As AI accelerated product development across Zocdoc, we saw an opportunity to raise the standard for trusted analytics workflows. Engineers could prototype and ship faster, product teams could test more ideas, and Data Science still had to support the measurement, interpretation, and decision-making that followed.
Better analytics helps teams make faster, more confident decisions about the experiences patients and providers rely on. Our goal was to improve the way Data Scientists worked without lowering the bar for trust.
AI-assisted analytics must be designed around privacy, data protection, and human accountability, particularly in healthcare. The work described here focused on improving analytical workflows while keeping those key principles embedded throughout.
Over the course of the year, we focused on four areas:
- Helping Data Scientists write better query drafts through trusted business context
- Streamlining repetitive workflow steps around analytical work
- Using AI to help Data Scientists investigate complex questions more systematically
- Building training and feedback loops so new ways of working could spread across the team
Across all four areas, the division of responsibility stayed clear: AI could assist with query drafts, organization, and hypothesis generation; Data Scientists owned framing, validation, interpretation, and recommendations.
1. Faster querying through trusted business context
One of the first places we focused was query writing.
Modern AI tools can generate Structured Query Language (SQL) quickly. But syntactically correct SQL is not necessarily useful SQL. Generic models do not automatically understand Zocdoc’s business definitions, analytical conventions, or the context behind a question.
So we made more of that context reusable.
We codified definitions and analytical conventions that experienced Zocdoc Data Scientists already relied on and made them available inside AI-assisted workflows. We also continued modernizing our semantic layer: the shared layer that standardizes business concepts, metrics, and relationships so people and tools can interpret them consistently.
Instead of asking a model to infer Zocdoc’s business logic from scratch, it could begin from a more reliable foundation. That improved the quality of first-pass query drafts and gave Data Scientists a better starting point for review.
It did not make every query correct, and validation remained essential. But it made AI more useful in the flow of real analytical work.
2. Reusable workflows simplified recurring work
A meaningful portion of a Data Scientist’s job involves turning loosely defined requests into structured work and moving that work through a repeatable process.
We built reusable AI-assisted workflows to help with those recurring steps.
An informal request could become a structured work item with the relevant context, decision to be informed, open questions, and expected output. Data Scientists could iterate with AI to sharpen the request, surface missing assumptions, and shape the analytical approach before beginning the work.
Additional workflows supported project setup, review, measurement planning, quality checks, and final readouts. The goal was not to automate the analysis itself, but to make the surrounding process more consistent and easier to navigate.
Individually, these steps are small. Together, they make the path from an initial question to a reviewed output more consistent. Final quality review and sign-off remain with a Data Scientist.
3. AI helped us investigate more, more quickly
We also explored where AI could help Data Scientists work through complex analytical questions.
Many investigations begin with a broad question and a large amount of evidence. A Data Scientist may need to examine many possible explanations before deciding which ones deserve deeper attention.
We built AI-assisted investigation tools to help with that first pass. They could scan a broader set of results, organize the evidence, and narrow a large hypothesis space into a smaller set of possibilities for a Data Scientist to investigate.
For recurring analytical workflows, we paired AI synthesis with deterministic systems. Deterministic logic handled structured calculations and checks; AI helped summarize the outputs and connect them with relevant operational context.
That combination proved important. The model could help decide where to look, but the underlying calculations still needed to be reproducible, and the Data Scientist still owned the interpretation.
The practical lesson was simple: AI was most useful when it expanded the amount of information a Data Scientist could consider, while deterministic checks and human judgment anchored the work.
4. Training and feedback loops helped the ways of working spread
Useful workflows alone were not enough. The team also needed a way to learn how to use them well.
We ran monthly hands-on sessions focused on AI-assisted coding and reusable analytical workflows. We established AI Champions across the Data Science team to demonstrate useful patterns, share examples, and help teammates adopt new workflows.
Domain-specific sessions, recurring updates, and onboarding materials helped turn individual discoveries into shared team knowledge.
One of the clearest lessons was that AI adoption could not be treated as a one-time rollout. The tools changed quickly, and the team’s own approaches evolved. Ongoing training and peer feedback became part of how the team worked.
More room for judgment, unchanged accountability
Across these efforts, we saw a consistent pattern: trusted context improved the quality of AI-assisted query drafts, reusable workflows simplified repetitive process steps, investigation tools helped Data Scientists narrow complex questions more quickly, and training helped those ways of working spread.
These approaches helped reduce some repetitive work and gave Data Scientists more room for judgment-heavy analysis. They remained responsible for framing the problem, selecting the method, validating the evidence, interpreting the result, and making recommendations.
That was the central lesson from the year: scaling Data Science with AI meant building better systems around Data Scientists so they could focus more of their attention on the decisions and analysis that required them most.