Verified and updated: November 19, 2026, 5:06 a.m. UTC.
The Department for Work and Pensions (DWP) has disclosed a limited live trial of a machine-learning model to help prioritise Universal Credit claims for possible Targeted Case Review (TCR). The announcement matters because it confirms a new role for machine learning in the review process, while drawing a clear boundary: the model does not decide whether a review goes ahead or determine anyone’s benefit entitlement, payment, suspension, claim closure or review outcome. Those decisions remain with DWP staff.
DWP’s latest annual report says the programme reviewed nearly 1.2 million Universal Credit claims in 2025–26 and identified and corrected about 250,000 awards. The department reports £1.1 billion in estimated savings for that year and £1.7 billion since the programme began. Those are DWP’s figures and estimates; they do not mean that every reviewed claim was wrong or that the machine-learning trial produced those results.
Destaques
- The new model is being tested to help prioritise claims for possible review.
- DWP says trained staff retain responsibility for decisions that affect claimants.
- TCR checks Universal Credit awards and can identify underpayments as well as overpayments.
- DWP has not publicly described the model’s inputs, weighting or individual claimant scores in its letter to Parliament.
What are DWP Targeted Case Reviews?
Targeted Case Review is a DWP programme established in 2022 to check whether Universal Credit awards are correct. The department may examine a claim and ask the claimant to clarify circumstances or provide information. A review can find an overpayment, an underpayment or no error requiring a change. DWP says the programme also gives it information that can inform improvements to policy, services and operational delivery.
The programme’s name can sound as if a claimant is suspected of wrongdoing. A review is a check of award accuracy; selection for review alone is not proof of fraud or an incorrect claim. DWP’s explanation says the model supports prioritisation of claims for potential review. It does not publish the model’s detailed selection criteria in the parliamentary letter.
How does the DWP machine-learning system work?
In its October 26, 2026 letter to the House of Commons Public Accounts Committee, DWP described the model as a prioritisation tool within TCR, also known to claimants as Universal Credit Claim Review. In practical terms, it is being tested to help staff decide where review capacity might be directed. DWP has not set out the model’s features, training data, thresholds or how a particular claim is ranked, so those details cannot be reliably explained from the published account.
The distinction between prioritisation and decision-making is central. DWP says staff decide whether a potential review proceeds and make decisions affecting a claimant after considering available evidence. The model itself does not decide entitlement or payment, nor does it suspend or close a claim. DWP says the trial includes governance, evaluation, human review and consideration of effectiveness and equality impacts. It says the evaluation will consider whether further safeguards or adjustments are needed, and acknowledges the trial may not deliver as hoped.
Read the DWP letter to the Public Accounts Committee for its account of the trial. DWP’s 2025–26 Annual Report and Accounts gives the latest programme totals.
What happens if your Universal Credit claim is reviewed?
The new model does not change the basic point that the review is handled through DWP’s claim-review process, with staff considering the case. If DWP contacts you, read the message carefully and use the contact route it gives you. Check what information is requested and the response date. If you are unsure what a request means, or need more time or a different way to provide information, contact DWP promptly and explain the difficulty.
Keep copies of documents and messages you send, and make a note of when you sent them. Provide accurate information about the circumstances DWP asks about; do not assume that being selected means a decision has already been made. If the eventual decision changes your award and you disagree, read the decision notice for the reasons and the process and deadline for asking DWP to reconsider it. For help with a complex case or difficulty responding, an independent welfare-rights adviser may be able to explain your options.
What do the latest review figures show?
DWP’s annual report describes TCR as operating at full scale in 2025–26, with nearly 1.2 million claims reviewed and approximately 250,000 awards identified for correction. The report gives an annual average “hit rate” of 21% for reviewed claims found incorrect. That measure includes the programme’s identification of incorrect awards; it should not be read as a fraud rate or as the probability that any one claimant’s award is wrong. DWP reports £1.1 billion in savings for the year, based on its accounting approach and assumptions about the impact of corrections.
Looking ahead, DWP says it will monitor and evaluate the machine-learning trial. The public letter does not provide a timeline for results or promise that the model will be adopted more widely. The department’s stated next steps are to assess how it operates and whether it is effective, proportionate and appropriately safeguarded. Until it publishes more detail, the model’s precise operation and its measured effects on claimants remain unclear.
Frequently asked questions
Does the AI decide whether my benefits stop?
No. DWP says the model does not decide entitlement, payment, suspension, claim closure or review outcomes. Staff make decisions affecting claimants.
Does being selected mean DWP suspects fraud?
No. Selection for an accuracy review is not proof of fraud or an incorrect award. A review may find an overpayment, underpayment or no error needing correction.
Has DWP published exactly how the model scores claims?
Not in its October 2026 letter. DWP describes the model’s purpose and safeguards, but does not disclose its specific inputs, weights or ranking method.