The Evidence Health Systems Are Not Yet Generating

We have built frameworks for measuring healthcare outcomes. We have not yet built the infrastructure that would tell us why those outcomes are what they are.

Healthcare governance in the United Kingdom has made progress over two decades. The audit culture, the regulatory frameworks, the outcome reporting infrastructure that now underpins everything from CQC inspection to NHS England oversight - these represent institutional investment and, in many cases, improvement in the visibility of care quality. It would be wrong to dismiss them.

But they share a blind spot which are processes. Between the inputs that any healthcare organisation receives (patients, resources, staff, guidelines) and the outcomes it produces (mortality, readmission, safety events, patient experience) lies a sequence of actions. Who did what, for whom, when, in what order, with what fidelity to the evidence-based standard. This is the process layer. And for most organisations, for most of the time, it is invisible.

Visible Only at the Extremes

This process layer is genuinely difficult to instrument. Care is distributed, highly-pressurised, and highly variable, not in the way that manufacturing processes are variable, but in the way that human judgment under uncertainty is variable. The patients, teams and shifts are different.

The consequence is that the process layer has been visible only at the extremes: when something goes wrong and a review is conducted; when a trust prepares for CQC inspection and data is assembled retrospectively; when a never-event triggers an investigation and the sequence of actions preceding it is reconstructed. In between these moments, the process is largely unobserved by those responsible for governing it.

The gap between guideline and practice — between what the standard says and what happened for this patient, on this shift — is where most preventable harm lives. We know this. We just cannot see it in real time.

The pattern in serious incident reports is, by now, depressingly familiar. What failed was not unusual. It was not a novel situation for which no guidance existed. It was a gap in adherence to a process that was specified, taught, and theoretically followed — a gap that, in retrospect, looks systematic rather than individual, the product of structural conditions rather than personal failure.

What Becomes Possible When the Process Layer Is Visible

Imagine a plumber able to see inside transparent pipes to work out the where a blockage or leak can be found and understand why the water pressure is too high or low. But visibility alone is not the end of the argument, it is the beginning.

When process data is collected at scale and linked to outcome data, something becomes possible that neither randomised trials nor retrospective audits can produce: provider-specific, population-specific, pathway-specific evidence about what works in this hospital, for these patients, under these operating conditions. Not a replacement for the clinical evidence base — rather, a complement to it: the localising layer that tells you whether the evidence-based process is producing evidence-based outcomes in your setting, with your cohorts, your staffing model, your EPR.

This evidence accumulates so a trust with two years of linked process-outcome data across its pathways knows something about the relationship between process adherence and patient safety and quality that it could not know without it. The data compounds in value over time: it captures seasonal variation, workforce change, the impact of EPR upgrades, the effect of policy revisions on actual practice.

This is a different kind of organisational asset from software capability, consultant insight, or regulatory compliance. It is evidence, generated from within, about how care works in particular instances.

The Agentic AI Question

The arrival of agentic AI systems in healthcare adds a new dimension to this argument - one that the field has barely begun to grapple with seriously.

Agentic systems can enforce processes at scale, handle routine complexity with consistency that human teams under pressure cannot always maintain, and free clinical judgment for the situations that genuinely require it.

But agentic systems also raise a governance question that no individual supplier is well-positioned to answer. When an AI system is running processes across thousands of patient interactions, who is responsible for knowing whether those processes are the right ones? Who tracks whether the actions taken by AI are collectively producing better outcomes or whether they are systematically missing a patient cohort, or failing to reflect an update to clinical guidance?

The governance of AI in healthcare is not primarily a question of explainability or data security. It is a question of process-outcome accountability, the same question we have always had about human care, now applied to automated systems at scale.

Each agentic supplier governs its own workflow. No supplier governs whether the collective behaviour of all the AI systems an organisation has deployed is improving patient safety. No supplier connects those automated actions to a board-level assurance picture. That connective, cross-system, outcome-linked governance layer is currently unoccupied in almost every health system.

Instrumented Governance

The organisations beginning to close this gap are building something that has not previously existed in health systems: a continuous, outcome-linked process intelligence function — what might be called instrumented governance. Not governance as periodic audit and retrospective review, but governance as a live, data-generating activity that connects process execution to outcome evidence in real time, across human and automated care alike.

This requires, at minimum, three things. A mechanism for making the process layer continuously visible rather than sporadically audited. A means of linking that process data to outcome data at the level of individual pathways and patient cohorts. And a layer of oversight that sits above individual systems — including AI systems — and translates their collective behaviour into the governance language that boards and regulators require.

An Investment in Knowledge

The argument for this kind of investment is not primarily about compliance. Compliance is a floor, not a ceiling, and organisations that build governance infrastructure to satisfy regulators tend to get less from it than those that build it to understand themselves.

The argument is about the quality of the decisions that governance makes possible. A board with continuous, outcome-linked process intelligence is not making judgments about care quality in the dark. An organisation that knows, from its own longitudinal data, which processes drive which outcomes for which patient cohorts, has a kind of operational knowledge that cannot be purchased from a consultant or downloaded from a national dataset.

The governance frameworks we have are appropriate for the information environment in which they were designed. That environment is changing rapidly enough that the gap between what it is now possible to know and what most governance processes currently know is, if anything, widening.

The question is whether the governance architecture is keeping pace or whether the most consequential layer of the system remains, for most organisations, still invisible.