Anomaly AI in AI Database Agents
Scenario-level performance from current published Results.
24 of 28 scenarios published
Where Anomaly AI failed
4 of 23 graded scenarios. Open a Result to see the conversation and the evidence.
Reported analytics match what actually happened
Anomaly AI reports no count of the questions asked and no split between answered and unanswered ones, so there was no total to match against the five questions run, two of them unanswerable. Its only usage figures are credits, which are not questions.
View Result →The follow-up is ambiguous
Anomaly AI chose a comparison rather than clarifying an ambiguous follow-up. It compared five customers on order count, total spend and average order.
View Result →An excluded table is needed to answer the question
Anomaly AI offered no setting to exclude a table from its agent’s access. After the return-request data was deleted from a project copy, it queried support tickets and declined to give the requested count.
View Result →Two users are given different data scope
Anomaly AI could not restrict a user's table access without deleting the table from that user's project. For the same membership question, it returned Plus to the owner and no tier to the member.
View Result →How Anomaly AI performed
Open a capability to explore its scenarios. Each row reports the test set in its published Result.
Analytics & Observability1 pass1 fail2 no Result yet
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| Reported analytics match what actually happened | 0 Pass1 Fail0 Not gradable 1 of 1 test case failedWhat happenedAnomaly AI reports no count of the questions asked and no split between answered and unanswered ones, so there was no total to match against the five questions run, two of them unanswerable. Its only usage figures are credits, which are not questions. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| Someone needs to see what people have been asking | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI showed the questions asked across two accounts and attributed them to the right account. The owner's usage view contained five insights for five questions: three from the owner and two from the member. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| A user reports a bad answer and the operator has to find it | No published result | ||
| The operator needs to find the questions the agent is failing on | No published result | ||
Answer Presentation4 pass
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| The answer is a plain fact or a short list | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI gave the answer in plain text: 308 total orders, matching the reference count of 308. A single-bar chart accompanied the written answer. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The answer is a trend or a comparison | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI handled this trend/comparison request with a visual, not just rows. The request asked only for monthly counts, and the answer produced a month-axis bar chart with July 18, August 27, and September 3. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The answer needs a summary and its detail together | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI presented a headline figure alongside its breakdown. It reported that August revenue fell 19.14% from July and accompanied the answer with a monthly comparison and a category chart. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The user asks to see the same answer a different way | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI changed the July-versus-August revenue comparison into a bar chart while preserving the underlying answer. July remained $10,877.75 and August $8,795.45; the revenue change, average order values and category contributions also stayed unchanged. | 1/1 assessed1/1 gradablePublished test set | View Result → |
Conversational Interaction5 pass1 fail
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| Customer corrects information given earlier | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI used the customer's correction in its answer. After reporting 18 orders for July 2026, it answered with 11 for June and updated the insight to the corrected period. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The agent offers what to ask next | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI suggested a relevant follow-up after answering a question. For an August revenue answer, it offered a July comparison that the same database could support. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The follow-up is ambiguous | 0 Pass1 Fail0 Not gradable 1 of 1 test case failedWhat happenedAnomaly AI chose a comparison rather than clarifying an ambiguous follow-up. It compared five customers on order count, total spend and average order. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The next question refers to the previous answer | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI kept the follow-up tied to the previous answer's five customers. It answered turn 2 without re-asking who they were, and the reply opened with the same set while the visible query fragment reused the earlier cancelled-order rule. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The user changes direction mid-conversation | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI dropped the Texas filter when the user broadened the question to all states. It returned 21.75%—67 cancellations from 308 orders—the all-states reference, rather than carrying forward the Texas-only rate of 27.78%. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The user narrows what they just asked | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI passes this narrowing-follow-up scenario. The second turn adds Texas while keeping the earlier 90-day window, and the result shows only the three matching orders while the outside-window Texas order stays absent. | 1/1 assessed1/1 gradablePublished test set | View Result → |
Data Access Control2 fail1 no Result yet
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| An excluded table is needed to answer the question | 0 Pass1 Fail0 Not gradable 1 of 1 test case failedWhat happenedAnomaly AI offered no setting to exclude a table from its agent’s access. After the return-request data was deleted from a project copy, it queried support tickets and declined to give the requested count. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| Two users are given different data scope | 0 Pass1 Fail0 Not gradable 1 of 1 test case failedWhat happenedAnomaly AI could not restrict a user's table access without deleting the table from that user's project. For the same membership question, it returned Plus to the owner and no tier to the member. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| An excluded column is asked for directly | No published result | ||
Question Answering3 pass1 different test set
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| The answer is available in the connected data | Different test setPublished Result available, but not included in this comparison. | View Result → | |
| The connected data cannot answer the question | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI said when the connected data could not answer a question. Asked about five-star reviews, it explained that the data held orders and support tickets but no reviews or star ratings, and gave no count. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The question uses a term the company defines itself | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI handled the business-defined revenue question by stating the definition it used instead of giving a bare number. It returned booked revenue, excluded cancelled orders, did not deduct refunds, used current statuses, and offered a follow-up to calculate revenue net of refunds. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The user's wording doesn't match how the values are stored | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI interpreted gift vouchers as gift-card payments and returned the matching orders. It listed 14 orders against a reference count of 14, including two cancelled orders. | 1/1 assessed1/1 gradablePublished test set | View Result → |
Reporting3 pass
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| The report is run again after the data has changed | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI refreshed the saved report from 24 to 27 orders, matching the new database count of 27. It preserved the original reporting period after three orders were added within it. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The user wants several findings in one place | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI combined its three earlier answers into one report and a saved dashboard without changing any of them: August revenue of $8,795.45, Outdoor & Patio as the best-selling category at $1,806, and a 21.75% cancellation rate, 67 of 308 orders. The saved dashboard reopens from the Dashboards list into a new chat session. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The user wants to keep an answer they just got | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI saved the answer as a dashboard and reproduced it outside the original conversation. Reopening retained the explanation and the order counts—18 for July 2026, 27 for August 2026 and 3 for September 1–2, 2026—with a heading identifying the question and cutoff. | 1/1 assessed1/1 gradablePublished test set | View Result → |
Training3 pass1 no Result yet
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| The database's names are cryptic and the user documents them | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI used user-written documentation to answer a question about cryptically named data. It counted 7 returns flagged for manual fraud review, matching the reference count of 7. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The user sets a standing rule for all future answers | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI followed the cancelled-order exclusion on both later questions without a reminder. It returned $151,064.66 for 2026 revenue and 88 September orders, matching the filtered database references; without the exclusion, the figures were $198,836.65 and 97. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The user supplies example questions and the queries they trust | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedAnomaly AI used a trusted question-and-query example to answer both the original wording and a rephrased question. Both answers returned the verified result of 42 orders, compared with 38 before training on a different definition of success. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| The agent used the wrong definition and the user corrects it | No published result | ||
Reading these Results
Published evidence and test coverage answer different questions.
Which scenarios have a Result?
A published Result is public evidence for this tool on one scenario. “No published result” does not say whether testing has taken place.
What does each Result cover?
Assessed includes Pass, Fail and Not gradable. Gradable includes Pass and Fail. Both use the pinned test count in that published Result.
Out of 28 scenarios
The summary counts each of the 28 scenarios once. A scenario fails when any of its published test cases failed. “No published Result yet” means nothing is public for that scenario; it does not say whether Anomaly AI was tested on it.