Dev48
Language
  • About
  • Services
  • Industries
  • Technologies
  • Articles
  • Contacts
Book a call
    Home/Articles/Healthcare ai starts with data you can trust
Dev48
info@dev48.org+375 29 320-40-96WhatsApp@isky_by

© 2026 · All rights reserved.

Healthcare AI starts with data you can trust

Source: dbt Labs•November 16, 2026 at 02:31 PM

Healthcare AI starts with data you can trust

Define metrics once, give agents context, and keep pipelines fresh. See how healthcare teams build AI their care teams can trust.

Ask four tools for your referral leakage rate, and you might get four separate answers: 22% from the business intelligence (BI) dashboard, 15% from an embedded app, 18% from a spreadsheet, and 27% from an AI chat. That leaves everyone scrambling to figure out which answer’s right.

In healthcare, these numbers answer questions about patient experiences and outcomes. The care teams acting on them need to trust them.

Self-service analytics has been an exciting possibility for years, but there’s still work to be done to bring it to fruition. AI-driven solutions can finally make it a true capability. But trust is still one of the biggest challenges, and it matters even more for the data you deliver to AI.

That's why we focus on reliable data, consistent definitions, and end-to-end lineage. In this article, we'll cover why trust is the bottleneck for healthcare AI, how to build a trusted data foundation, and how Astrodata's AXIOM application uses it to take a medical director from population trends to a patient's care pathway.

Trust is the bottleneck for healthcare AI

JB Brooks, principal data and AI architect at Astrodata, has seen the problem from both sides: as a healthcare practitioner and, on the data side, as someone trying to keep up with the business.

One of the main challenges has always been reconciling disparate data spread across electronic health records (EHRs), electronic medical records (EMRs), hospital information systems, and internal Google Sheets. These silos make it hard to combine business context and get fresh data to make decisions.

Governance and procedures around data sharing exist for good reason, but they further slow down healthcare. According to a 2024 study by Arcadia and the Healthcare Information and Management Systems Society (HIMSS), 47% of healthcare data on average is underutilized in clinical and business decisions.

Not long ago, Brooks was faxing paper records across the office. He says the pre-2010 era is "somewhat pervasive still today."

Digitization and connectivity have improved since then. The next stage is agentic AI that surfaces real-time recommendations, and that requires unified, governed data.

The bottleneck is trust. When data is spread across disparate systems and transformations are siloed or poorly tested, the numbers look off. People lose faith in the data, and work grinds to a halt.

But Brooks says a system delivering highly trustworthy data and business context to AI is possible with tools like Fivetran and dbt, when used with strong implementation patterns that earn stakeholder trust.

Four pillars of trusted data

  • One number, everywhere. If AI is querying a metric that's also used in a dashboard to make decisions, everyone needs to be talking about the same number. Otherwise, data isn't helping you arrive at a consistent conclusion or drive alignment.
  • Governance. You and your agents need to know how a number is calculated, who defines it, and which team controls it. That ties back to how the data team version controls and evolves its modeling.
  • Freshness. If a number isn't fresh, it has drifted. It's no longer accurate, even if it once was. We’ve put a lot of effort into making the Fivetran platform reliable, so you don't end up with gaps where data becomes stale.
  • Secure by design. Features like column blocking and column hashing give you control of data as it flows through. You can compare data without exposing the data itself.

An end-to-end architecture for healthcare data

For an organization that wants to combine EHR, financial data, and patient experience surveys, it might aim to optimize staffing, study whether data points from visits correlate with survey outcomes, or optimize finances based on what happens in those visits.

Fivetran and dbt sit in the middle. All connections are encrypted. Fivetran encrypts data in motion and at rest. With Fivetran's customer-managed keys, you manage encryption and maintain control and visibility. Fivetran can also work with external key vaults for security teams that want to manage keys themselves.

Fivetran ingests the data into a data warehouse. Then dbt transforms and models it. With Fivetran Activations, you can write the modeled data back out to source systems, so it doesn't just go to dashboards or agentic experiences. Practitioners get direct access to their data in the tools they use every day.

Protect sensitive data in the modeling step

Data from a system like an EHR is very raw. It's probably more than most people need to answer their questions, and most of those questions don't need identifiable data.

During the modeling step, you can anonymize and pre-aggregate data, then expose it downstream without risking sensitive information. Some anonymization can also happen during ingestion, as a pre-transformation.

Transformation also lets the data team govern how data is first exposed: a gold layer, or data mart, that's easy to query, with packaged datasets that fit what your users need.

Define a metric once so the dashboard and the agent agree

Analytics used to mean ingestion, transformation, and dashboards. Now, agents are accessing data. People across the organization can directly access data via SQL and semantic layer interfaces.

Agent chat is powerful, but it can mislead people and cause misalignment. The solution is governed metrics defined in a dbt project. When an agent uses a governed metric, you can tie its answer back to that specific definition. Instead of debating the number, you look at the definition and confirm it's correct.

Take referral leakage rate: out-of-network referrals divided by all referrals, by specialty and month. You define it once in code, on trusted EHR referral, payer claims, and provider data. Every tool returns the same 19%, including an AI agent connected through the dbt Model Context Protocol (MCP) server. Changes go through a pull request. Lineage shows exactly which data the metric depends on.

Spencer Taylor, co-founder of Astrodata, calls consistent, governed answers table stakes for any embedded or internal analytics solution. With dbt and the dbt Semantic Layer, end users get the same answer "regardless of the exact nuance of how they asked the question or what the LLM decided to generate in terms of a SQL query."

"We're removing a lot of that squishy, non-deterministic methodology," Taylor says, "and replacing it with deterministic, codified logic that represents business goals, represents business logic."

Context tells your agents why a number moved

AI agents don't really have a memory or learn between interactions. You have to give them the knowledge and skills to understand the organization: which metric to query, how to query it correctly, and how it fits into the whole.

Fivetran Context Layer is designed to take a question like, "Why is cardiology referral leakage up six points?" and place it in context. It draws on how parts of the organization work together and where the data points come from to explain why the metric might have moved and what to do about it.

That context spans metrics; relationships between providers, sites, and plans; documents like contracts and prior authorization rules; events like provider departures; lineage; and access rules for protected health information (PHI).

By default, Fivetran Context Layer lands this context in your warehouse. You own the contextual and unstructured data about your organization the same way you own the structured tables you use for analytics.

Even without an MCP server, your agents can query it directly. Fivetran's Agent Context MCP is designed to query it more efficiently, applying indexing so agents get quick, responsive access to the complete picture.

From population trends to a patient's care pathway

Astrodata, a roughly 30-person team of senior to principal data practitioners, built AXIOM on this foundation. It's a high-risk-patient command center for value-based care teams that blends analytical and operational workflows, two work streams historically owned by different people and systems.

AXIOM combines a number of elements in its architecture:

  • A React front end
  • Embedded Omni dashboards
  • dbt Semantic Layer
  • Fivetran Context Layer
  • The Tuva Project, a set of prepackaged dbt marts for clinical and claims data

The AXIOM Assistant is a custom agent that talks to all of these sources and queries the internet and connected repos like Google Drive and GitHub. It also takes actions, like navigation and kicking off workflows.

Picture a medical director trying to rein in rising costs from elevated readmissions. With a context layer built in Fivetran and dbt, an agent can point to the cohort whose costs are climbing fastest, then recommend care pathway steps that lower a patient’s projected readmission risk.

See how this system runs end to end in Beyond the Dashboard: Building Healthcare AI You Can Trust.

Evals and guardrails keep AI answers in bounds

AI is a probabilistic solution. Astrodata evaluates and grades the output, so any upgrades or changes keep to the same high standard. "It's really nice that we can actually measure that," Brooks says.

Healthcare is a sensitive ecosystem, so the output needs guardrails. Guardrails on large language model (LLM) calls ensure the data surfaced stays within the solution's governance and privacy requirements.

You don't need to wait years to get data ready for a solution like this. Start with a proof of concept, or a lean minimum viable product (MVP) on your first governed metrics, to prove the possibility to your health system.

From there, you can productionize and scale it, bringing the long-promised dream of self-service data in highly sensitive environments into reality.

To see AXIOM go from a population heatmap to a patient's care pathway, watch the full webinar on demand. To see how Fivetran and dbt fit your own healthcare data, talk to an expert about a one-on-one dbt platform demo.

← All articles

What this article says

More in Data & Analytics

All →
AI-powered lineage for Python scripts
Collibra

AI-powered lineage for Python scripts

Gain control of your MCP ecosystem with AI Command Center
Collibra

Gain control of your MCP ecosystem with AI Command Center

JSONBench is back, and ClickHouse still leads
ClickHouse

JSONBench is back, and ClickHouse still leads

How ISVs are unlocking European workloads on Google Cloud Dedicated
Looker

How ISVs are unlocking European workloads on Google Cloud Dedicated

Inside the new Data Boundary experience
Looker

Inside the new Data Boundary experience

Every flag needs evidence: how ReviewerZero AI traces integrity issues back to the page
Datalab

Every flag needs evidence: how ReviewerZero AI traces integrity issues back to the page

More from dbt Labs

How to assess your organization’s AI readiness
dbt Labs

How to assess your organization’s AI readiness

Making the economic case for warehouse-first AI context
dbt Labs

Making the economic case for warehouse-first AI context

Fivetran + dbt Labs Complete Merger to Create the Data Infrastructure for Trusted AI Agents
dbt Labs

Fivetran + dbt Labs Complete Merger to Create the Data Infrastructure for Trusted AI Agents

Everything we announced at dbt Summit and why it matters
dbt Labs

Everything we announced at dbt Summit and why it matters