Holland Bloorview Kids Rehabilitation Hospital's electronic medical record holds about 51 million lines of clinical text. When the hospital put an artificial intelligence (AI) coding agent to work on that record, the decisive work had already been done years earlier, in governance, stewardship, and shared definitions. The model mattered far less. That is the lesson for any organization planning data foundations for agentic AI: an agent is only as trustworthy as the data, context, and decision records underneath it. The foundation is the strategy.
This article draws on what we learned building that coding copilot alongside Holland Bloorview and Snowflake, and on what we see across healthcare data programs. We’ll set out five foundations, a way to tell which stage you are in, and the order in which to build.
What are the necessary data foundations for agentic AI?
Data foundations for agentic AI are the governed data, shared context, and decision records an AI agent needs to act reliably on an organization's behalf.
Analytics foundations were built to answer questions. A dashboard shows a number, and a person decides what it means. Agents go further. They read documents, propose actions and, in some workflows, carry them out. Every gap a person once filled with judgment (which definition applies, which document counts, what this unit calls a service) now has to be written down where the agent can find it.
That raises the bar in three places. The data has to be trusted at its source, the meaning behind the data has to be explicit, and every decision the agent influences has to be traceable, so people can see what happened and the system can improve. Organizations that have already invested in governance for reporting are closer than they think. The work ahead is to extend those foundations, not replace them.
How Holland Bloorview Kids Rehabilitation Hospital Made it Trustworthy
Holland Bloorview is Canada's largest pediatric rehabilitation hospital. Its patients often live with several diagnoses at once, which makes medical coding demanding: a critical diagnosis can sit in an admission note, a consult, a nursing note or the discharge summary. The hospital asked a focused question. Could AI help coders find the right evidence faster while keeping coding quality high?
The answer depended on groundwork. The hospital runs as a learning health system, cycling data into knowledge, knowledge into practice and practice back into data. It built data stewardship frameworks that define who owns each dataset and how quality is managed across its full lifecycle. On Snowflake, the team automated a full medallion architecture and brought clinical, HR, finance, patient experience and research data into one governed platform. It then wrote down the context: what a service such as complex continuing care means, and how occupancy is defined for a service versus a physical unit.
The coding copilot sits on top of all of this. It reviews source documents, proposes ICD-10-CA codes with confidence, and links each suggestion to the sentence that supports it. Deterministic, rule-based automation fills roughly half of each abstract, and a machine learning model and a retrieval-augmented generation model handle the rest. Coders accept, replace, reject or add codes, and every decision is recorded with a reason.
16,000+ ICD-10-CA codes in the classification coders work across 20 to 25 minutes to code and abstract one inpatient record About 19,000 records coded each year Source: Keyrus, Holland Bloorview Kids Rehabilitation Hospital, and Snowflake webinar, 2026
Understanding Where Your Organization is Today
Most organizations fall into one of three stages, and each stage has a clear next step.
Holland Bloorview's own path offers a useful map. Before 2023, its analytics work was largely request-driven. In 2024 and 2025, it shifted to self-serve products backed by strong governance. It now runs agent-enabled delivery across the hospital. We see the same three stages across the healthcare organizations we work with.
Stage 1: Request-driven | Stage 2: Governed self-serve | Stage 3: Agent-ready | |
How work starts | A ticket to the data team | Users open a governed product | Users describe a problem; agents and products respond |
Where definitions live | In analysts' heads and report logic | In a shared glossary and certified models | In a context layer agents read directly |
Data in scope | Structured records, mostly one system | Structured data across systems | Structured data plus notes, PDFs and forms |
Role of AI | Pilots and forecasting models | Predictive models inside products | Agents that propose and act, with people deciding |
Next step | Stewardship and quality ownership | Explicit context and decision capture | Feedback loops and performance monitoring |
If you recognize yourself in Stage 1, the fastest progress comes from naming data owners and agreeing on definitions. Stage 2 organizations are closest to agents, because the hardest cultural work is done. Stage 3 becomes an operating rhythm that keeps improving with every decision.
5 Foundations That Make Data Ready for Agents
What worked at Holland Bloorview applies well beyond coding and pediatrics. It rests on five foundations that any organization can assess and build. Together, they describe the base an organization needs to reach the Human Orchestrated Model™ (HOM™), where intelligence foundations, human command, and performance steering work as one system.
Trusted data: Every critical dataset has a named owner, a steward, and a managed quality lifecycle, starting at the moment data is entered.
A unified platform: Structured records sit alongside clinical notes, PDFs and referral forms in one governed environment, refined in layers so each use case draws on data at the right level of maturity.
Engineered context: Definitions, terminology and local rules are written down where agents can read them, so an agent knows what your organization means by its own words.
Evidence by design: Data stays inside your own account, identifiers are removed before a model sees the text, and every suggestion points to the sentence it came from. People make the decision.
Decision memory: Each accept, change, or rejection is captured with a reason, reviewed for patterns, and used to test and improve the agent against gold-standard data.
Context Engineering Matters More than the Model You Pick
The first instinct on most AI projects is to choose the strongest model and let it work things out. In practice, even a strong model performs poorly with the wrong context. Context engineering is the discipline of deciding which documents an agent sees, in what form, and which local rules it must follow. It is where most of the effort in an agentic project should go.
The coding copilot showed why. Some knowledge lives only in coders' heads. Holland Bloorview codes an admitting diagnosis for every inpatient to capture the clinical reason for admission, and its clinicians abbreviate brain injuries in ways a general model will not recognize. The team supplied those rules and abbreviations directly. Early on, the agent also tended to over-code, flagging every mention of a symptom. A note that a patient is at risk of falling is different from a fall during the stay, and context is what teaches the agent that difference.
Context also drives cost. Every document an agent reads consumes tokens on every chart. Our rule of thumb is simple: get the context right on a mid-sized model first and move to a larger one only when the evidence is in place and answers still miss.
Trust is earned one recorded decision at a time
Trust starts with privacy. Whatever the jurisdiction, the questions are the same: who sees the record, where does it go, and can you show who did what? In this design, the chart, the suggestions and every coder decision stay inside the hospital's own Snowflake account, where the model also runs. The pipeline removes identifiers, so the model reads clinical text only, while coders see the full chart under role-based access and logging.
"When you put AI in front of anybody, especially in health care, if it gives a wrong answer, it's not going to be trusted." Joe Warbington, Industry Principal for Healthcare, Snowflake (2026)
Trust grows through evidence. Every suggestion shows the exact sentence it came from, in the document coders already know, and each code is checked against the full ICD-10 classification so an invented code cannot pass. Confidence scores guide the coder's attention, and the coder still decides. Small frictions matter too. Hidden evidence or an extra click gives people a reason to stop using the tool.
Trust lasts through memory. Each decision is stored with a reason, the source sentence, a name, and a timestamp. Recurring corrections surface as candidate rules, and Holland Bloorview's health information management team decides each month which ones reach the model. Behind that sits a full AI governance lifecycle: model outputs are tested against gold-standard data, monitored against thresholds that tell the team when to pause, and reported openly to leadership.
How to build the foundations in the right order
The order matters as much as the parts. Organizations that follow Keyrus’ HOM™ tend to build its three layers in sequence, using one well-understood workflow as the proving ground.
Choose a workflow your experts know well. Run detailed discovery sessions with the people who do the work, and confirm the process can be recreated from defined steps and information you already hold. In our experience, design becomes the easier part once discovery is thorough.
Lay the Intelligence Foundations. Name owners for the data the workflow depends on, bring structured and unstructured sources onto one governed platform, and write down the context: definitions, abbreviations and local rules.
Keep people in command. Keep the data behind your own sign-on, show the source sentence for every suggestion, and give users clear choices to accept, change or reject, each with a reason.
Steer performance from day one. Capture a baseline before launch, record every decision, review patterns on a regular rhythm, and test the agent against gold-standard data.
Starting points vary. One regional partnership of four hospitals we work with in Ontario shares a single electronic health record across autonomous organizations. It chose a federated model: definitions, privacy and technology standards are set regionally, while each site keeps control of local workflows. That shared language is exactly the base an agent will need when the partnership is ready for one.
Getting Started: Start with one workflow your experts know well
The fastest path to agentic AI runs through the foundations you already have. If you are deciding where your first agent should go, our healthcare team can help you assess your data foundations for agentic AI against the five foundations and identify the workflow most ready to start. Talk to our healthcare team.
AI does not transform businesses. Architected intelligence does.
Usually no. Most organizations extend the governed platform they already run by adding unstructured data and an explicit context layer.
Start with a mid-sized model and invest in context. Move to a larger model only when the evidence is in place and results still fall short.
Keep the data and the model inside your own environment, remove identifiers before the model reads the text, and log who did what. Each province and state has its own rules, so involve your privacy office early.
Test its outputs against gold-standard data and track how often experts accept, change or reject suggestions. Measure time per task against a baseline captured before launch.
No. The agent does the searching and drafting, while experts apply standards and make every final decision.
CoCo is the builder's agent: it creates semantic models, code, and governed data assets inside Snowflake's operating context. CoWork is the business user's agent: it consumes those assets to answer questions, run investigations, and trigger actions. CoCo builds the capability; CoWork puts it in the hands of the business.
Snowflake CoCo (formerly Cortex Code) is a context-aware coding agent built natively into the Snowflake platform. It works in Snowsight, a desktop app for macOS and Windows, and a CLI, and understands your specific environment, including schemas, RBAC permissions, governance policies, and data relationships, to generate and modify code and semantic assets through natural-language prompts, with human review remaining important before production deployment.
Snowflake CoWork (formerly Snowflake Intelligence) is a conversational agent that lets business users investigate, analyze, and act on governed Snowflake data using natural language. It runs multi-step research, produces reusable artifacts, schedules recurring automations, and can take action in external tools like Google Drive, Salesforce, and Slack through MCP connectors.
The Human Orchestrated Model™ (HOM™) is Keyrus' proprietary structured framework to design, operate, and scale intelligence within organizations. HOM™ is built on three interdependent layers: 1. Intelligence Foundation 2. Human in Command 3. Performance Steering Three layers that turn data, agents, and human governance into measurable performance. Data enables intelligence. Governance directs it. Measurement improves it.
No. AI amplifies whatever foundation it's built on. Ungoverned data leads to untrustworthy AI outputs, regardless of how sophisticated the model is. Governance is what makes an organization genuinely AI-ready.
In most industries, conflicting data is a scheduling headache. In healthcare, it can delay staffing decisions, cause operating room underutilization, or point care teams toward the wrong intervention, making governance a patient-safety and operational issue, not just a technical one.
Governance should always come first. A dashboard built on ungoverned data just delivers faster access to unreliable numbers. Establishing shared definitions, data lineage, and cataloging must happen before any reporting layer is built.
Most health systems are "data-rich, insight-poor." Digitization created massive amounts of structured data, but without governance and shared definitions, that data stays fragmented across departments, making it hard to turn into reliable decisions.
A unified data platform is the infrastructure that connects fragmented healthcare data, clinical, operational, financial, and specialty systems, into a single, governed source of truth. It replaces inconsistent departmental definitions and one-off reporting with certified data models everyone can trust.