
How an AI-First Data Platform Enables Self-Service Analytics
A sales development manager with no data science background built a call-analysis model in a matter of weeks using AI. According to Gina Jenks, AI Knowledge and Analytics Manager at Synthesia, a project of similar complexity would once have taken a data scientist three to five months.
The example illustrates a broader shift in how organisations can use data, moving some analytical work beyond specialist teams and into the hands of employees across the business. But making that possible takes more than access to AI tools. It requires data foundations that allow people to find, understand and use information effectively.
Jenks joined Christina Stathopoulos, founder of Dare to Data, on Don’t Panic! It’s Just Data to discuss how Synthesia rebuilt its data platform to support this shift with Fivetran, and what it takes to make data more accessible and useful across an organisation.
Why a modern data platform mattered
Synthesia makes AI video for business, so data is part of the product. The company was growing fast and producing huge volumes of it. It was also changing who it hired. As Synthesia became a scale-up, more non-technical people joined, and they needed to use data without a builder’s skills. The starting point was modest with a simple reporting tool, plus a lot of analysis scattered across spreadsheets and the CRM. Jenks says the goal was to give people “all of their reporting available in one place” as a single source of truth.
A stack built for a small team
The data team was tiny, so it had no time to tend the pipes. It chose Fivetran for ingestion, which now pulls from about 50 systems. Fivetran handles schema changes automatically, moving roughly six terabytes of data and 500,000 daily changes into the warehouse. Transformation on dbt Labs came with documentation and tests built in. As a result, the team went with a self-service approach and knew the trade-offs. Anyone could create their own metrics, which meant metric sprawl and some ungoverned definitions. But adoption was fast, and for the first time people could answer basic questions, like how many videos were generated yesterday, from one place.
From gatekeepers to a platforms team
Before AI, the data team was part analyst, part platform builder, still making dashboards and answering questions by hand. Jenks says that has changed. The business context analysts once carried in their heads can now live in a context layer that helps the AI decide what data to show whom. “No more gatekeeper,” Jenks agreed when Stathopoulos put it that way. The team now spends its time improving the platform instead of fielding one-off requests.
What business users built with AI
Jenks says the most interesting results are the tools nobody on her team would have predicted:
- A book-of-business planner. A sales operations lead built a scenario tool to balance account assignments across reps. It lets her audit books at scale, and reps can check it themselves when they question their allocation.
- A coaching report. A sales development manager inherited a team already hitting 90 to 100 per cent of target, and his goal was to push it to 120 or 130 per cent. He built a red-amber-green report he can read in five minutes before Monday one-to-ones.
- A conversation analysis model. He then used call transcripts to compare last week’s conversations with those that became sales-qualified opportunities, to spot where reps stumble, such as objection handling.
Jenks sees two problems solved at once. Niche requests rarely make the priority list. And there is always a translation gap between a business user and an analyst who doesn’t know their daily job. “AI has sort of collapsed that interaction,” she says, so people can build exactly what they need.
Semantic layer, context and documentation
Getting AI to return reliable answers took real work. Jenks’s first fix was the semantic layer, because AI, she says, reads things “very literally.” The old setup was organised around entities such as opportunities and accounts. Users joined everything together and bolted on endless fields. Even humans found it a mess. The team switched to a use-case model, with sets such as open pipeline and bookings, built around who in the business looks at what.
Next came context, layered at three levels, namely: overall guidance, individual datasets and single fields. The team also needed far more documentation on its data models, a chore nobody wanted. They adopted an AI coding assistant that can read the whole codebase and work out what each table and field does, and used it to write descriptions at scale. Those descriptions flow through to the analytics layer.
Testing AI accuracy with evals
Early results weren’t always right, so the team set up evals, which are a set of prompts, such as “What is my open pipeline?”, each paired with an expected answer. An eval runner executes them, compares the output with the expectation and gives a pass or fail. Jenks says this is how you can trust the AI at scale, and it shows where the context needs adjusting.
Balancing freedom and control
Jenks admits it’s “quite unnerving” to stop controlling the flow of information. Several things help. Synthesia already had strong data literacy, and AI adoption pushed even sceptics to learn more. The team answers questions directly, and an AI assistant in a dedicated Slack channel can read the codebase and reply on the team’s behalf. An “AI celebrations” channel lets people show off what they’ve built and borrow ideas from each other.
Dashboards haven’t gone away either. Teams still rely on them for monthly and quarterly reporting, and they give people a known source of truth to check AI answers against. Jenks says they also show which metrics people care about most.
Advice for building an AI-first data platform
Her advice to organisations looking to implement this type of solution is all about foundations: fresh data, version control across everything, schemas that are easy to change, and thorough documentation. Investing in data foundations, she says, is “absolutely critical” if you want to self-serve data with AI. If you would like to learn more, visit fivetran.com and synthesia.io or connect with Gina Jenks on LinkedIn.
Takeaways
- Synthesia's early data platform and growth challenges.
- Transition to a modern, scalable data architecture.
- Role of Fivetran, Snowflake, and dbt Labs in data ingestion and transformation.
- Layering business context and documentation for AI effectiveness.
- User-driven tool creation and AI's role in democratising data access.
- Balancing control and freedom in self-service AI environments.
Chapters
00:00 Introduction to Synthesia's Data Journey
02:37 Early Data Platform and Growth Challenges
04:47 Building a Modern Data Platform with Fivetran and dbt Labs
06:29 Self-Service Analytics and User Empowerment
08:38 The Role of AI in Scaling Data Access
10:16 Innovative Tools Built by Business Users
16:35 Behind the Scenes: Data Infrastructure and AI Integration
21:21 Layering Context and Documentation for AI Effectiveness
22:21 Evaluating AI Results with Evals
23:16 Balancing Control and User Freedom in AI Data Use
26:40 Trust, Dashboards, and Data Governance
27:25 Advice for Building AI-First Data Platforms
Weitere Episoden von „Don't Panic! It's Just Data“



Verpasse keine Episode von “Don't Panic! It's Just Data” und abonniere ihn in der kostenlosen GetPodcast App.








