AI in NHS Diagnostics: Delays, Legacy IT, and Lessons from Real-World Rollout

Hurdles for AI Implementation: What's Holding Back NHS Digital Transformation

AI in NHS hospitals is proving harder to implement than expected. A new UK study in The Lancet eClinicalMedicine, led by researchers at UCL with NIHR funding, shows the gap between promise and practical delivery.

The 2023 NHS England programme funded £21 million to roll out AI for chest diagnostics across 66 trusts in 12 imaging networks. By June 2025-18 months after contracting was meant to finish-23 trusts still weren't using the tools in clinical practice.

What the study found

• Contracting took 4-10 months longer than planned.
• Integration with ageing, varied IT systems slowed progress.
• Clinicians' workloads made engagement and governance approvals difficult.
• Staff scepticism and uncertainty about accountability reduced early enthusiasm.
• Training often failed to address decision-making, safety, and liability concerns.

Despite delays, national programme leadership, shared learning across imaging networks, committed local teams, and dedicated project management helped several sites move forward.

Why this matters for healthcare leaders

AI tools can support prioritisation and flag abnormalities, potentially improving diagnostic flow. But the study is clear: these tools will not relieve service pressures as simply as hoped unless leaders fix procurement, integration, and training fundamentals.

Where projects stalled

• Procurement overload: teams faced a flood of technical detail without clear comparators.
• Legacy IT: variable PACS, RIS, and data pipelines made standard deployment unrealistic.
• Governance drag: approvals, DPIAs, and clinical safety cases took longer than anticipated.
• Trust and accountability: concerns about AI "making decisions" without clinician oversight.
• Change saturation: busy services struggled to prioritise AI alongside core pressures.

What helped progress

• National guidance and central support to unblock local issues.
• Imaging networks sharing procurement lessons and technical templates.
• Dedicated project managers coordinating IT, clinical, and vendor tasks.
• Strong local clinical champions bridging supplier promises and on-the-ground reality.

Practical actions for the next 6-12 months

• Fund project management upfront: ring-fence capacity for procurement, IT integration, governance, and training.
• Adopt a national supplier shortlist: standardise due diligence, data security, and clinical safety requirements to reduce local burden.
• Use templated contracts: include service-levels, model updates, drift monitoring, explainability, and exit clauses.
• Build for interoperability: insist on vendor-neutral integration with PACS/RIS and standard formats (e.g., DICOM SR).
• Define a clinical operating model: where AI fits in workflow, who reviews outputs, and how exceptions are handled.
• Clarify accountability: document roles for clinical sign-off, escalation, and incident reporting.
• Run meaningful training: address use cases, limitations, bias, failure modes, and medico-legal questions-not just product clicks.
• Plan data and monitoring: establish baselines, prospective audit, false positive/negative tracking, and periodic model validation.
• Sequence sites: pilot in 1-2 trusts, stabilise processes, then scale across the network.
• Engage patients and carers: incorporate perspectives on consent, fairness, and transparency early.

Governance and risk controls to put in place

• Information governance: DPIAs, data minimisation, access controls, and audit trails.
• Clinical safety: hazard logs, SBARs, and safety cases aligned to DCB standards.
• Bias and equity: monitor performance by subgroup; define thresholds for corrective action.
• Model lifecycle: version control, change notifications, rollback plans, and end-of-life strategy.

Procurement checklist (short)

• Evidence: peer-reviewed performance in comparable settings; prospective validation plans.
• Integration: supported environments, APIs, and vendor responsibilities spelled out.
• Economics: clear cost model (licences, compute, support) with measurable ROI metrics.
• Support: onboarding timelines, training delivery, and response times for incidents.

What success looks like

• Time to report reduced for urgent cases without increased error rates.
• Stable integration with minimal manual workarounds.
• High clinician adoption with transparent oversight.
• Routine safety and performance monitoring with published outcomes.

For background on imaging networks, see NHS England's overview of imaging networks here. For the clinical research context, visit eClinicalMedicine.

Upskilling your teams

If your roadmap includes wider AI adoption, plan structured education for managers, clinicians, and IT. Curated options by job role can reduce time to competency-see courses by job.

Key takeaway

AI can support diagnostics, but only if leaders treat implementation as an operational change programme-not a software install. Resource it, standardise it, and train for it, and the benefits become reachable.

References (selected)

• Ramsay AIG, Crellin N, Lawrence R et al. eClinicalMedicine, 2025.
• Liu M et al. PLOS One, 2023.
• Joy Mathew C, David AM, Joy Mathew CM. EXCLI Journal, 2020.
• Chiu HY, Chao HS, Chen YM. Cancers, 2022.
• Lawrence R, Dodsworth E, Massou E et al. eClinicalMedicine, 2025.

Source: University College London, 13.09.2025