Sharan Senthilvasan is director, Data Platform Engineering at ModMed.
As the director of data platforms for a $5.6 billion SaaS company selling EHR solutions to doctors and patients, I’ve seen firsthand how important—and how challenging—it is to build a data foundation for AI-enabled workflows.
Like most companies, we have data spread across multiple systems. A data platform brings those sources together and transforms the information for reporting and other uses. But just putting data in one place doesn’t make it trustworthy. The platform needs rules that ensure the same question produces the same answer every time, regardless of who asks it or how the original data was structured.
Unify Meaning First
The first instinct companies often have is to pull data from different systems and put it all in one place: a data lake. That gives you unified storage, but you still need shared definitions, documented calculations, ownership, access rules and consistent ways to transform the data.
One billing system might store a claim as a single record, while another breaks it into service lines, so one table has one record per claim and another can have 10. Joining the tables without accounting for that difference can give incorrect financial numbers even if every record is correct.
A trustworthy platform defines how tables relate and makes those rules reusable. Technical teams need consistent design patterns, while nontechnical users should be able to easily understand a table through a catalog or interface.
The platform must also protect the data, and when information is subject to HIPAA, those requirements need to be built into the architecture, whether it’s a person or an AI model using it. A platform can provide one environment containing protected information and another in which identifying details are masked. But changing a patient’s name is not enough if claims, insurance records and diagnoses aren’t connected correctly.
Start Small
Before extending a platform across an organization, it’s critical to validate the architecture with one complete use case. At ModMed, we started with financial data, bringing information from several systems into shared storage and then building rules on top of it. Once that pattern was established, we could apply the same approach to clinical data. Even if the rules differ for different kinds of data, the underlying design patterns can still be reused.
The first use case should reconcile with its source systems and have a named owner. If a reported revenue figure does not match the source, the team needs to understand why before other teams build on top of it. Data structures and locations can change; revenue might be stored in one table today and another tomorrow, or it may move between systems. If the rules don’t keep up with those changes, the revenue number everyone’s used to can end up meaning something very different.
AI Needs Clear Context
Even when data is available, AI does not automatically know what it means in an organization. The data layer provides information, while a context layer gives an AI workflow guidance on how to use it.
Data harmonization is the middle step, after data comes out of the source systems and before anyone uses it. We clean and align data for reporting, AI consumption or a nontechnical user’s questions. But what needs cleaning depends on the purpose, and there is no single process that fits every use. Preparing data for a report may require different choices from preparing it for AI. For AI context, in reality, you don’t have to do a lot.
If billing and reporting show different balances for the same account, someone has to determine whether the difference comes from timing or different definitions. AI might make a disagreement sound settled, so it shouldn’t decide which number is right when the organization hasn’t resolved the question. The workflow needs explicit rules for deciding which information applies and when uncertainty should be escalated. Otherwise, you might get an AI answer that sounds confident even though a person still has to make the call.
Designers who build the data platform need to understand what the end user and the AI need. A data platform serves both people and AI tools, but they don’t use information in the same way. People need consistent definitions and results they can trust. AI also needs a context layer that explains how to interpret data and what to do when the answer is unclear. So measure what they need first.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?