Today, 18:00

Participant context in your AI assistant, six new agents for hotels, retail and banks, and faster outlines

Features

Ask your AI assistant where a participant came from. Through the Perspective MCP server, conversations now carry the participant context you passed in, like source, campaign or a CRM ID, with the same full values you already get in exports and webhooks. Pull it for a single conversation, or switch it on when listing many. Read participant metadata through MCP.
Six new agents are ready to adapt for hotels, retailers and banks: a hotel guest satisfaction survey that gets the reason behind every score, guest journey research from booking to checkout, a review recovery check-in that catches a problem while the guest is still there to help, shopper insights research on why people chose you and what almost sent them elsewhere, a bank customer experience survey for the whole relationship, and a transactional NPS survey for banks that captures the why right after a branch visit, call or app task.

New outlines are ready in about half the time. Outline reviews in the design chat also grade more strictly than before, so expect more parts of an outline to be flagged as needing attention.
Conversation summaries stick closer to what participants actually said.
If you chose Claude Sonnet as your text interviewer, your conversations now run on Claude Sonnet 5.5. GPT-6.1 Sol is now the default for agent design, analysis sessions and scheduled insights. Anything still set to the retired GPT-6 Sol moves to it on its own, and models you picked yourself stay as they are.
Starting an agent from a written brief picks the right agent type in about half a second, and trust scores on finished conversations are ready in about a fifth of the time.
Wondering how much a share link can carry? The URL parameters guide now spells it out: there is no app limit on how many custom parameters you add, though the whole link still has to fit what browsers, email tools and servers accept. Values are trimmed to 50 characters in what the agent sees and in MCP batch transcripts, while exports, webhooks and MCP participant metadata keep the full value.

Outline reviews now score every dimension instead of leaving some marked unavailable.
A results field that asks whether a conversation happened by voice or by text now gets voice conversations right, instead of marking every one as text.
Results only record an option when the participant actually picked it. A reason that fits none of your options now goes into your "other" field instead of disappearing.
On Concierge agents, a visitor who asks a real question and then leaves is no longer written off as abandoned. What they asked for shows up in your results, and the summary no longer says they never engaged.
Agents told to be helpful no longer take that as permission to do a visitor's own work, like writing their reports. They say what the conversation is for and carry on. Agents built to produce something for the participant still do.
When someone answers through their AI assistant, the agent now asks a few related questions per turn from the start, so more of what you need is collected before the assistant stops.
Design sessions that were set to GPT-6 Sol no longer open with no model selected in the picker.