Artificial intelligence is now embedded across a growing range of healthcare workflows. As of 2025, healthcare captures about 43% of all AI investments, representing roughly $1.5 billion (€1.38 billion) in spending. Silicon Valley Bank data indicates that 46% of healthcare investment is directed toward AI technology, while Deloitte reports that organisations dedicate an average of 36% of their digital initiative budgets to AI-driven projects. AI-enabled tools are already used in clinical and administrative areas, including clinical decision support and revenue cycle management. These tools are intended to improve patient care and financial performance, but their effectiveness depends on the quality of the data on which they are built. When data is incomplete, inaccurate or inconsistent, AI can reproduce those weaknesses at speed and scale. The result is not only poorer performance, but also greater operational risk, reduced confidence in outputs and more difficulty in expanding AI beyond isolated use cases. Data quality remains central to whether AI delivers value or magnifies existing problems.
Poor Data Quality Creates Four Immediate Risks
Poor data integrity remains a major obstacle to AI adoption. A Q4 2025 survey of revenue cycle leaders found that 74% identified poor data quality as the primary barrier to successful AI adoption. The problem lies not in the complexity of the neural network, but in the integrity, completeness and accuracy of the underlying data.
Bias remains one of the clearest risks. Models trained on large aggregate datasets or on narrow, specific datasets can still be unrepresentative of the local patient population. Data from large urban academic medical centres may fail to translate to rural or community-based settings. When a model encounters unfamiliar clinical markers or documentation norms, it may ignore critical signals or display algorithmic overconfidence despite lacking context. Missing data creates an added problem. The absence of a data point does not necessarily mean the absence of a clinical issue, but AI trained on poor data may interpret it that way.
Documentation gaps create a second risk. AI is often presented as a way to close clinical documentation gaps, yet it cannot recognise what it has not been trained to see. If specific patient populations or rare care pathways fall outside historical norms, omissions may go unflagged. In non-standard cases, over-reliance on automation can create a false sense of security where human clinical intuition remains necessary.
Errors and Mistrust Can Spread at Scale
Scale is one of AI’s main value propositions, but scale also amplifies risk. A documentation or coding error made by a human is typically localised to a single encounter. When that same error exists within a training dataset, AI can propagate the inaccuracy across thousands of encounters at machine speed. Without robust data governance, inaccuracies can become system-wide failures that are difficult and costly to remediate.
Trust is affected as well. Frontline experts can lose confidence in AI solutions when outputs generate hallucinations, false positives or questionable clinical recommendations. Once clinicians and other end users lose trust in one output, scepticism can spread to other digital transformation initiatives. Credibility depends on consistent accuracy.
Must read: Data Quality in Federated Health Networks
Autonomous coding illustrates how historical data quality shapes AI performance. These systems are typically trained on historical charts coded by humans, which means they inherit both the strengths and weaknesses of those records. If historical accuracy averages 90%, the AI cannot reach a 95% clean claim standard without significant intervention. That creates a data quality gap that must be bridged before moving to direct-to-bill automation. Without auditing and correcting historical data, full automation is delayed, requiring costly validation steps and hybrid workflows that offset the original return on investment of the AI.
Broader Rollout Depends on Data Stewardship
McKinsey’s 2025 State of AI report found that nearly 2/3 of organisations have yet to scale their AI projects across the enterprise. Broader implementation depends on a disciplined data-first approach.
Four priorities define that path. The first is auditing historical records to identify systemic inaccuracies in the data used for model training or tuning. The second is establishing accuracy baselines by defining what success looks like in a manual environment before measuring AI performance. The third is remediating known gaps by addressing inconsistencies in documentation and coding standards prior to automation. The fourth is maintaining human-in-the-loop oversight so that clinical and technical governance remains central to the deployment lifecycle.
These priorities place data stewardship at the centre of implementation. AI acts as a force multiplier, but it only multiplies the quality of the foundation beneath it. Stronger data integrity supports more reliable rollout. Weaker data integrity allows bias, omissions and inaccuracies to spread faster across workflows. The central issue is therefore not only whether AI can be introduced into more settings, but whether the underlying data can support wider deployment without amplifying error and uncertainty.
Healthcare continues to invest heavily in AI, yet broader adoption remains constrained by the quality of the data beneath these systems. Poor data can scale bias, preserve documentation gaps, industrialise errors and weaken clinical and operational trust. Autonomous coding shows how historical inaccuracies can delay automation and reduce expected returns by creating a need for added validation and hybrid workflows. A more dependable path to scale begins with auditing historical records, setting accuracy baselines, addressing known inconsistencies and maintaining human oversight during deployment. AI can improve clinical and administrative performance, but only when the data foundation is accurate, complete and fit for use.
Source: MedCity News
Image Credit: iStock
