When restraint makes government more efficient

As outlined in Government Efficiency in the Age of AI, system-level efficiency comes from how well resources are used to produce outcomes for society. Democratic principles such as transparency, subsidiarity and accountability support this. Data science adds another: restraint, the deliberate choice not to intervene where it will not change the outcome.

In the welfare state, the usual response to pressure is to add capacity. That is understandable: action is visible and effort is often taken as evidence of care. But under tighter budgets and rising expectations, agencies need to become more selective, concentrating effort where it changes outcomes. A resource spent on an intervention that makes no difference is still a resource lost.

Data helps make that distinction. Its value lies in testing whether professional assumptions still match reality.

Data shows where intervention matters

At Estonia’s Unemployment Insurance Fund, Töötukassa, a machine-learning model called OTT has assessed jobseekers’ employment prospects since 2020. It shows counsellors the main factors shaping each person’s outlook, helping them decide how and when to intervene.

Some of its early findings challenged expectations. In several cases, the model identified prison release as a positive predictor of employment. For counsellors who regularly worked with released prisoners, this was unsurprising. Many knew that these individuals were already supported through probation, social services, in-prison training and employer subsidies. For others, with less direct exposure, the finding seemed counterintuitive.

Support was already in place. An additional intervention from the employment service would add little. Breaking silos is usually discussed in terms of the citizen experience. It should also help the state see where it is already acting and where another layer of support adds little.

When experience no longer matches reality

Professional judgement remains essential, but it is shaped by experience, and experience can become dated. Labour markets change, support systems change and citizens’ circumstances change. A counsellor whose understanding was formed several years ago may be working from a view of the system that no longer holds.

Models have limits, yet they can recalibrate more frequently and more systematically than professional intuition. Take the prison example. Cross-agency support for people leaving prison has grown over the last decade, and the data picked it up before counsellor intuition did.

Used well, data gives the counsellor a sharper lens on a changing reality. It is most useful where direct experience is limited, uneven or out of date.

Designing where to intervene

The labour market has also changed. People change jobs more often than they used to, and registering with an employment service no longer signals what it once did. Jobseekers now include senior specialists between contracts, professionals making lateral moves and executives looking for their next role. Many are fully capable of finding work without mandatory state-provided counselling.

Applied indiscriminately, mandatory counselling adds little value. It also consumes time that could be spent on people at greater risk of long-term unemployment.

This is where data science can move from diagnosis to service design. If a model can reliably identify jobseekers who are unlikely to need mandatory counselling, the obligation can be lifted. Counsellors’ time can then be reallocated to those who need it most. For a jobseeker at high risk of long-term unemployment, that can mean more frequent contact, a tailored plan and case management coordinated with other services.

Efficiency and impact then align. Restraint becomes part of system design, rather than case-by-case judgement made under pressure.

But what if someone slips through?

The harder concern is safety. Few would argue that a senior specialist between roles needs the same support as someone facing structural barriers to work. The relevant question is what happens if someone who does need help is missed.

The answer depends on how the system is designed. People must be able to decide for themselves, and getting back into support must be easy. In Töötukassa’s case, any jobseeker can book a counselling appointment through the self-service portal in a few clicks, provided there is availability on the other side. The model is not there to keep people out of support. It is there to avoid imposing support where it is not needed. Restraint is only legitimate when the route back into help remains open.

Where the state adds value

The central question for public services is where effort changes the outcome. That question applies far beyond employment services. Every social welfare agency, healthcare system and public employment service faces a version of it. Some citizens need more intensive support. Others need less. Some interventions change outcomes. Others mainly add process. Better service design means knowing the difference.

For senior public leaders, the challenge is straightforward: Where is the state acting without changing outcomes, and where could it do more by targeting effort better?

Those themes, along with a practical framework and roadmap for policymakers, are explored in Nortal’s white paper Government Efficiency in the Age of AI, developed with partners from the University of Tartu, the Estonian government and the Tony Blair Institute for Global Change.