusing in particular on user outcomes including citizen satisfaction, users’ long-term prospects and reducing bias in provision. The DWP should be able to collect granular feedback from service users and harness AI to generate actionable improvement suggestions for work coaches, operational managers and policymakers.

• Driving efforts to link and analyse data across government to address intersecting needs of DWP service recipients. For example, the DWP could work with the NHS to make better use of medical records to accelerate and improve the accuracy of PIP claim assessments, support people with health conditions to enter long-term employment through collaboration^[_] with integrated care boards and local governments, and work more closely with the Department for Education to align training offers to local employment needs.

Adopting a collaborative approach underpinned by the efficient use of data and AI would allow the DWP to escape the present constraints of short-term, crisis-driven decision-making. By working with other departments to address complex needs, the DWP would be able to target a wider set of citizens who can benefit from its support, progressing into better-paying, lasting and more fulfilling careers, and gaining greater financial independence.

A more productive workforce should contribute to greater economic growth, as well as increased tax revenue. Addressing operational inefficiencies and tackling the issue of fraud could help significantly reduce the department’s expenditure.

By embracing AI in all aspects of its work, the DWP stands to gain significant advantages in terms of operational efficiency, expenditure, productivity and, most importantly, the quality of the services it provides. TBI analysis shows that AI can help save up to 40 per cent of DWP staff time, equivalent to £947 million.

Instead of serving as a sticky spider’s web in which citizens become caught, the DWP can leverage these productivity gains to advance users towards greater security, economic opportunity and prosperity. To achieve this, the department should set itself three ambitious missions:

1. Within a year, reduce backlogs to zero to give every citizen the support they need when they need it.

2. Reimagine job centres by introducing a digital employment assistant for every claimant so they can find the right job or training to progress in their career and gain financial independence.

3. Turn the DWP into an “AI exemplar” department that spurs cross-governmental collaboration to drive economic growth and reduce the long-term cost of benefits.

Ordinarily, these missions would be difficult – perhaps impossible – to achieve. But the technology available to us today puts them within our grasp. Embracing them is the only way to modernise a crumbling public realm. To show their impact on a department such as the DWP and the citizens it serves is to demonstrate in practice what it means to govern well in the age of AI.