Figures show the time these tasks take today; actual savings depend on the work. ¹ Anaconda, State of Data Science 2022 (about 3,500 respondents). ² McKinsey Global Institute, The Social Economy (2012). ³ ThoughtSpot BI survey, January 2015 (60 IT/BI leaders; a vendor survey). ⁴ Our banked token census, 2026-10-03: 18 runs of three multi-table questions on two hosted models took 11 s to 6.7 min each (median 16 s on gpt-oss-20b, 4.7 min on gpt-oss-120b on a slow provider day); the same questions ran in 7–33 s the day before. Speed depends on the model and the provider's load.
A question costs one credit, a deck export ten; a question that fails is refunded. Every plan gets the same safety rails: read-only SQL, validated charts, full trace. You pay for capacity and privacy, never for safety.
| ProLyt | Wren AI | Julius AI | Vanna AI | |
|---|---|---|---|---|
| Entry price | €29 per user per month (Analyst) | $179/month, billed annually1 | €18/month (Plus)2 | $50/month3 |
| Free plan | ✓ | — | — | — |
| Published on-prem price | ✓ | ✗ | ✗ | ✗ |
| Self-hosted option | ✓ | ✓ | ✓ | ✓ |
| Use your own model/key | ✓ | ✓ | ✗ | ✓ |
| Frontier models included without own key | ✗ | — | ✓ | — |
| Data cleaning built in | ✓ | — | — | — |
| Document search | ✓ | — | — | — |
| Dashboards | ✓ | ✓ | — | — |
| Slide decks | ✓ | — | ✓ | — |
| Published token-cost measurement | ✓ | — | — | — |
✓ yes · ✗ no · — not stated publicly. Marks follow each vendor's own public pages (sources below); a dash is a gap in what is published, not a judgement.
From public pricing and product pages, October 2026: 1 getwren.ai/pricing · 2 julius.ai/pricing · 3 vanna.ai · 4 public directory listings. Plans change; check each vendor for current terms.
The AI model writes the query and the database computes the numbers. Fixed rules around the model check the answer before anyone sees it.
What scaffolding cannot prove: that the model chose the right query. That is why every answer shows its SQL and its sources.