AI (text-to-SQL, insights, Q&A) has a budget with reservation→settlement, controlled egress and provenance in its answers. Data cost comes from processing in the datalake, with a cost guard and cache. See Privacy, egress and budget and Usage and quotas.
AI (text-to-SQL, insights, Q&A)#
| Symptom | Likely cause | How to check | Fix | Prevention |
|---|---|---|---|---|
| AI refuses the action | AI credits exhausted (fail-closed ledger) | check the balance in Usage | top up credits | keep an eye on the monthly allowance |
| AI "doesn't respond" / unavailable | LLM not configured or kill switch on | see Status | wait; confirm the configuration | monitor the capability's status |
| AI answer without numbers/explained refusal | Q&A requires evidence/citation | read the answer's provenance | rephrase with a clear metric/period | ask about published data (Gold) |
| Generated SQL blocked | text-to-SQL defenses (SELECT only, published marts) | see the 403 message | fit the question to the allowed scope | query only published marts |
| AI usage higher than expected | broad/repeated questions | see the Audit log (cost per action) | narrow the scope; reuse answers | encourage objective questions |
Cost and performance#
| Symptom | Likely cause | How to check | Fix | Prevention |
|---|---|---|---|---|
| Query cut off due to cost | hit the per-query volume cap | see the run's error/cost | filter/aggregate; shorten the period | model a lean Gold; partition |
| High projected bill/cost | large table scan | see bytes in the Audit log | filter/aggregate; use partitions | avoid paging through a whole table |
| Paging the API doesn't reduce cost | LIMIT/OFFSET don't cut bytes | billing is by actual bytes | use $select/$filter (OData) | prefer pushdown over raw paging |
| Dashboard slow on refresh | heavy query / no cache | compare the 1st run vs. a repeat | identical queries are served from cache (R$ 0) | schedule refresh off-peak |
Manual refresh blocked (402) | overdue payment (not performance) | see Suspension | the owner settles it | keep an eye on billing |
| Query over billions of rows is slow | volume too large for a direct query | \1 | aggregate in Gold and use the Import model | size with headroom |
BRL values are projections today
x-ingestia-cost-brl and the per-event cost are a projection, not a guaranteed
billed price. The per-query volume cap limits the cost of each query.
Quick cost check#
- Open the Audit log and sort by cost.
- Expand the SQL of the expensive event; check the bytes processed.
- Rewrite with filtering/aggregation; confirm that the repeat is served from cache.
Security and privacy#
- Controlled egress: AI doesn't leak data beyond what's necessary; answers carry provenance — see Privacy, egress and budget.
- Anti-injection and kill switches protect the AI path.
Related#
- Text-to-SQL · Insights · Q&A
- Privacy, egress and budget
- Usage and quotas
- Capabilities and kill switch
Last reviewed: 2026-09-08. Cost values on screen are estimates; the amount charged is the one on the invoice.