Bridging the healthcare AI ROI gap

Despite significant investment in AI across the healthcare sector, a substantial proportion of AI initiatives remain confined to pilot programs, failing to deliver measurable return on investment at scale. Sathiyan Kutty, chief AI officer at Emids, examines the structural and operational factors that have widened this ‘ROI gap’, explores where genuine value creation is beginning to emerge, and sets out the conditions required for healthcare organisations to move from fragmented early wins to repeatable, enterprise-wide outcomes.

The healthcare sector has been pouring money into AI for years now and, if anything, that spending is accelerating. However, if we look at how many of those investments have delivered real, measurable value at scale, the picture is considerably less impressive. According to a recent study by MIT, 80 per cent or more of healthcare AI projects never move beyond pilot phase, with some analyses finding that 95 per cent of enterprise AI pilots fail to show any meaningful ROI.

Healthcare is a harder environment for AI than most sectors because the space is more regulated, more fragmented, and deeply dependent on institutional knowledge that has been built up over decades and is rarely documented in structured form. Other sectors like retail and financial services have had an easier run of it than healthcare organisations, and that reality deserves more honest conversation than it typically gets.

The ROI gap has become more visible as the sector has matured. In the early years of healthcare AI, the absence of measurable returns could be explained by the novelty of the technology and the expectation that deployment pipelines would improve over time. That explanation has become harder to sustain. Adoption has accelerated and tools have become more sophisticated, yet the gap between the promise of AI-driven transformation and the reality of operational outcomes at scale remains wide. Understanding why that gap exists is the first step toward closing it.

The learning gap between AI systems and enterprise workflows

One of the least discussed contributors to the ROI gap is the mismatch between how AI systems learn and how healthcare organisations operate. Adoption has risen sharply, with 22 per cent of healthcare organisations now implementing domain-specific AI tools, a seven-fold increase over 2024, but what has not kept pace is the integration of those tools into the broader operational fabric of the organisations using them. Staff are logging in, running queries, and interacting with these systems, but widespread usage is not the same as structural integration.

Most AI systems in healthcare today operate with limited context. They can complete discrete tasks but struggle to account for the rules, exceptions, and dependencies that define real-world workflows. In a sector where process complexity is the norm, this becomes a fundamental limitation.

A deeper issue is that these systems rarely learn in ways that mirror how enterprises evolve. In many cases, feedback loops are weak or non-existent. Decisions made by the system are not consistently captured, evaluated, and fed back into model improvement in a structured way. As a result, systems remain static even as the environments they operate in continue to change. In healthcare, where policies, regulations, and operational practices are constantly evolving, this lack of adaptive learning creates a widening gap between system outputs and operational reality.

This also highlights the difference between task intelligence and workflow intelligence. Task-level AI can execute a defined function, but workflow intelligence requires an understanding of sequence, dependency, and consequence across multiple steps. Without this, systems cannot effectively manage real-world variability, which is where most operational complexity resides.

The result is adoption without transformation. High engagement can create the impression of success, while masking the absence of meaningful operational change. Tools that operate in isolation may generate activity, but they rarely reshape how work gets done.

Where ROI Is beginning to emerge

Despite the persistence of the ROI gap, there are areas where genuine value creation is becoming visible. The clearest early signals are coming not from clinical AI applications, which carry the heaviest regulatory and integration burden, but from operational and back-office workflows. Patient intake, documentation classification, prior authorisation, and revenue cycle management are all areas where AI-driven automation is beginning to show consistent, measurable results.

For payer organisations, this value is most visible 
in prior authorisation and claims operations, where 
AI is beginning to support decision-making rather 
than just task execution. Systems that can interpret 
policy rules, validate coverage, and identify discrepancies earlier in the process are helping reduce administrative effort while improving turnaround times. 
The impact extends beyond efficiency, influencing provider relationships and member experience in meaningful ways.

High-volume workflows

On the provider side, similar gains are emerging across intake and revenue cycle workflows. AI is being used to verify eligibility in real time, structure unorganised documentation, and identify gaps before submission, shifting error detection upstream in the revenue cycle. 
This reduces downstream rework and improves the 
overall efficiency of financial operations. In many provider environments, even small improvements in first-pass acceptance rates can translate into meaningful financial impact, given the volume and value of claims processed.

There is a structural reason for this. These workflows are high-volume, largely rule-governed, and tied to clear performance metrics such as processing time, accuracy, and reimbursement outcomes. They are also sufficiently removed from direct clinical decision-making, allowing organisations to introduce improvements without the same level of regulatory friction.

Revenue cycle management has become a proving ground for healthcare AI. The combination of complex rules, transaction volume, and financial exposure creates an environment where even incremental improvements translate into measurable returns.

For payers, this extends into payment integrity, where AI systems are being used to detect anomalous billing patterns, identify duplicate claims, and prioritise audit efforts with greater precision.

For providers, the impact is most visible in denial management and coding accuracy, where AI-enabled systems are helping improve first-pass acceptance rates and strengthen revenue predictability.

The shift toward workflow-native and agentic AI systems

Understanding where early ROI is emerging helps clarify what kind of AI architecture is required to deliver it at scale. The dominant model of AI deployment in healthcare to date has been what might be described as a alongside-workflow. These are tools that receive input from existing systems, perform defined tasks, and return outputs for human review. This model has value, but it has reached its limits as a vehicle for enterprise-wide transformation.

The limitation is not just technical, but structural. Systems that operate alongside workflows cannot coordinate across steps, manage dependencies, or adapt dynamically to exceptions. As the number of AI-enabled tasks increases, the operational burden of stitching these outputs together often shifts back to human teams, limiting the overall impact of automation.

What is beginning to replace this model is a workflow-native approach, in which AI systems are designed to operate within operational processes rather than alongside them. These systems manage multi-step execution, handle exceptions, and adapt based on context, functioning as part of the operational infrastructure itself.

In payer environments, this enables a more cohesive approach to prior authorisation, where AI systems can manage the lifecycle of a request from intake and validation to exception handling and compliance tracking.

In provider settings, a similar shift is visible in access and scheduling workflows, where systems coordinate across referral requirements, clinical appropriateness, and network constraints. These processes, which previously depended on manual coordination, are beginning to operate with greater consistency and speed.

The emergence of agentic AI represents a broader shift in how systems interact with operations. Instead of responding to individual prompts, they are capable of navigating complex processes with a degree of autonomy. This creates new opportunities for end-to-end optimisation, while also raising important questions around governance, accountability, and oversight.

The rise of forward-deployed context engineering

Healthcare workflows are shaped by a complex, often undocumented mix of payer rules, regulatory requirements, legacy systems, and institutional practices. These are not variables that can be standardised easily or addressed through generic configurations. They require careful interpretation and translation into system behaviour.

This has led to the emergence of forward-deployed context engineering, teams embedded within healthcare organisations to translate operational complexity into production-ready AI systems.

In payer environments, this is particularly important where policy rules vary across plans, products, and geographies. Without this embedded understanding, systems risk producing outputs that are technically correct but operationally misaligned.

In provider settings, workflows such as clinical documentation improvement and coding are shaped by specialty-specific practices and local variations that are rarely captured in structured form. Embedding this context ensures that systems reflect how work is performed, rather than how it is assumed to function.

Organisations investing in this capability are finding that it significantly improves the transition from pilot to scale by addressing the contextual gaps that often derail implementation.

The conditions for scalable ROI

A pilot that works well in one environment does not automatically succeed in another. The organisations that achieve repeatable ROI tend to share a few common foundations.

Domain expertise is the anchor. AI systems that are not grounded in regulatory, clinical, and operational realities will struggle to deliver meaningful outcomes.

For payer organisations, this often means embedding expertise around benefit design, policy interpretation, and claims adjudication logic. For providers, it requires a deep understanding of clinical workflows, documentation patterns, and coding dependencies that directly influence reimbursement outcomes.

Change management is equally critical. In payer environments, trust in AI-driven decision-making must be built carefully among teams responsible for utilisation management and claims review. In provider settings, the challenge is more immediate, as workflow disruptions can directly affect patient experience if not introduced thoughtfully.

Compliance and governance are ongoing commitments rather than one-time requirements. As AI systems take on greater operational responsibility, the need for explainability, auditability, and accountability becomes central to sustained adoption.

Finally, scalable architecture enables repeatability. Organisations that sustain ROI build systems that can extend and evolve over time, rather than requiring reinvention with each deployment.

From adoption to operationalisation

The question in healthcare AI has moved on. It is no longer whether organisations should invest in AI, or even whether they are using it. The question is whether those investments are producing sustained, scalable value.

About 60 per cent of healthcare leaders who have implemented AI solutions are either already seeing a positive ROI or expect to, which is an encouraging sign. Early success in areas like revenue cycle and intake demonstrates that value is achievable, but not automatic.

Across payer and provider organisations, a consistent pattern is emerging. The deployments that deliver sustained value are not those that automate isolated tasks, but those that reshape workflows end to end and embed intelligence directly into decision points.

At its core, the ROI gap is not a technology problem, but an execution problem. The organisations that are closing it are not necessarily those with the most advanced models, but those that have aligned architecture, operations, and domain expertise in a way that allows AI to function as part of the enterprise system rather than as an external capability.

What separates success from stalled pilots is disciplined execution. Leading organisations make deliberate architectural choices, invest in domain-aligned expertise, and treat governance and adoption as ongoing priorities. The result is not just improved outcomes from individual deployments, but the creation of an AI capability that strengthens over time, one that is embedded into the fabric of how work actually gets done.