Automate, assist or retain manual assessment of a benefits caseload

A public body with a growing backlog decides whether to automate entitlement assessments, use a system to assist caseworkers, or leave the work manual. The log shows a legal position under the UK GDPR...

Published content

node: Context. The organisation is a public body administering a means-tested support scheme, receiving approximately 61,000 applications a year against a caseworker establishment able to issue about 52,000 decisions. The backlog stands at 14,000 cases and median time to decision has risen from 19 to 34 working days over three years. An applicant waiting for a decision is waiting for money. A six-week pilot of an assisted assessment has just concluded, which is why the choice is being made now.
node: Should assessment of the benefits caseload be automated, used to assist caseworkers, or left as a manual decision by a caseworker?
node: Judgement. A caseworker can recognise the circumstances the rules did not anticipate, and the discretionary element in the scheme exists precisely because the rules cannot enumerate every household.
node: Pilot. In a six-week pilot on 900 cases, caseworkers departed from the recommendation in 12 per cent of cases, and in four fifths of those the departure concerned household composition or a discretionary element rather than an arithmetic error.
node: Judgement. On the approximately 55 per cent of applications identified as straightforward in the case-mix assumption, with a single income source, undisputed household composition and no discretionary element, the rules determine the outcome arithmetically, and caseworker judgement adds nothing a reader of the regulations could not reproduce.
node: Decision-maker. The Director of Operations, with the Senior Information Risk Owner holding a separate approval on the data protection position. The Data Protection Officer advises and does not decide.
node: Risk. The model was fitted on past caseworker decisions, so it reproduces their errors as faithfully as their correct judgements, and the overturn rate measured against them cannot detect that. The function holds no independent record of what the right answer was.
node: Scope. The decision covers how an entitlement assessment is made on a complete application. It does not cover the eligibility rules themselves, which are set in regulations, nor the appeal mechanism beyond internal review, which sits with an independent tribunal.
node: Consulted. Caseworkers and their recognised union on the change in role; the applicant representative panel on how an automated or assisted decision should be explained; and legal advisers on the Article 22 position of each option.
node: Risk. Capacity is below intake, so this option does not hold the position — it worsens it. On current trend the backlog grows by about 9,000 cases a year, reaching about 23,000 within a year and 32,000 within two, and median time to decision continues to rise. The people bearing that delay are by definition those with the least means to absorb it.
node: Risk. The legal position and the quality benefit both rest on caseworkers genuinely departing from the recommendation. A throughput target applied to the same role would erode that without any decision being taken to change it, and under Article 22A a review without real authority to depart is not meaningful human involvement.
node: Risk. The legal position is contested rather than merely complex. Proceeding would put the scheme's determinations at risk of challenge as a class rather than case by case, and a successful challenge would fall on decisions already issued.
node: Risk. The option is not available until the regulations are amended, which is outside the function's control and on published timetables takes twelve to eighteen months. The backlog grows throughout.
node: Action. Seek the regulatory amendment that would expressly authorise automated determination for a defined straightforward class, with the safeguards Article 22C requires set out on the face of it. Until that is made, the automated option is unavailable rather than merely risky.
node: Constraint. A decision on entitlement to means-tested support has a legal or similarly significant effect on the applicant, and the assessment uses health and disability information, which is special category data. A solely automated decision on these applications therefore falls under the restriction in Article 22B of the UK GDPR as reformed. Explicit consent cannot be freely given to a public body on which the applicant depends for income, and there is no contract, so the only route is an authorisation in law. The legal position of each option is a criterion rather than an afterthought.
node: Risk. A recommendation shown before the caseworker forms a view anchors the assessment. The pilot did not test presenting the recommendation after the caseworker's provisional decision, which would test the effect but slow the process.
node: Risk. The capacity gain is partly illusory. Review demand more than doubles, and a review is more expensive than a first decision because it examines one that has already been made and explains why it is being changed.
node: Risk. The boundary between straightforward and referred is itself a determination affecting the applicant, and it would be made automatically. An applicant wrongly routed as straightforward receives an automated decision on a case that needed judgement, and the routing rule needs the same scrutiny as the assessment.
node: Action. Do not apply a throughput target to caseworkers operating the assisted process. The legal position of the recommended option depends on their authority to depart from the recommendation being real, and a target is the most likely way to remove it without anyone deciding to.
node: Constraint. The function cannot recruit its way out of the backlog. The establishment is fixed by the current settlement and two of the last four caseworker vacancies took more than six months to fill.
node: Action. Record the departure rate from the recommendation as a standing operational measure from the first week, so that the assumption the legal position rests on is observed rather than assumed.
node: Regulation. The Data (Use and Access) Act 2025 replaced Article 22 of the UK GDPR with Articles 22A to 22D, and this decision is taken under the reformed provisions. A significant decision is one with a legal or similarly significant effect on the data subject, and it is based solely on automated processing where there is no meaningful human involvement in taking it (Article 22A). Where such a decision is based entirely or partly on special category data, Article 22B permits it only with explicit consent, or where it is necessary for a contract or required or authorised by law and the substantial public interest condition applies. Wherever a solely automated significant decision is taken, Article 22C requires safeguards that inform the data subject and let them make representations, obtain human intervention and contest the decision.
node: Official guidance. The Information Commissioner's Office guidance on rights related to automated decision-making sets out that the additional rules apply to solely automated decisions with legal or similarly significant effects, that human involvement must be meaningful rather than a token gesture, and what safeguards a controller must provide where such processing is permitted.
node: Review trigger. If the departure rate from the recommendation falls below 5 per cent for two consecutive quarters, treat the assisted option as having become a solely automated one in substance and reassess the Article 22 position immediately. That is the threshold at which the recommended option quietly becomes an option that was not recommended.
node: Regulation. Article 22A makes meaningful human involvement the test of whether a decision is solely automated, consistent with earlier regulatory guidance that a token human step does not change the character of the decision. A human step without real authority or capacity to depart from the automated output is not meaningful involvement. A caseworker who nominally signs a recommendation but has neither the time nor the standing to reject it does not convert an automated decision into a human one.
node: Review trigger. If the regulatory amendment is made, reconsider automating the straightforward class, including the scrutiny owed to the routing rule that decides which cases qualify.
node: Assumption. The overturn rate on internal review is the best available proxy for decision quality, but it measures only decisions an applicant challenged. An applicant who accepts a wrong refusal does not appear in it, and there is reason to think refusals are challenged less often than reductions. Every option is compared on the same imperfect measure.
node: Review trigger. If internal review demand rises more than 20 per cent above the modelled figure, reassess: the capacity case for the recommended option assumes reviews grow with volume and not faster.
node: Assumption. Approximately 55 per cent of applications are straightforward on the rules as drafted: a single income source, no disputed household composition, and no discretionary element. This proportion is taken from a sample of 400 cases and drives the capacity figures for both partial-automation options.
node: Retain manual assessment of every application
node: Assist: the system produces a recommended assessment, a caseworker decides
node: Automate every assessment, with human review available on request
node: Automate the straightforward cases under an express legal authorisation, refer the rest to a caseworker
node: Recommendation. Adopt the assisted option now, and seek the regulatory amendment that would permit automating the straightforward class. The assisted option is the only one that raises capacity above intake without relying on a legal position the function cannot presently support. Full automation is not recommended at any speed advantage while its authorisation under Article 22B is contested, because a determination successfully challenged as unlawful is worse for the applicant than a slow one.