TLDR: Healthcare technology management owns the structured equipment data most AI projects lack, so integration and governance — far more than the algorithm — decide whether it reaches the P&L.
The advantage is sitting in the equipment inventory
Most hospital functions start their AI journey by building the dataset. Healthcare technology management (HTM) — the discipline that keeps a hospital’s medical equipment fleet safe, available, and compliant — starts with one already built.
Every infusion pump, ventilator, and magnetic resonance imaging (MRI) scanner in a modern hospital already carries a service history: install date, work orders, parts replaced, failure codes, utilisation logs, recall notices. That record lives in a computerised maintenance management system (CMMS), the operational spine of every clinical engineering department. It is structured, longitudinal, and tied to physical assets with serial numbers. For machine-learning models that thrive on clean, labelled, time-series data, that is a rare starting position inside a hospital.
This is the real reason the HTM function keeps surfacing in conversations about practical AI. The draw is practical: the equipment inventory is one of the few places in healthcare where the data is already organised enough to act on. The webinar circuit has noticed: sessions on “unleashing AI in HTM” now promise to eliminate hidden inefficiencies, sharpen data-driven decisions, and turn modern equipment inventories into an adoption advantage. The premise is correct. The optimism is premature.
The 95 per cent failure rate follows AI onto the equipment floor
The hardest number in enterprise AI comes from an academic lab. MIT’s Project NANDA found that 95 per cent of enterprise generative-AI pilots left the profit-and-loss statement untouched, despite an estimated 30 to 40 billion dollars in spend. The researchers were explicit about the cause, and it sits in operations: brittle workflows, weak integration, and builds that start before the outcome is defined. The models themselves perform. Healthcare drew a specific mention, because clinical information is so context-specific that generic tooling struggles to land.
HTM inherits that risk directly. A clean CMMS schema and clean CMMS data are different things. The AAMI HTM benchmarking programme — the field’s reference for peer-to-peer performance comparison — has documented how messy the underlying records still are, surfacing outliers such as a reported 460 dollars per hour of internal labour cost, 30 million dollars of support attributed to a single technician, and cost-of-service ratios of 0.3 per cent that sit far below operational reality. AAMI’s own read is blunt: clinical engineering departments and CMMS vendors need to collect and provide far more accurate data before the metrics can be trusted. An algorithm trained on that record predicts failures with the confidence of a system fed fiction.
The pattern mirrors what Kainjoo has argued about transformation more broadly: data liquidity precedes AI capability. Before a model can transform maintenance, the data has to flow — accessible, interoperable, and clean across the CMMS, the asset tags, the building management system, and the original equipment manufacturer (OEM) telemetry feeds. Frozen, siloed asset records defeat sophisticated models the same way they defeat simple dashboards.
What “good” looks like: predictive maintenance that shows up in the P&L
When the data does flow, the payoff is concrete and measurable — which is exactly why HTM is a better proving ground than most AI showcases. GE HealthCare’s OnWatch Predict service applies machine learning to real-time scanner telemetry, and the company reports it helps improve MRI uptime by 4.5 days per year and cut unplanned downtime by around 40 per cent. The economics are easy to follow: GE HealthCare estimates a single day of unplanned MRI downtime forces roughly 18 cancelled scans and near 25,000 euros in lost revenue. Multiply that across an imaging fleet and predictive maintenance stops being an IT line item and becomes a revenue-protection strategy.
That distinction matters, because it is the one MIT identified as the difference between the 5 per cent that worked and the rest. Predictive maintenance succeeds when it is wired into the work that already happens: the model builds an intervention queue, the queue triggers a parts order, the parts order schedules a maintenance window during low-census hours, and the avoided downtime lands in a finance report. It fails when it produces a dashboard that a biomedical technician glances at and a vendor invoices for.
The honest caveat belongs here too. The headline ranges that circulate in the trade press — 40 to 50 per cent downtime reductions and 25 to 40 per cent maintenance-cost savings — are industry estimates drawn from favourable deployments. They mark a ceiling that organisations reach after fixing their data and their workflows first: a target to work toward, and a reward for the groundwork.
Beyond the wrench: the same data answers the capital question
Predictive maintenance is the obvious application, but it is the narrow one. The more valuable use of a clean equipment record is the decision that sits above the repair: what to buy, what to retire, and what to redeploy. Those are capital questions, and most hospitals answer them from a depreciation schedule and a budget cycle, while the evidence to answer them better already sits in the CMMS.
A CMMS that captures utilisation, failure frequency, and cost-of-service ratio (COSR) — the standard HTM metric for what a device costs to keep running relative to its acquisition price — turns the repair-versus-replace decision into an analysis instead of an argument. A scanner with a rising COSR and falling utilisation is a sell signal; an identical model running hot in another department is a redeployment opportunity that the budget process routinely overlooks. AAMI’s benchmarking work exists precisely so departments can measure these metrics against peers rather than guess.
This is where AI-driven analytics earn their place in HTM beyond the maintenance bay: right-sizing a fleet, timing capital purchases to real failure curves, and defending an alternative equipment maintenance (AEM) programme with the device’s own history. The same data liquidity that powers a downtime prediction powers a multi-million-franc capital decision — and the second is where the larger money is.
The security tax every connection adds
Connecting medical equipment to a network is the precondition for predictive maintenance. It is also the precondition for the largest unpriced risk in the hospital. Cynerio’s analysis of the connected-device footprint across more than 300 hospitals found that 53 per cent of connected medical devices carried at least one unpatched critical vulnerability. Intravenous (IV) pumps, the most common device on the network at roughly 38 per cent of the footprint, were the worst exposed: 73 per cent carried a vulnerability capable of threatening patient safety or service availability.
Regulators have moved to close the gap. The US Food and Drug Administration (FDA) issued its final premarket cybersecurity guidance in June 2025, setting lifecycle expectations from design through decommissioning and asking manufacturers to evidence penetration testing, patch-deployment timelines, and defect density. In Europe, the Medical Device Regulation (MDR) already treats security as a safety requirement, and the EU AI Act layers obligations onto AI systems that qualify as medical devices. The same data pipeline that enables an AI maintenance model is now a regulated, auditable attack surface — and HTM owns it.
The constructive read is that the fix is well understood. Cynerio’s data showed that network segmentation alone addresses more than 90 per cent of the critical risk presented by connected devices. Segmentation, an accurate asset inventory, and a patch-governance process are the unglamorous foundations that make an AI programme defensible. A predictive-maintenance project that ignores them is borrowing against patient safety to buy a dashboard.
The behavioural trap: why the pilot feels like the safe choice
If the path is this clear, why do hospitals keep running pilots that stall before they scale? The answer lies in behaviour, well ahead of technology. Loss aversion — the finding by Daniel Kahneman and Amos Tversky that people weigh a potential loss far more heavily than an equivalent gain — pushes decision-makers toward the option that feels reversible. A pilot feels safe because it is bounded; a full integration feels like exposure because it touches the CMMS, the network, the clinical schedule, and the budget at once.
So the pilot becomes the destination. Kainjoo has named this the “death by pilot” effect: experimentation that protects the sponsor’s downside while quietly keeping the value trapped inside the experiment. In HTM the trap is sharper, because the function’s entire value is reliability. The same instinct that makes a clinical engineering team excellent at planned maintenance — caution, reversibility, documentation — makes it conservative about the organisational change that scaling AI demands. Recognising the bias is the first step to designing around it: define the P&L outcome before the build, commit to the integration over the experiment, and measure the avoided loss in the same report as the spend.
Europe’s constraint is a head start
The instinct in regulated markets is to treat compliance as a brake on AI. The HTM case argues the opposite. The disciplines that European and Swiss healthcare already impose — device traceability, post-market surveillance, structured incident reporting — are precisely the data practices that make AI dependable. A fleet managed to MDR standards is a fleet with the audit trail a model needs.
Switzerland’s own infrastructure work makes the point. The federal DigiSanté programme — a ten-year, roughly 400-million-franc national effort running to 2034 under the Federal Office of Public Health — is wrestling with the same data-liquidity problem at national scale: clinical and operational data scattered across incompatible systems, with interoperability as the precondition for anything intelligent built on top. Kainjoo’s work inside that environment, alongside its Allegory Capital and Kainjoo Ventures portfolio, is grounded in a single conviction — that the regulated-market constraint is a forcing function for the integration discipline the rest of the market is now scrambling to learn.
What to do now, by role
The webinar framing is right that AI belongs in HTM. The work is in sequencing it correctly, and the next move depends on the seat.
HTM and clinical engineering directors: audit the CMMS before you audit vendors. Run the AAMI benchmarking metrics against your own records and fix the data-quality outliers first; a model is only as good as the work-order history beneath it. Then pick one high-downtime, high-revenue asset class — imaging is the obvious candidate — and build a single predictive-maintenance workflow end to end, from prediction to parts order to a finance line, before buying anything broader.
CIOs and CISOs: run the AI maintenance roadmap and the device-security roadmap as a single programme. Segmentation, asset inventory accuracy, and patch governance are the entry ticket, and the FDA’s 2025 lifecycle expectations now make them a formal audit item.
OEMs and service vendors: the model is now table stakes. The differentiator is integration — proving that a service writes back into the customer’s CMMS, respects MDR and FDA obligations, and reports avoided downtime in the customer’s own financial terms.
The C-suite: stop counting AI projects and start reading the P&L. Fund the HTM AI that shows up as protected revenue or a lower cost-of-service ratio — in the financial statements, where it counts.
The advantage converts only after the groundwork
HTM holds the cleanest operational data in the hospital, and that is a genuine, structural advantage in a market where most AI fails for want of usable data. On its own, it is the starting line. The advantage converts to value once the data is made to flow, the security foundation is laid, the behavioural pull of the pilot is overcome, and the outcome is defined in financial terms before a model is trained. That makes it an integration problem and a governance problem first, and an algorithm problem a distant second — exactly the work Kainjoo argues regulated industries should take on now, while the rest of the market is still admiring the shiny object.
References
- Fortune. MIT report: 95% of generative AI pilots at companies are failing. https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
- Healthcare IT News. MIT: 95% of enterprise AI pilots fail to deliver measurable ROI. https://www.healthcareitnews.com/news/mit-95-enterprise-ai-pilots-fail-deliver-measurable-roi
- AAMI / Biomedical Instrumentation & Technology. AAMI’s Benchmarking Solution: Analysis of Cost of Service Ratio and Other Metrics. https://array.aami.org/doi/full/10.2345/0899-8205-44.4.346
- Kainjoo. The AI Adoption Challenge: Go Beyond the Shiny Object in 2026. https://kainjoo.life/the-ai-adoption-challenge-go-beyond-the-shiny-object-in-2026/
- GE HealthCare. OnWatch Predict. https://www.gehealthcare.com/services/onwatch-predict
- InterMed. Emerging Trends in HTM — What Hospitals Need to Know in 2025. https://intermed1.com/emerging-trends-in-healthcare-technology-management-htm-what-hospitals-need-to-know-in-2025/
- Cynerio. Visibility Is Not Enough: Key Takeaways from Cynerio’s 2022 State of Healthcare IoT Device Security Report. https://www.cynerio.com/blog/visibility-is-not-enough-key-takeaways-from-cynerios-2022-state-of-healthcare-iot-device-security-report
- HIPAA Journal. More Than Half of All Healthcare IoT Devices Have a Known, Unpatched Critical Vulnerability. https://www.hipaajournal.com/more-than-half-of-all-healthcare-iot-devices-have-a-known-unpatched-critical-vulnerability/
- US Food and Drug Administration. Cybersecurity in Medical Devices: Quality System Considerations and Content of Premarket Submissions (final guidance, June 2025). https://www.fda.gov/regulatory-information/search-fda-guidance-documents/cybersecurity-medical-devices-quality-management-system-considerations-and-content-premarket
- AAMI. Healthcare Technology Management (HTM) focus area. https://aami.org/focus-area/healthcare-technology-management/
- BioAlps. CHF 400 Million for DigiSanté, a programme to transform the Swiss healthcare system. https://bioalps.org/chf-400-million-digisante/
