How AI Is Evolving in Healthcare: From the 2024 Predictions to What Actually Happened

Scoring the 2024 predictions

The 2024 consensus forecast was that AI would transform early detection, personalise treatment planning, compress drug discovery and make remote monitoring continuous. Two years on, the honest assessment is mixed in a very telling way. The predictions closest to existing clinical workflow have come true; the predictions that required new evidence, new regulation or new liability arrangements have not.

Early detection: real progress, quiet deployment

Early detection was the headline prediction, and it's the one where real progress is easiest to miss because of how it was delivered. Rather than a system announcing a diagnosis, what shipped was prioritisation: models reading studies as they come in and reordering the queue so that the concerning ones are seen first.

It's a smaller claim than the 2024 forecasts made, and a far more useful one. Detecting a finding two days earlier simply because it moved to the front of the worklist is a genuine clinical improvement, achieved without asking a regulator to approve an autonomous diagnosis or a hospital to accept unfamiliar liability.

• Screening triage. Models pre-read screening studies and flag the subset that requires priority attention, compressing time-to-review where volume is highest.
• Deterioration prediction on the ward. Vital-sign and lab trajectories feeding early-warning scores — effective when calibrated to a particular population and then monitored, unreliable when transplanted to a different hospital.
• Opportunistic findings. Detecting incidental but consequential findings in studies acquired for some other purpose — a genuinely new capability rather than just an acceleration of an existing one.
• The false-positive economy. Every early-detection system has to trade sensitivity against the downstream cost of investigating the cases it flags. That trade-off is a clinical decision, not a threshold to optimise in isolation.

Personalised treatment: why it slipped

The most confident of the 2024 predictions concerned treatment personalisation: models looking at a patient's history, genomics and response data to recommend the therapy most likely to work for that individual. The modelling was never the obstacle. The obstacle is that recommending a treatment requires evidence that the recommendation improves outcomes, and that evidence has to be generated prospectively.

• Retrospective data encodes past decisions. A model trained on which treatments patients actually received learns the prescribing patterns of the clinicians in that dataset, including their biases and constraints — not which treatment would have worked best.
• Confounding is structural. Sicker patients get different treatments. Without careful causal design, a model can conclude that a treatment causes bad outcomes when in fact it was given because the patients were already sicker.
• The evidence bar is a trial. You can't demonstrate that an AI-guided treatment pathway leads to better outcomes without running a study. That takes years, and it doesn't speed up because the model has improved.
• Narrow beats general. Where personalisation has genuinely taken hold — specific therapy selection in oncology guided by well-characterised markers — it's narrow, heavily validated and unglamorous.

Drug discovery: acceleration in the wrong half

Computational drug discovery produced some of the most impressive technical results of the period. Structure prediction became routine, generative models proposed viable candidate molecules, and screening that once took months was dramatically accelerated.

The complication is arithmetic. Discovery and preclinical work are a minority of the time and cost of bringing a drug to market; clinical trials dominate both. Halving the discovery phase moves the total timeline far less than the headlines suggested, and trials didn't get faster because the candidate was found computationally.

• Target identification improved genuinely. Models that integrate multi-omic and literature evidence surface plausible targets much faster than manual review, and this compounds across a portfolio.
• Candidate generation is real. Generative chemistry produces synthesisable molecules with the desired properties, shifting the effort from designing candidates to filtering them.
• Trial design benefits quietly. Patient stratification, site selection and recruitment forecasting are ordinary predictive modelling applied where delay is enormously expensive — and this may matter more than the chemistry.
• Validation still governs. A computationally proposed molecule still has to get through assays, animal models and human trials, and the failure rate at each stage has not meaningfully changed.

Remote monitoring: succeeded by getting smaller

Remote patient monitoring is the prediction that quietly worked, and it worked by abandoning the grand vision. The 2024 vision was continuous general health inference from wearable signals. What actually deployed is much narrower: specific signals feeding specific, validated alerts for specific conditions.

The lesson generalises well beyond healthcare. A system that answers one question reliably gets deployed; a system that attempts general inference produces alerts no-one can act on, and gets switched off.

• Alert fatigue is the failure mode. A monitoring system that generates more alerts than staff can triage is worse than no monitoring at all, as it trains people to dismiss the whole category.
• Signal quality dominates. Consumer-grade sensors generate artefacts constantly. Most of the engineering effort goes into distinguishing a genuine arrhythmia from a loose strap.
• Someone must be responsible for the alert. Continuous monitoring without a defined clinical response pathway generates liability rather than benefit — it's a workflow question, not a model question.
• Chronic conditions fit best. Managing a known condition with well-defined thresholds is a tractable problem. Detecting unknown conditions from ambient signals largely is not.

What this predicts about the next two years

If the 2024 forecasts erred by assuming model capability was the constraint, the useful correction is to forecast against the real constraints: evidence generation, workflow integration, regulatory clarity and who is ultimately responsible for the output.

• Expect the administrative layer to keep compounding. Prior authorisation, coding, referral triage and scheduling involve no clinical claim, deliver immediate measurable value and transfer no liability. That's where near-term progress will concentrate.
• Expect agentic systems in operations before clinical use. Multi-step agents that coordinate scheduling, follow-up and documentation are arriving. Agents that make clinical decisions are not, and shouldn't be rushed.
• Expect evidence infrastructure to become the bottleneck worth solving. The organisations that can run prospective evaluation efficiently will deploy clinical AI years before those that can't.
• Expect monitoring to become mandatory in practice. As deployments age, silent degradation across sites is the failure that leads to incidents, and the requirement to detect it will harden.

The through-line is that healthcare AI rewards narrow, well-specified systems with a human accountable for the output. That's not a temporary state of the technology; it's what deploying software into clinical responsibility looks like.

How SoftmaxAI approaches healthcare AI

We build our healthcare systems the same way we build everything else. We have a written, client-approved specification prior to any code being written. We have an agent-accelerated build led by senior engineers, and every decision that matters is owned by a human.

• Forecast on constraints, not capability. During the specification we identify what will actually gate deployment — evidence, integration, regulatory status — so the timeline reflects reality rather than model progress.
• Start where the claim is smallest. Operational and administrative applications return value in weeks, and that in turn builds the institutional confidence that clinical applications later require.
• Clinical ownership throughout. A named clinician co-authors the specification and owns the escalation, the refusal behaviour and what success means in the clinic.
• Built on your cloud. As a Google Cloud Partner we build natively on Vertex AI and Gemini where that fits your compliance posture, deployed into your own environment.

See it in practice in our patient onboarding and scheduling case study, or start with the AI Opportunity Sprint for an honest feasibility assessment before you commit to a build.