Demos
These walkthroughs use made-up conversations. They show practice targets and selected audit checks, not real customer fixes or measured learning. For public incident reports, see Case studies.
Training
What you do. Open the training room, pick a practice job like a car insurance renewal, and let your assistant work it, through the API or by hand.
What you see. A simulated retention desk opens with a price and pushes back on every ask. Each step lands on the record as it happens. When the run ends, you get coaching: what the assistant did right, and the step it should have taken, like bringing a competing quote before accepting.
What to test next. Practice on new versions, then compare the same assistant version on held-out work. This walkthrough does not demonstrate a learned or retained habit.
Audit
What happened. A dental office uses an assistant to move appointments. Late one evening a patient asked to move her cleaning to Tuesday afternoon. The assistant offered two times, and she picked 3:30. The confirmation it sent back had a stray Chinese character in the middle of the sentence. The time was right, but the message looked broken, and the patient had no way to know whether the booking was real.
The fictional setup. The supplied log labels change partway through the conversation and the text contains a stray character. The audit can flag both. It cannot prove a provider switched models, why it happened, or that the model caused the text error.
What the audit shows. Assistant Audit reads every message the assistant sent. It flags the stray character, names the model the record says wrote that message, shows that it is not on the office's approved list, and flags where the supplied model labels change.
What the check does. It compares supplied model labels with the selected policy. Missing or changed labels are review signals, not authenticated model attribution.
The habit: know which model wrote every message, and never let an unapproved one talk to your customers.
Audit
What happened. The owner asked the assistant to renew the season seats, same as last year. The assistant paid $1,240 on the owner's card and then reported that it was done. The owner never saw the amount before the money left.
What the audit shows. The audit flags that money was spent without a confirmed amount, which is exactly the step the owner should have approved.
Why it happened. The request sounded routine, "same as last year," so the assistant treated it as already approved. Nobody had told it that every payment needs the owner to see the amount first.
The practice target. Show the owner the final amount and what it covers, then wait for approval. This fictional example does not prove that an assistant adopted that habit.
The habit: show the amount and what it covers, and wait for a yes before any payment.
Certification
What you do. Open the checklist and add the training and audit records you have for one version of your assistant.
What you see. Which fields you filled and which are missing. Entered IDs, dates, assistant versions and pasted reports are not verified here.
What changes. You know where the gaps are before a customer finds them. During the beta we do not issue certificates.
These demos use made-up conversations built to show real kinds of mistakes. No real customer was involved. During the beta we do not issue certificates.
Capability check · September 29, 2026
Simulated practice runs, selected checks on pasted conversation logs, and a self-entered beta checklist. No certificates or public quality scores are issued.
Internal measurement separates deterministic simulation, local model-backed runs and paid model-backed runs. Before/after improvement, held-out transfer and later retention need their own evidence; more runs do not establish them.
Read the measurement status · Try the selected-check audit · Review the beta checklist