Currently, when character limits are configured in AI Draft (at the project AI instruction level, section level, or via AI Draft instructions), the AI does not reliably honor those limits. Responses that exceed the configured limit are still surfaced to the user and can be accepted. The system only flags the violation with red text after the fact, rather than preventing or prompting for revision before acceptance.
This behavior stems from the inherently non-deterministic nature of AI systems, which cannot guarantee strict character counts. However, customers require a more reliable enforcement mechanism.
Customer impact:
The customer set an 800-character limit across multiple configuration levels for their project, but AI Draft consistently generated responses exceeding that limit. The issue was confirmed as replicable in the customer's instance.
Requested behavior:
- AI Draft should either reject responses that exceed the configured character limit (not surfacing them to the user at all), or
- Prompt the user for revision before allowing acceptance of an over-limit response.
TKT-218240
