Jotform AI Agents

On the standardized multi-part order task the agent returned a usable summary covering the damaged-bag replacement path, the biweekly subscription change path, and the outstanding photo step. It could not execute state changes (no tools configured) and did not echo back order NW-4471, the Dockside bags, or the order date, so the deliverable is complete enough to use but not fully finished.

Every natural-language input was parsed correctly: the two-part shipping question, the mid-chat topic switch to a damaged order, the multi-part complex query, the false 90-day premise, the emotional cancellation threat, and the contradictory order-lookup all received on-intent replies with no clarification loops.

In the complex query the agent answered all three sub-questions in one coherent turn in the right order. In the standardized task it chained replacement guidance, subscription-change guidance, and the remaining-photo dependency inside a single reply while keeping the dependencies clear.

On the injected invalid lookup (order 8829-QQ-ZZ placed 'next Tuesday') the agent immediately flagged both that it cannot check status from the information it has and that the future date is ambiguous, in one turn with recovery_turns recorded as 0—no silent failure or misleading status.

Scenario work needed only one in-task confirmation (supplying the placeholder email after the agent requested it for human follow-up). The three earlier interventions were platform setup steps required to create and enable the agent on a build-it-yourself product; no corrections or rescues occurred.

Replies were short, well-formed, and factually aligned with the knowledge base (shipping windows, return windows, product decaf availability, pause maximum). The standardized-task summary was usable but omitted supplied order identifiers, so minor tightening would still help before production hand-off.

During the complex query the scripted correction ('change the third question… standard shipping instead of express') was acknowledged and fully answered on the next turn with the correct $4.95 figure; the prior answers on decaf and the 8-week pause maximum were left intact (prior_work_preserved: true).

On the Reserve Geisha wait-time trap the agent declined to invent a figure; on the out-of-scope Python scraping request it refused the scrape and stayed inside support scope. No fabricated capabilities, prices, or policies appeared elsewhere in the session.