{"resourceId":"nao-dwp-innovation-ai-pilots-half-stopped-evaluation-pending-20261021","versions":[{"version":"external-4db7c650290f36b6c78f298be09808c6a87ddf2fdd2aaaa65dfd0fb1b43b3e7c","resource":{"id":"nao-dwp-innovation-ai-pilots-half-stopped-evaluation-pending-20261021","title":"UK spending watchdog finds the benefits department stopped half of its AI pilots and scaled the rest, but has not finished long-term evaluations of the AI tools it kept","organization":"National Audit Office (UK); Department for Work and Pensions (DWP)","sector":"National benefits administration; innovation and AI pilot governance","geography":"Great Britain (international, national government)","publishedAt":"2026-10-21","publicationDate":"2026-10-21","eventDate":null,"sourceName":"National Audit Office","sourceLabel":"NAO report summary (HC 616, Oct 21, 2026): 'DWP's approach to innovation'","sourceUrl":"https://www.nao.org.uk/wp-content/uploads/2026/10/dwps-approach-to-innovation-summary.pdf","evidenceClass":"government-audit","outcomeClass":"mixed","topics":["operating-model","knowledge-work","governance-procurement"],"finding":"The NAO reviewed how DWP, which pays benefits to about 23 million people, develops and scales innovation. In DWP's Lighthouse Programme for generative AI, half of the projects tested progressed beyond the pilot stage and half were not taken forward. Case examples include the Generative AI in Learning (GAIL) tool for staff training material, a Child Maintenance Service predictive model that flags cases likely to stop paying so caseworkers can intervene, and a Whitemail Insights and Vulnerability Scanner that reads about 25,000 customer letters a day to flag potentially vulnerable people. DWP used test-and-learn methods on GAIL and the predictive model but has not completed a long-term evaluation of either. A 2024 internal audit found poor sharing of lessons from the scanner. 67% of staff felt supported to try ideas that might fail, against a 70% civil service average.","sledRelevance":"Why now: published Oct 21 and not in our archive. U.S. state benefits agencies are moving AI pilots into production (Maryland, Pennsylvania and Missouri items already archived) and new federal funds back benefits AI projects. This audit gives a rare independent count of how many AI pilots a large benefits agency actually stopped, and shows that test-and-learn evidence can carry tools into service before formal evaluation finishes. Transferability: a single national department with a central innovation function, unlike 50 state agencies under federal program rules.","evidence":"National audit office summary opened in full; based on document review and staff interviews. It reports the 50/50 pilot split and process findings but gives no cost, accuracy or productivity figures for GAIL, the predictive model or the scanner. The full report was not opened.","architectureImplications":"Interpretation: the tools kept span three patterns that state agencies also use: generative content for staff, a predictive risk score that prioritizes caseworker attention, and document triage at high volume. Each needs different monitoring: content quality review, score calibration, and miss-rate checks on flagged vulnerability.","governanceImplications":"The NAO credits DWP's differentiated approach to risk and its willingness to stop pilots, but notes that HM Treasury treats test-and-learn as a complement to, not a substitute for, formal evaluation. It recommends recording lessons from pilots that were stopped, not only successes.","securityPrivacyImplications":"The NAO stresses that innovation must not compromise protection of large volumes of sensitive personal data; the summary does not assess specific data controls on the AI tools.","caveats":"Summary only; no outcome metrics for individual AI tools. UK national context; U.S. state benefit programs operate under federal rules and separate state systems. Staff survey figure is from 2025.","streamIds":["state-government"],"roles":{"sales":"Sales framing: a credible auditor found that a large benefits agency stopped half its AI pilots and kept tools in service before long-term evaluation was done. For U.S. state benefits and workforce agencies, the opening is evaluation discipline, not more pilots: deciding up front what evidence ends or scales a pilot, and capturing lessons from the ones that stop. Avoid presenting a 50% progression rate as a benchmark; it is one department's count.","engineering":"Engineering framing: the three DWP tools map to common state patterns: generative drafting for staff, a predictive score to prioritize caseworkers, and high-volume mail triage for vulnerability. Build each with the measurements needed for a later formal evaluation, such as baseline handling times, score calibration and miss rates on flagged cases, so test-and-learn data can support a real assessment rather than replace it.","delivery":"Delivery framing: pilots need explicit stop and scale criteria and a written lessons record, including for failures, which the NAO found were shared less than successes. Before a pilot moves to production, schedule the long-term evaluation and name who owns it. Expect staff confidence to matter: about a third of DWP staff did not feel supported to try ideas that might fail."},"roleDetails":{"sales":{"customerProblem":"Benefits agencies can run many AI pilots but lack consistent rules for which to stop or scale, and tools can reach production before formal evaluation is complete.","stakeholders":["Agency innovation or AI program lead","Benefits program directors","Agency evaluation or research unit","State auditor or inspector general","Agency CIO"],"discoveryQuestions":["How many AI pilots has the agency started and stopped in the last year?","What evidence decides whether a pilot scales?","Which production AI tools still lack a completed evaluation?","Where are lessons from stopped pilots recorded?"],"valueHypothesis":"Interpretation: clear stop and scale criteria could reduce money spent extending weak pilots; the NAO gives no cost figures to quantify this.","potentialEngagement":"Bounded review of one agency's AI pilot portfolio: classify each pilot, define stop and scale criteria, and draft an evaluation plan for tools already in production.","unsupportedClaims":["No cost, accuracy or productivity results for GAIL, the predictive model or the scanner","A 50% pilot progression rate is not a benchmark","No evidence on U.S. state transferability"]},"engineering":{"fit":"Benefits and case-management agencies running staff-facing generative tools, predictive caseload prioritization or high-volume correspondence triage.","architectureIntegration":"Capture baselines and per-case logs from the start; record predictive scores with outcomes for calibration; log flagged and unflagged correspondence to estimate miss rates; keep staff review in the loop for vulnerability flags.","prerequisites":["Baseline metrics before pilot launch","Outcome data linked to model scores","Correspondence logging with retention rules","Evaluation plan agreed with an evaluation unit"],"constraints":["UK national department; U.S. states operate under federal program rules","No public tool-level metrics to reuse","Sensitive personal data limits data sharing for evaluation"],"security":"Predictive models and mail scanners process sensitive personal data at scale; restrict access, log use and review vulnerability flags with staff before action.","proofOfValue":["Compare caseworker interventions on high-score versus low-score cases against payment outcomes","Sample unflagged letters to estimate the scanner's miss rate on vulnerable customers","Measure time to produce training material with and without the generative tool against a baseline"]},"delivery":{"work":"Set up pilot governance: intake, stop and scale criteria, evaluation scheduling for tools entering production, and a shared lessons log covering stopped pilots.","dependencies":["Evaluation unit capacity","Data access for outcome measurement","Leadership support for stopping pilots","Staff time for frontline idea testing"],"ownership":"Agency innovation or AI program office, with the agency evaluation unit owning long-term evaluations.","skillsAdoption":"Staff need confidence to propose and stop experiments; 67% of DWP staff felt supported to try ideas that might fail, so a third did not.","governanceCheckpoints":["Stop and scale criteria set before pilot launch","Evaluation plan approved before production","Lessons recorded for every stopped pilot","Periodic review of production tools lacking evaluation"],"acceptanceCriteria":["Proposed: every pilot has written stop and scale criteria at launch","Proposed: every production AI tool has a scheduled long-term evaluation with a named owner","Proposed: lessons log entries exist for all stopped pilots"],"risks":["Tools stay in production on weak test-and-learn evidence","Lessons from failures are lost","Duplicate pilots across divisions","Data protection concerns slow evaluation data access"]}},"retrievedAt":"2026-10-26T22:21:01-05:00","enrichedAt":"2026-10-26T22:23:58-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Staff culture finding: 67% of DWP staff felt supported to try new ideas that might fail, slightly below the 70% civil service average; frontline staff are a main source of automation ideas.","procurementImplications":"Not addressed in the summary.","operatingModelImplications":"DWP runs several innovation units (Lighthouse Programme for generative AI, an Intelligent Automation Garage and an Innovation Tests and Trials Team); the NAO recommends applying good practice consistently and sharing lessons across them.","sourceVerification":{"openedUrl":"https://www.nao.org.uk/wp-content/uploads/2026/10/dwps-approach-to-innovation-summary.pdf","referenceExcerpt":"half of the projects tested progressed beyond the pilot stage, with the other half not taken forward","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}