For the last decade, building anything meaningful on a platform like Snowflake required a specialist. If a quality director wanted a HEDIS gap report, the request went into a queue. A data engineer wrote the SQL. A report writer formatted the output. A BI analyst validated it. The work was slow, but it passed through hands that understood both the data and its consequences.
That model is ending faster than most health plans planned for, and it is not the doing of any single vendor. Claude has Claude Code, Epic has Agent Factory, Salesforce has Agentforce Vibes, and the list goes on. While this isn't breaking news, it's certainly becoming the new normal. Across the platform landscape, the major players have made the same bet: that the next wave of value comes from the operational users sitting closest to the work.
With its Cortex Code and Snowflake Intelligence releases, Snowflake has extended build functions beyond data experts to domain experts: the people who know what a RAF score means but have never written a line of SQL. A care management lead can now ask a governed AI agent a question in plain English and get an answer grounded in the plan's own data, while a quality manager assembles an agent that reasons over claims, labs, and notes so each member's risks and open gaps are accounted for.
The change reaches further than the obvious technical roles. It is not only software developers and report writers who will feel it. Business analysts who translate requests for the data team, quality-assurance testers who validated the output, and anyone whose job assumed a technical handoff may find that handoff collapsing into a single conversation with a tool. Some of that work disappears. Much of it changes shape into reviewing, governing, and validating what others build.
For most health plans, the appeal is obvious. When everybody becomes a builder, the analytics backlog that has throttled quality, payment integrity, and care management for years could dissolve. A department manager who has waited weeks for a number that shaped a decision due tomorrow can suddenly get that answer without filing a ticket.
It is also where the hard questions begin. The specialist queue was slow, but it was also a quality gate. Removing the gate without replacing the function it served does not eliminate risk. It distributes it across every department in the building.
This is bigger than any one vendor
It would be a mistake to read this as a story about any individual vendor. The spread of AI tools is a structural shift, and the scale of it is still being absorbed. Three things are driving it:
- Natural-language interfaces: ask a question or describe an agent in plain English, and the tool does the rest. No SQL, no pipeline, no ticket.
- Connectors that wire AI directly into source systems: increasingly through the open Model Context Protocol (MCP), which lets a tool reach into SharePoint, a claims platform, or an EHR feed using the end user's own permissions.
- Agents that non-technical users can assemble: reusable workflows that read from, and sometimes write to, the systems the business actually runs on.
The strategic implication is the part most plans have not fully absorbed: the front door to your data is no longer a single warehouse guarded by a single IT gatekeeper. It is a dozen tools, each able to reach into source systems on behalf of whichever operational user is holding the keyboard. Governing one platform was hard enough. Governing the pattern across vendors, with controls that do not match each other, is the real challenge.
The risks that don't show up in the demo
The product demonstrations show what the technology does. They are quieter about what must be true for it to be effective and safe at scale. Health plans should be working through the following risk areas as they map out their AI approach.
Accuracy: confident answers, wrong numbers
Health plan data is full of nuance: overlapping eligibility spans, ICD-to-HCC mapping rules that change each plan year, and fields across systems that represent similar concepts but are not identical. Health plans have wrestled with data quality for years, so this shift does not create the problem so much as scale it. It is already common for a single organization to maintain several data lakes, and running the same algorithm against different sources can return different answers. An operational user who does not know these landmines can prompt a natural-language agent in Snowflake, ChatGPT, or Claude and get a clean-looking answer that is quietly, dangerously wrong. That answer may then travel to:
- Internal decision-makers: who set strategy on the assumption the data is sound.
- Members: in benefit, coverage, or care communications.
- CMS and other regulatory agencies: in submissions where accuracy is a compliance obligation.
A fluent, confident-sounding answer makes a wrong one harder to catch.
Fragmentation: ten departments, ten versions of the truth
Without governance and oversight, self-service does not dissolve silos. It can entrench them. When every team builds for itself, across several different tools, two departments will produce the same metric with different definitions and different answers, each solving the problem for its corner instead of for the enterprise. Multiply the silos by the number of platforms and builders in play and the divergence only widens. The result is not faster decisions but reconciliation meetings and eroded trust in the data.
Cost control: invisible consumption, now across many tools
Every agent, query, and AI seat consumes resources. On a data platform like Snowflake, that is warehouse compute. Gen2 warehouses run meaningfully faster on heavy workloads, but they bill at a premium of around 30 percent more per second. A workload only saves money if it finishes proportionally faster. On top of that sit per-seat and per-token charges for ChatGPT Enterprise, Claude, and Copilot. Spread across hundreds of new builders and several tools at once, spend compounds in places where no single owner is watching. The platforms increasingly provide budgets and usage analytics and admin consoles, but they only help if someone owns them across vendors, not one tool at a time.
Adoption and limits: the tool can't fix what the data never had
Two things get overlooked in the rollout. First, without operational buy-in, even the best capability becomes shelfware as people return to the spreadsheet they trust. Second, the data limitations you have now do not vanish. If your chart-chasing process cannot retrieve the record today, an agent asked to close that gap will hit the same wall, no matter how advanced the underlying model.
None of this is an argument against the technology. It is an argument for treating the shift as an operating-model change that spans your whole tool portfolio, not a software rollout.
How leaders can get ahead of this
The plans that win with self-service AI will not be the ones that adopt it fastest. They will be the ones that pair adoption with structure that holds regardless of which vendor's logo is on the screen. Four moves separate the two.
Stand up governance before you scale access
Governance is not a brake. It is what lets you press the accelerator with confidence. Done well, it also gives you the framework to measure results: the baselines, metric definitions, and tracking that let you prove ROI. Before a wave of new builders arrives, decide the rules of the road and write them to be tool-agnostic, applicable equally to Snowflake, ChatGPT, Claude, Copilot, or whatever you license next:
- Appropriate use: what AI is and is not fit for, and where a human decision is non-negotiable.
- Who can create an agent: clear roles for who builds, who reviews, and who publishes, with admin approval of new connectors so AI cannot quietly reach into source systems.
- Quality assurance: a validation step every tool must clear before its output reaches a member, regulator, or executive.
- Cost and consumption: owned budgets and monitoring across every platform, with a single named owner accountable for spend.
- Ethics and bias: explicit fairness checks for any model touching member care or coverage.
- Measurement: a process to monitor performance, accuracy, and drift on an ongoing basis.
The platforms supply real tools for this: row-level data isolation, spend budgets, and evaluation frameworks on the data side; role-based access control and connector approvals and audit logs on the AI-assistant side. Governance decides how they are used and lines them up across vendors.
Train operations like builders, not just users
A login is not a skill. The operational users you are enabling need two new competencies: enough technology literacy to choose the right tool for the right problem, and real fluency in prompt engineering — knowing how to frame a prompt so the output is accurate and auditable. This is teachable, and it pays for itself the first time it prevents a bad number from leaving the building.
Decide your IT and tooling model deliberately
Centralize or decentralize? The honest answer is usually both, on purpose. The deeper shift is cultural: AI has to be owned across the organization, not quarantined in IT. Many plans are embedding technical talent directly inside business units — people who know both the tool and the domain it serves. A parallel decision is how many tools you actually sanction. Unmanaged sprawl multiplies every risk above; a deliberate, governed tool portfolio beats a different favorite in every department.
It also means naming owners, not just policies. Cost control needs a single accountable person: a FinOps lead, the data-platform owner, or the AI governance body. The same goes for quality, access approvals, and measurement. Each needs a person, not just a policy. This is also the moment to repurpose your data specialists rather than cut them. The report writer who knows your data forward and backward is exactly the person you want validating what operations now builds.
Earn buy-in with one focused pilot
Pick one department where the appetite for AI is already high and the data is well understood. Risk adjustment is often the right place to start, given how data-driven that function already is. Let that team demonstrate measurable results under real governance, on a sanctioned tool, build the playbook, and bring the rest of the organization along. Operational buy-in spreads through proof, not mandates.
Where i2 Health Advisors comes in
Anyone can stand up Snowflake Cortex, switch on ChatGPT Enterprise connectors, or deploy a Claude agent. The harder problem is what determines whether self-service AI creates value or liability: knowing whether the answer coming out the other end is right for a regulated health plan, regardless of which tool produced it.
We understand the technology. We understand the industry too. We are vendor-agnostic by design: we pair fluency across the leading data and AI platforms (Snowflake, OpenAI, Anthropic, Microsoft, and others) with hands-on payer operations expertise across risk adjustment, HEDIS, utilization management, payment integrity, and care management. We build and operate the governance infrastructure, training, and workflows that let the wider organization scale safely across every tool in your portfolio, so speed never comes at the cost of accuracy.
Talk to us before your first agent goes live, not after. Contact i2 Health Advisors to start the conversation.
Frequently Asked Questions
What are the biggest risks when health plan staff start building their own AI-powered reports and workflows?
We see three primary risks when health plans expand AI access to their teams: output accuracy, context fragmentation, and tooling costs. AI tools produce confident-sounding answers even when the underlying data is wrong, and health plan data is layered enough that a staff member without deep data knowledge can generate a bad number that travels to members or regulators before anyone catches it. Fragmentation happens when departments build independently across multiple tools, ending up with different definitions for the same metric and no shared source of truth. Costs compound faster than most plans expect, as per-seat and per-token charges across platforms add up quickly when no single person owns spend at the portfolio level.
Can health plans implement self-service AI without replacing core admin systems like Facets or QNXT?
Yes, and that's exactly how most plans should approach it. Self-service AI tools connect to source systems through APIs, MCP connectors, and standard feeds, so a plan running Facets, QNXT, or TriZetto can extend access to operational users without touching the core.
How should a health plan structure AI governance to work across multiple vendors?
The most durable frameworks are tool-agnostic. At minimum, governance should define who can build and publish agents, what validation any tool must clear before its output reaches members or regulators, how spend is owned and monitored across platforms, and how fairness checks apply to any model touching care or coverage decisions. Each area needs a named owner, not just a policy document. Plans that skip this step early almost always rebuild it under pressure later.
