In our recent webinar, "Healthcare Foundations for Agentic AI: Healthcare Coding Accelerator in Action," Keyrus was joined by Holland Bloorview Kids Rehabilitation Hospital and Snowflake to show how a pediatric hospital moved AI-assisted medical coding from idea to daily practice. The session brought together Sri Vijay, Director of Decision Support and Data Analytics; Viji Thayalan, Health Information Manager; and Nikki Ramos, Senior Health Information Analyst, all from Holland Bloorview. They were joined by Joe Warrington, Industry Principal for Healthcare at Snowflake, and from Keyrus, Kyriakos Bilias, Managing Consultant, and Leslie Potgieter, Director of Data Solutions and Alliances, North America.
The message running through the hour was clear. AI coding agents perform when the foundations underneath them are trusted, the context is engineered with care, and coders stay firmly in command of every decision. In case you missed the live session, here is a recap of the foundations, the live demo, and the lessons learned along the way. You can watch the replay here.
Industry perspective: foundations come first
Holland Bloorview is Canada's largest pediatric rehabilitation hospital, with 67 funded inpatient beds and roughly 60,000 outpatient visits each year. Many of its patients live with acquired brain injury, cerebral palsy, orthopedic conditions, or autism spectrum disorder, often alongside several other diagnoses. That clinical complexity shapes everything, including how records are coded.
Sri Vijay explained that the hospital's agentic work rests on years of groundwork. Holland Bloorview operates as a learning health system, moving continuously from data to knowledge to practice and back again. It is currently the only hospital in Canada at HIMSS Analytics Maturity Model (AMAM) Stage 7, and it has reached HIMSS EMRAM Stage 6 for electronic medical record adoption.
The journey followed three clear phases:
Before 2023: analytics and reporting were largely request-driven, with significant manual effort spent transforming data for each question.
2024 and 2025: a strategic shift toward self-serve analytics products for clinicians and operational leaders, built on a strong data governance and stewardship framework covering ownership and the full data quality lifecycle.
2026 and beyond: agent-enabled delivery across the hospital, from internal pipeline and analytics agents to domain-specific agents like the coding copilot.
Key challenges addressed
Viji Thayalan walked through why coding is far more demanding than it sounds. Coders translate clinical documentation into ICD-10-CA diagnosis codes that feed provincial and national databases, supporting reporting, funding decisions, system planning, quality improvement and research. At Holland Bloorview, that work spans inpatient rehabilitation, complex continuing care and ambulatory programs.
Evidence spread across the chart: a critical diagnosis can appear in an admission note, a consult, nursing documentation, a progress note, or the discharge summary.
A vast classification: coders navigate more than 16,000 ICD-10-CA codes while applying detailed coding standards and diagnosis typing rules.
Volume and time: each inpatient abstract takes roughly 20 to 25 minutes and each ambulatory abstract 10 to 12 minutes, across nearly 19,000 records a year.
Scale of documentation: the electronic medical record holds about 1.3 million documents and 51 million lines of clinical text.
During the pilot review, the team saw that both human coders and AI could miss opportunities in large, complex records. That led to a focused question: could AI help coders find relevant evidence faster while keeping coding quality high?
The business case and the future-state workflow
The business case rested on three goals. Efficiency: less time searching through charts. Quality: every suggestion linked to supporting evidence so coders can validate it consistently. New value: richer data for decision-making and broader clinical coverage. Above all, the copilot was designed to enhance coders' expertise, with the coder remaining responsible for review and the final decision.
In the future-state workflow, clinical documentation remains the source of truth. The AI reviews the source documents, identifies potential coding opportunities, and links each suggestion to its evidence. Coders then validate, apply the standards, and make the call. Quality review and submission stay unchanged, so coders spend less time searching and more time applying what they know.
The foundation underneath: data, context and governance
Sri described the infrastructure that makes this possible. Holland Bloorview's enterprise data platform brings together its Meditech electronic health record alongside HR, finance, patient experience and research data. On Snowflake, the team automated its full medallion architecture, from bronze staging through transformation to presentation layers that serve analytics and AI applications.
The layer Sri emphasized most was context. Years of governance work defined the metadata, terminology, and business definitions that an agent needs to reason well: what complex continuing care means as a service, or how occupancy rate differs by service versus physical location. The platform also now goes beyond structured data to clinical notes, discharge documents and referral forms. On top sits an agent mesh of engineering, analytical, and domain-specific agents, surfaced to users through an enterprise catalog that routes each person to the right product for their question, with governance ensuring the right person sees the right data at the right time.
The foundation underneath: data, context and governance
Sri described the infrastructure that makes this possible. Holland Bloorview's enterprise data platform brings together its Meditech electronic health record alongside HR, finance, patient experience and research data. On Snowflake, the team automated its full medallion architecture, from bronze staging through transformation to presentation layers that serve analytics and AI applications.
The layer Sri emphasized most was context. Years of governance work defined the metadata, terminology and business definitions that an agent needs to reason well: what complex continuing care means as a service, or how occupancy rate differs by service versus physical location. The platform also now goes beyond structured data to clinical notes, discharge documents and referral forms. On top sits an agent mesh of engineering, analytical and domain-specific agents, surfaced to users through an enterprise catalog that routes each person to the right product for their question, with governance ensuring the right person sees the right data at the right time.
The AI coding copilot in action
Nikki Ramos then shared her screen for a live walkthrough of the copilot as coders use it every day.
Workload in one place. Before a chart is even opened, coders and managers see which accounts are ready to abstract, who they are assigned to, their workflow status, and how many days have passed since discharge. Managers gain clear visibility into backlogs and turnaround, and coders spend less time tracking work through emails and spreadsheets.
One workspace per chart. Each account opens to an encounter summary (unit, admission and discharge dates, length of stay, service and readmission status), clinical documentation pulled directly from Meditech, and a built-in ICD-10 code finder. Notes can be read in a Meditech-style layout or as transcribed text.
Suggestions with evidence attached. For every suggested diagnosis, coders see the proposed code, the diagnosis type, a confidence level, and the supporting documentation that explains why it was suggested. In one example, the copilot proposed spastic cerebral palsy (G80.1) as the admitting diagnosis with high confidence, backed by the provider's problem list and the discharge summary. The coder accepted it.
Coders stay in command. Each suggestion can be accepted, replaced, rejected, or supplemented with a missing code. In a second example, the copilot suggested an unspecified fall with low confidence and no supporting evidence. Using the built-in search, the coder quickly found that "fall" referred to a patient at risk of falling, not an actual fall during the encounter, and rejected the code. Every decision captures a reason, how far off the suggestion was, and the exact line of documentation behind it, which feeds the model's learning.
Those decisions build the abstract, including diagnosis types, prefixes, and clusters, with pre-submission checks completed along the way. For now, coders still key the final abstract into Meditech and follow the usual submission routine, with more integrated submission features on the roadmap. As Nikki put it, the copilot and the coder's expertise work together to produce a more accurate abstract that reflects each patient's complexity.
Sri closed this segment with an invitation. Holland Bloorview built the solution as a module other hospitals can plug into their own platforms, starting with Meditech sites across Ontario and Canada, and the foundational workflow applies well beyond Canada wherever medical coding matters.
Technology spotlight: trust at the center
Joe Warbington of Snowflake reflected on what the demo represents: technology that steps out of the way and gives coders a single pane of glass instead of a swivel chair across systems. Snowflake supports more than 1,500 healthcare and life sciences organizations worldwide, from health systems and pediatric institutions to insurers, pharma and medical device companies.
For Joe, trust is the foundation. In healthcare, an AI that gives a wrong answer will not be used, so data must be unified, governed, and compliant with regulations such as HIPAA and HITRUST certification. He highlighted common provider use cases, from patient 360 to the medical coding agent shown in the session, and connected the copilot to the quintuple aim: better outcomes, broader population coverage, lower cost, and a better experience for patients and for staff. He also welcomed Holland Bloorview's offer to share its work and noted that Snowflake will help bring the approach to other EHRs.
The Keyrus perspective: context before model
Kyriakos Bilias, who worked alongside the Holland Bloorview team to build the application, shared what the project taught Keyrus about operationalizing coding agents. It reflects how Keyrus approaches every engagement through the Human Orchestrated Modelâ„¢: intelligence foundations first, humans in command of decisions, and performance measured continuously.
Privacy comes first. Whatever the jurisdiction (in Ontario, PHIPA), the questions are the same: who sees the chart, where does it go, and can you show who did what? Here, the chart, the suggestions, and coders' decisions all stay inside the hospital's own Snowflake account, where the model also runs. The pipeline strips identifiers so the model sees only clinical text, while coders see the full chart, supported by role-based access, sign-on, and logging.
Context before model. The instinct is often to reach for the strongest model. In practice, even a strong model performs poorly with the wrong context. Context engineering defines which documents the agent sees, in what form, and the facility rules coders know by heart. Because every document consumes tokens, context also drives cost. A good rule of thumb: get the context right on a mid-sized model, and reach for a larger one only when the evidence is in place and answers still miss.
A workflow coders trust. Every suggestion shows the exact sentence it came from, in the document coders already know. Small frictions such as hidden evidence or extra clicks erode trust quickly, so the workflow was designed to remove them.
Kyriakos also explained how coder feedback becomes compounded intelligence. Every accept, replace, or reject decision is captured with a reason, the source sentence, a name, and a timestamp. Those decisions are stored and indexed so the next similar chart shows what coders chose before. Recurring corrections surface as candidate rules that a person approves before they reach the model, reviewed on a weekly or monthly rhythm. Each code is checked against the full ICD-10 classification so invented codes cannot pass. As he put it, trust comes from the evidence, not the size of the model: confidence is a signal, never a verdict, and the coder decides.
His three recommendations for getting started: keep the data behind your own sign-on, show the evidence in the documents users know, and measure from day one.
Q&A highlights
How do you know the tool has not missed codes? Nikki noted that coders still review discharge summaries for now. In practice, the tool tends to over-code rather than miss, picking up every mention of a condition or symptom, so the team is teaching it that context determines whether a mention was clinically significant to the episode of care.
How quickly is coder feedback incorporated? Sri explained that two models sit behind the copilot: a machine learning model and a retrieval-augmented generation (RAG) model. At month-end, the health information management team reviews the feedback and decides which corrections are deployed, separating one-off errors from fundamental patterns.
How does the AI assign diagnosis typing? Early scoping sessions captured the facility's coding rules. Deterministic, rule-based automation now populates about half of each abstract, and the machine learning model handles the rest from the documentation, including type 5 admitting diagnoses, which the hospital codes to capture the clinical reason for admission.
What about abbreviations with more than one meaning? The team supplied a list of common clinical abbreviations, such as DAI for diffuse axonal injury in brain injury rehab. Ambiguous cases are resolved by reading each abbreviation in the context of all source documents for that encounter.
Where should an organization start? Kyriakos recommended detailed discovery sessions with end users to confirm the process can be recreated from defined steps and available information. Design follows naturally from there. Context engineering is where the real effort lies.
How is success measured? Sri described a full AI governance lifecycle at Holland Bloorview. Both the machine learning and RAG outputs are tested against gold-standard data, monitored against thresholds that guide when to pause or continue, and reported transparently to leadership, alongside ongoing work to build AI literacy across every deployment.
Ready to bring AI-assisted coding to your organization?
Whether you are a hospital looking to give coders back time for expert judgment, or any healthcare organization building the foundations for agentic AI, Keyrus can help you move from idea to operational reality. Keyrus and Snowflake are offering two funded workshops covering how to advise, implement, and operate a coding agent like this one. Share your use case or idea with us by November 16 to be considered.
Contact us to get started, or explore our healthcare industry work.