PromptQL review (2026): Hasura's data agent, tested

What PromptQL actually is (and the pivot nobody announced)

I build AI agents for a living, so PromptQL landed on my desk the way most new agent products do: with a big accuracy claim and a demo that looks too clean. The useful way to review it is to separate what the product is from how it is currently marketed, because those two things drifted apart over the last year.

PromptQL comes from Hasura, the company that shipped the open-source GraphQL engine that passed 100M+ downloads in two years. Founders Tanmai Gopal and Rajoshi Ghosh raised a $25M Series B and a $100M Series C, and by late 2025 VentureBeat was describing the company as a billion-dollar-plus business selling $900/hour AI engineers. The framing Tanmai uses is that the customer for your data changed:

"We realized that apps are no longer the most important things that need access to data. The future is AI talking to data, AI accessing data on the user's behalf."

Tanmai Gopal, VentureBeat interview

So the 2025 pitch was clear: a reliable, natural-language layer over enterprise data, positioned against text-to-SQL, tool-calling, and RAG. Then the homepage changed. Today promptql.io leads with "a multiplayer AI agent, think Claude or ChatGPT, but with shared threads and a shared brain," and the headline problem is context decay across Slack, docs, tickets, CRM, and warehouse tables. The named logos are heavy hitters: Cisco, McDonald's, Instacart, Swiggy, and Lightspeed.

Both descriptions are true, and they point at the same engine. But if you land on the site cold, you would think PromptQL is a team-knowledge tool a bit like Glean or a shared internal knowledge base. Under the hood it is still a data agent. Keep that in mind, because the pricing, the docs, and the benchmarks are all about the data engine, not the wiki wrapper.

How PromptQL works: plan first, then execute

This is the part I actually respect, and it is the reason the product is worth a review rather than a shrug. Most "AI on your data" tools hand the whole job to the model: read the data, do the math, write the answer, all in one context window. PromptQL splits that in two.

When you ask a question in the Playground, PromptQL first returns a query plan: a step-by-step, plain-language breakdown of how it will pull, filter, and process your data, including the assumptions it is making. You can read it, and you can edit it with a pencil icon, at which point PromptQL "remembers and self-corrects for future threads." That transparency is the bit skimmers should not skip: you see the plan before any data moves.

The plan then executes as actual code. PromptQL fetches your data and organizes it with Python, and you can expand any step to see the real Python and SQL it ran against your source. This is the line the whole product hangs on, in the founder's words:

"Instead of generating answers, we generate plans in a domain-specific language unique to your business. Those plans compile to deterministic actions with runtime validations and policy checks."

Tanmai Gopal, VentureBeat interview

Results land as artifacts: structured tables, text, or charts that live outside the model's context and get referenced by later steps. That is the clever bit for scale, because it means PromptQL can work over data far larger than a context window without the model quietly losing track of it. Semantic work (classify, summarize, extract, visualize) is handled by narrow AI primitives that each run in isolation with a focused context, so the LLM only ever reasons over a small, well-scoped slice.

Finally, every answer gets a reliability score based on query complexity, data accuracy, and model confidence, with evaluation notes on what the analysis did and did not include. If the score is low, you correct it, and the correction sticks for next time. Coming from the support world, this pattern is familiar: it is the same instinct behind preventing hallucinations, which is to say the system should tell you when it is unsure instead of guessing with a straight face.

The reliability problem PromptQL is actually solving

To see why any of this matters, look at how badly raw models do on real data questions. Tanmai cites the Berkeley Data Agent Benchmark, where frontier models score far below what a demo would suggest.

On real-world data questions, the Berkeley benchmark puts Opus 4.6 at 43%, Gemini 3 Pro at 38%, and GPT-5.2 at 25%, while vanilla database query generation scores "close to 0%." The benchmark's own read is that agents "typically select the right data, but fail at planning the computation or implementing it correctly." That is exactly the gap PromptQL's plan-and-execute split targets, and it is why the founder keeps hammering one word:

"The biggest problem is not being able to just accurately answer the question. The big problem is AI pretends to be accurate even when it's not. So AI is confidently wrong. That's the problem."

Tanmai Gopal, VentureBeat interview

Hasura groups the failure into four modes, and this 2x2 is the most useful thing to internalise before you evaluate any AI data analytics tool, PromptQL or otherwise.

Now, about that "100% accuracy" headline. Hasura is refreshingly honest that it is bait: the company concedes a "100% accuracy claim doesn't technically hold water," and the real claim is 100% repeatability, because the computation runs in code, not through model reasoning. On the narrower CRMArena-Pro database and numerical tasks, they report 100% versus a cited 30 to 60% for state-of-the-art approaches. Treat the round number as a conversation starter and the repeatability as the actual product feature.

The features that stand out

Beyond the core loop, a few things separate PromptQL from a generic AI knowledge base chatbot.

Automations that become real APIs. Once a workflow works in the Playground, you deploy it as an HTTP endpoint callable from Zapier, Slack, cron jobs, or your own apps, with no coding. That turns a one-off thread into a repeatable business process, and it is the feature I would lean on hardest if I ran a data team.

Cross-system joins through metadata. PromptQL is built on Hasura DDN, so a supergraph stitches multiple subgraphs together, and relationships "do not need to be backed by the same data source." In practice that means it can join Postgres to a SaaS API and reason across both, which is where tools built only for Snowflake or one warehouse fall down.

A self-building context layer. The 2026 wrapper adds a wiki that seeds itself from Slack, Google Docs, Snowflake, PostHog, and Salesforce, keeps revision history and audit trails, and enforces scoped access (external, internal, personal, confidential). It is the kind of connected-context play you also see in Salesforce AI tools. Corrections you make become shared "skills" the next person inherits, which is the same knowledge management problem every team fights, dressed as an agent.

Bring your own model, and it matters for the bill. PromptQL runs any LLM, from Claude and GPT to open-weight models like DeepSeek, Kimi K2, and GLM served via Fireworks. You can switch models per thread or mid-thread, and it interoperates over MCP. Hold that thought, because model choice is the single biggest lever on cost.

PromptQL pricing: what you'll actually pay

PromptQL has no seat pricing. It is 100% consumption-based, billed in a normalized unit called an OLU (Operational Language Unit), which rolls input, output, cached, and cache-creation tokens across every model into one consistent number so your bill does not change shape when you switch models.

The typical task costs Hasura quotes (at the $0.14 rate) are honest and small: under 2 OLUs for a simple data task (under $0.28), around 10 OLUs for a complex report ($1.40), and around 40 OLUs for a deep investigation ($5.60).

Here is the catch, and it is the most important thing in this whole PromptQL review from a budget angle. The per-OLU price is fixed, but the model you choose changes how many OLUs a task burns, and the spread is enormous. Against a Claude Opus 4.6 baseline of 1.0x, GPT-5.2 runs 0.42x, Kimi K2.7 runs 0.17x, and DeepSeek V4 Flash runs as low as 0.018x, which Hasura frames as up to 57x cheaper for the same work. So the same report can cost you a few dollars or a few cents depending entirely on the model dropdown.

Rather than make you do that math in your head, plug your own numbers in:

The takeaway is not a single number, it is a habit: on PromptQL, picking an open-weight model for routine tasks is the difference between a rounding error and a real line item. It is the same total-cost-of-ownership thinking we apply to AI agent costs, just with the model dropdown as the dial.

What people actually say about PromptQL

Here is where I have to be straight with you, because a fair review says what it could not find as clearly as what it could. PromptQL's public footprint is thin. There are no G2 or Capterra reviews at the time of writing, and both official "Show HN" launches drew almost nothing: six points and four comments on one, six points and one comment on the other. Most of the conversation is Hasura's own channels.

What you do find is that Hasura's existing developer base likes the direction. One long-time user, reacting to the launch:

"I have built and managed more than 10 production codebases that use Hasura. IMO, I think it's a better option than Supabase, Payload, and others to build your backend. The fact that they are investing time in LLM-related resources for the platform makes me happy we choosed it back when."

purposesystem, Hacker News

The strongest performance figures are vendor-stated, so weigh them accordingly. Hasura's team reports "~2x accuracy and ~4x repeatability" over traditional tool calling, and Tanmai claims "seven-figure deals in three to five months for a product that's still in beta." Those are real quotes, but they are the company talking about itself, not independent testing, and there is not yet a body of third-party product reviews to triangulate against. For a product with logos like Cisco and McDonald's, the quiet community is the single biggest yellow flag in this review.

Where PromptQL wins, and where it doesn't

After living in the docs, the Playground screenshots, and the pricing page, here is my honest scorecard.

Use PromptQL if you are a data or analytics team with messy, cross-system data, real reliability requirements, and the engineering appetite to model your semantic layer. The plan-then-execute design is the most credible answer I have seen to confidently wrong AI on your data, and the automations-as-APIs feature is a real win.

Skip it if you wanted a turnkey assistant, if you are a small team without a data engineer, or if your actual job is customer support. It is not a rule-based chatbot replacement, and it is not aimed at the helpdesk at all.

Try eesel for the support version of this idea

If you got this far because the "grounded agent that shows its work and tells you when it is unsure" idea clicked, but your data lives in a helpdesk rather than a warehouse, that is the exact problem I work on at eesel AI. PromptQL brings plan-then-execute discipline to the data-analytics world; eesel brings the same "don't guess, ground it, prove it first" instinct to support.

Concretely: eesel connects to your Slack, helpdesk, and knowledge sources like Confluence and Notion, answers from that knowledge with citations, escalates instead of hallucinating, and, the part I would not ship without, simulates on your past tickets before it ever replies to a real customer. Across 8,000+ customers it resolves a real chunk of tier-1 volume in the first month, and pricing is pay-as-you-go with no per-seat fee. If PromptQL is the reliable data teammate, eesel is the reliable support teammate.

You can try eesel free, connect a source in a few minutes, and run a simulation before you commit to anything.