This review is part of a larger series of LinkedIn newsletters titled [The Human Side of AI: Cutting through the AI noise to show you how AI can be a powerful tool for your creativity, efficiency, and strategy].
PromptQL, from Hasura (the team behind the open-source GraphQL engine), is an AI data agent. Researchers point it at their databases and warehouses, ask questions in plain English, and get answers back as tables, charts, and short written findings. It is positioned as a “multiplayer” agent with shared threads and a shared team brain.
PromptQL works on structured data from databases and warehouses. It fits the quantitative corner of a research workflow and stays there without pretending otherwise.
Key strengths
Answers computed as code you can inspect
PromptQL uses the model to write a query plan, then runs that plan as real Python and SQL outside the model. For a researcher analyzing survey data, that matters. A crosstab or a segment breakdown gets computed by code you can read, so the arithmetic never depends on a model doing math in its head. Every answer carries a reliability score, and when a client challenges a number in a findings deck, the result traces back to the exact rows and the query that produced them. On the Berkeley Data Agent Benchmark, frontier models on their own tend to land in a low band on real data questions, which is the gap this plan-then-execute design targets.
Joining across fragmented, structured sources
Riding on Hasura’s data layer, PromptQL can join a Postgres table to a warehouse table to a SaaS API and reason across them in one thread. For win-loss or go-to-market work, a researcher can line up deal outcomes against firmographics without waiting days for a data team to hand-build the query. Native connectors cover PostgreSQL, Snowflake, BigQuery, Databricks, and most of the warehouses a client is likely to run.
A confidentiality posture built for NDA data
For research, data handling decides whether a tool is usable at all. PromptQL is designed for bring-your-own-cloud deployment, so data can stay in the firm’s own cloud account (AWS, GCP, or Azure) rather than a shared vendor database. Permissions are enforced at the data layer, which means the agent only ever sees what the person running it is cleared to see. The security posture is aimed at regulated and NDA-bound data, but confirm the specific certifications (SOC 2, ISO 27001, HIPAA) with the vendor before relying on them in regulated work.
Where it could improve
The hard limit is scope. PromptQL will not read a stack of interview transcripts for themes, tag patterns across IDIs, or turn a dozen win-loss conversations into a persona. That unstructured, qualitative work is the heart of most research engagements, and this is the wrong tool for it. NotebookLM, Claude, and ChatGPT remain the right choices there.
Setup is the other cost. Standing it up is a genuine data-engineering project: connectors, metadata modeling, a semantic layer, and command-line work. A solo researcher or a small team without a data engineer cannot self-serve. The security features a research firm would actually lean on, including bring-your-own-cloud, single sign-on, and audit trails, sit behind the quote-gated Enterprise tier. And the public footprint is still thin, with no G2 or Capterra reviews yet, so an evaluation rests on vendor material and a handful of early hands-on accounts.
Ratings
| Dimension | Rating | Rationale |
|---|---|---|
| Usability | 2.5 / 5 | Readable query plans are approachable once the system is live, but standing it up is a data-engineering project a researcher cannot do alone. |
| Power | 4.5 / 5 | On structured, quantitative work it is strong: deterministic and auditable down to the query that produced each number. |
| Flexibility | 2.5 / 5 | Wide connector coverage but a narrow task range for research. Strong on quantitative work, absent on qualitative. |
| Cost | 3.5 / 5 | Consumption pricing is fair and small per query, with generous free credits, but the real cost is setup and the gated Enterprise tier. |
Conclusion
PromptQL earns a place in a research stack on one job: reproducible analysis of structured, cross-system data that has to survive client scrutiny. On that job it is good, and its willingness to show its work gives it an edge over a general model answering the same question blind. Most research is qualitative, though, and there the researcher’s judgment and a different toolset still carry the load. The future of this work is Human + AI as the operating model, with the person directing the analysis and the tool handling the parts it can prove.
Last updated: 8/5/2026.