> ## Content Index
> Fetch the complete content index at: https://www.healthdatacon.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# Eddie Pérez-Ruberté Gives Health Systems a Practical Path to AI Governance
- URL: https://www.healthdatacon.io/articles/eddie-perez-ruberte-gives-health-systems-a-practical-path-to-ai-governance/
- Published: 2026-10-16T16:25:24.000Z
- Updated: 2026-10-16T16:25:24.000Z
- Description: At HealthDataCon 2026, Eddie Pérez-Ruberté outlined a practical approach to AI governance built on system inventories, risk assessment, accountable oversight and a 90-day implementation plan.
- Author: HDM Staff
- Tags: Asia-Pacific

A chief information officer receives another request for an AI tool. A physician wants faster documentation. Another team wants help identifying deteriorating patients. Each demonstration looks promising, but the CIO cannot establish what data the system uses, who owns its quality or how the organization will manage failure. The requests remain on hold.

Eddie Pérez-Ruberté, MSAAE, LSSBB, founder and CEO of Areito Group, used that fictional scenario to open his HealthDataCon 2026 presentation, From Paralysis to Practice: A Structured Way to Identify, Quantify, and Manage AI Risk in Health Systems. His argument was that hesitation often reflects missing decision structures, rather than resistance to technology.

“I call governance the runway from which your innovations take off, one after the other, fast and safely,” Pérez-Ruberté said. His proposed route forward combined an inventory, multidisciplinary oversight, risk classification and continuing review, with a 90-day plan for making those practices operational.

## Find the AI already inside the organization

Waiting to approve a new purchase does not necessarily prevent AI from entering a health system. Pérez-Ruberté noted that vendors can introduce capabilities through updates to existing products. Governance therefore begins with discovery, including tools embedded in familiar workflows and systems being used beyond their original purpose.

“We cannot govern what we cannot see,” he said. An inventory should identify each system, its vendor, its owner and its intended use. It should also document data sources, protected health information, connected systems, affected populations and mechanisms for human intervention. Procurement records, software licenses, department surveys and interviews provide complementary starting points.

He paired that inventory with a maturity assessment across eight domains, from strategy and data quality to human oversight and vendor management. One example showed strong self-reported data governance alongside weaker vendor oversight. The assessment illustrates where teams perceive gaps; it should not be mistaken for independent assurance that a system is safe.

Pérez-Ruberté grounded his approach in NIST’s AI Risk Management Framework and ISO/IEC 42001\. The [ISO standard](https://www.iso.org/standard/42001?ref=healthdatacon.io) establishes requirements for an AI management system, giving organizations a structure for policies, responsibilities and improvement. A preliminary questionnaire can help organize that work, but completing one is different from demonstrating conformity to the standard.

![Illustrative maturity assessment shows stronger data governance and weaker vendor and human oversight scores.](https://storage.ghost.io/c/91/db/91db0927-dd9a-4578-9944-921da70f6885/content/images/2026/10/ai-governance-maturity-radar.jpeg)

Pérez-Ruberté used a self-reported maturity assessment to identify governance gaps. Slide credit: Eddie Pérez-Ruberté / HealthDataCon 2026.

## Make data choices part of the risk assessment

The data community has a direct role even when a vendor controls the model. Pérez-Ruberté identified three areas organizations can examine: the quality of their input data, its provenance and lineage, and its representation of the patients who will be affected. Completeness and accuracy remain essential when software acts on what it receives.

His example was the widely studied population health algorithm that used healthcare spending as a proxy for need. A [2019 study](https://pubmed.ncbi.nlm.nih.gov/31649194/?ref=healthdatacon.io) found that Black patients were substantially sicker than White patients at the same risk score. Unequal spending made the proxy misleading. Removing race as an explicit input did not remove the bias embedded in the prediction target.

For leaders, the lesson is to ask what a model is actually predicting and whether that target fits the intended decision. A technically accurate estimate of spending can still be unsuitable for allocating clinical support. Validation must examine the relationship between the model’s output, the local population and the action users take.

The [NIST framework](https://www.nist.gov/itl/ai-risk-management-framework/ai-risk-management-framework-faqs?ref=healthdatacon.io) describes seven interconnected characteristics of trustworthy AI, including validity and reliability, safety, transparency, privacy and fairness with harmful bias managed. Pérez-Ruberté connected these characteristics to data stewardship. Lineage supports accountability; representative data supports meaningful evaluation. Neither replaces testing within the workflow where the tool will operate.

## Use a risk score to structure judgment

Pérez-Ruberté proposed an AI strategic council with clinical, technology, privacy, legal, informatics, operational and quality expertise. Its purpose is to make accountable decisions with fewer blind spots. The council also needs a policy defining permitted uses, prohibited uses, data handling and the organization’s tolerance for risk.

His proposed risk classification number multiplies three ratings: severity of potential harm, the likelihood that failure goes unnoticed, and reach across patients or staff. Each factor receives a score from one to three, producing a total between one and 27\. His thresholds classify scores of six or below as low risk, seven through 15 as medium risk, and 16 or above as high risk.

Those thresholds belong to the speaker’s method, not to a universal NIST or ISO classification. They can guide the intensity of review, but multiplying ordinal judgments does not produce a calibrated probability of harm. A severe, unusual failure still deserves explicit attention even when other factors lower the combined score.

“It is a sort of a subjective analysis,” Pérez-Ruberté acknowledged during questions. He emphasized bringing knowledgeable people into the assessment and defining what each rating means. Recording the reasoning and uncertainty behind a score makes that discussion more useful than the number alone.

Intended use also matters. Asked about a tool moving from an administrative task into clinical care, he recommended reassessment: the severity rating may change, requiring stronger controls. Impact assessments then examine both anticipated benefits and potential harms. NIST’s [AI playbook](https://www.nist.gov/itl/ai-risk-management-framework/nist-ai-rmf-playbook?ref=healthdatacon.io) offers suggested actions across Govern, Map, Measure and Manage, reinforcing an ongoing process rather than a single approval event.

## Give oversight authority and a reporting channel

Approval is the beginning of operational accountability. Model updates, changing data and new workflows can alter performance even when a deployed model does not continuously learn. Pérez-Ruberté called for oversight requirements matched to the risk tier, supported by staff who understand the output and can act on concerns.

“We give humans the authority to actually override the AI,” he said. Authority alone is insufficient if reviewers lack the competence or information to exercise it. The [transparency principles](https://www.fda.gov/medical-devices/artificial-intelligence-enabled-medical-devices/transparency-machine-learning-enabled-medical-devices-guiding-principles?ref=healthdatacon.io) jointly developed by the FDA, Health Canada and the UK’s MHRA for machine learning-enabled medical devices emphasize the performance of the human-AI team and information appropriate to users.

His incident examples included inaccurate outputs, unexpected behavior, patient data exposure, bias and use outside the approved purpose. He proposed nonpunitive reporting, triage and authority to suspend tools. These processes complement existing patient safety reporting; they do not replace it. Response timing must reflect the actual harm, rather than treating a sample review window as permission to delay urgent action.

He also distinguished early concerns from confirmed incidents, encouraging low-barrier reporting and feedback to identified submitters. This supports the broader principle in [WHO guidance](https://www.who.int/news/item/28-06-2021-who-issues-first-global-report-on-ai-in-health-and-six-guiding-principles-for-its-design-and-use/?ref=healthdatacon.io) that humans should remain in control of healthcare systems and medical decisions.

![A 90-day AI governance plan moves from foundations to controls and operational review.](https://storage.ghost.io/c/91/db/91db0927-dd9a-4578-9944-921da70f6885/content/images/2026/10/ai-governance-90-day-plan.jpeg)

Pérez-Ruberté’s proposed 90-day plan connects assessment, controls and continuing review. Slide credit: Eddie Pérez-Ruberté / HealthDataCon 2026.

## Turn the first 90 days into operating practice

Vendor oversight was a recurring concern. Pérez-Ruberté urged teams to ask about training and validation data, local applicability, storage and whether customer data will train models. Legal and procurement teams can then address appropriate audit rights, data rights and responsibilities. Existing tools deserve this scrutiny alongside proposed purchases.

His 90-day roadmap starts with assessment, inventory, leadership commitment, policy and assigned roles. The second month introduces impact assessments, vendor review and documented controls. The third puts human oversight, monitoring, internal review and improvement into practice. Workforce education should explain approved uses, data safeguards and escalation routes in terms people can apply.

For organizations trapped in endless review, his smallest next step was an inventory paired with a current-state assessment. “We need to understand what we have so that we can govern it,” he said. That work gives the fictional CIO a concrete basis for deciding what can proceed, what needs additional evidence and who will remain accountable after deployment.