Mark Hay, TextQL’s CTO and co-founder, joins Angelina Yang on TwoSetAI for a 37-minute conversation about AI data analysis, shared definitions, and changing data-team responsibilities.

Start with a question, then improve the plumbing

  • Hay recommends trying a model with a database schema before building a company-wide data pipeline, provided security is enforced from day one.
  • Experienced analysts can inspect early results; executives need more support before relying on answers.
  • Existing tools for moving and transforming data can stay. Add connections and clean up the parts that actual questions expose.
  • The product’s added value includes governed connections, enough computing capacity, a sandbox for analysis, and an interface business teams can use.

Explore first, save verified definitions

  • The demo asks TextQL’s Anna to research the host online and build a dashboard using seven connected data sources.
  • The host questions the advertising budget and possible confusion with another channel. A polished dashboard still needs factual review.
  • Hay describes an ontology: saved definitions of business entities, metrics, formulas, and relationships that later analyses can reuse.
  • Validate an analysis before saving those definitions. Reusing a mistaken definition can make an error consistent.
  • Teams can share definitions while restricting access and keeping different business units’ meanings separate.

Test answers and control the workload

  • Hay says enterprise pilots use customers’ questions and expected answers, then expand access after meeting an agreed accuracy bar.
  • The interview supplies no numerical accuracy result, test-set size, or independent benchmark.
  • More analysis agents can explore more questions, but source databases, query bills, and human attention still limit useful throughput.
  • His “next trillion-token use case” framing is a prediction about demand for analysis, not a measured deployment.

Data work shifts toward maintaining the system

  • Hay reports about 40 employees and 15–16 engineers at TextQL, with analytics shared across functions and no dedicated data team.
  • His example starts with an engineer documenting a new feature’s queries, then sales staff investigating adoption and potential beta customers.
  • He expects data specialists to spend more time maintaining reliable environments, definitions, and tests for AI-assisted analysis.
  • That is a founder’s account and forecast; the interview does not establish that large enterprises can remove their data teams.

Hire and learn for judgment

  • TextQL allows AI on take-homes and asks candidates to improve a small product feature beyond a model-only baseline.
  • Interviewers still check code fluency, algorithms, systems design, performance constraints, and visualization judgment.
  • Hay recommends separate practice sessions: push AI tools hard, then study and write code without them to build independent understanding.

“Your data is a disaster, but we can work with it.” — Mark Hay