How accurate is our AI?
Accuracy claims mean little without a method. Here is ours: a labeled test set, the numbers it produces, and what it does not tell you.
AI marketing review
30 marketing pieces written to test the SEC Marketing Rule (206(4)-1) and FINRA Rule 2210, each labeled with the findings a compliance reviewer should raise. We run the full set before any change to the model or the instructions, and keep every run. Latest run: October 18, 2026.
100%
Violations found
Every planted violation across 30 labeled pieces was flagged (recall 1.00).
0%
False flags on compliant pieces
Compliant pieces, including ones built with traps, drew no serious finding.
96.7%
Right overall verdict
29 of 30. The one miss was too strict: a private placement summary that defers to the PPM was marked for revision.
100%
Required disclosures identified
Every missing disclosure the case expected was named.
A review takes about 75 seconds on average.
What the 30 cases cover
- Performance: net vs gross, 1/5/10-year periods, hypothetical, extracted and predecessor performance, the private-fund exception
- Testimonials, paid endorsements and third-party ratings
- Guarantees, unsubstantiated claims, fair and balanced presentation, omitted conflicts
- Firm-specific checks: a fee that contradicts the firm’s own policy, a stale disclaimer
- FINRA Rule 2210 retail communications and correspondence; dual-registrant capacity
- Five compliant pieces with traps, to measure false flags
AI compliance assistant
24 questions asked against three test firms: an SEC adviser with a December fiscal year and its own published policies, an SEC adviser with a June fiscal year, and a state-registered adviser. An answer passes only if it states the right rule and dates for that firm and cites the firm's own policy where one applies. Latest run: October 18, 2026, 24 of 24.
The first run scored 18 of 24. It caught real errors, such as telling a state-registered adviser it was regulated by the SEC and assuming a December fiscal year. We fixed them before publishing this page.
What these numbers do not tell you
- They come from our own test set. Real marketing varies more than 30 cases can.
- The same piece reviewed twice can occasionally land on a different overall verdict while raising the same concerns.
- The AI drafts; your CCO decides. Every finding names the rule, quotes the sentence and explains why, so it can be checked, and the record keeps both the AI's assessment and the CCO's decision.
Want to see it on your own material? Start a 14-day trial, or open the interactive demo.
AI you can check
Every finding cites the rule and quotes your text. From $99/mo, 14-day free trial, no card.