Check what your chatbot captured, test an agenda, and read its stats

When software fills in your CRM for you, the fair question is how do you know that's right? Every value an agenda captures records where it came from.

What the Inbox shows

Open a chat conversation and the panel on the right lists what the bot has gathered — each field, its value, and whether it was asked yet. Values the bot inferred with AI are marked AI; hover for the model and its confidence. Everything else matched a recognised format — an email, a phone number, an amount, a menu choice — and is as reliable as the visitor's typing.

That distinction is the point. Before you take a conversation over, you can see which details are worth double-checking instead of treating everything as an equally solid fact.

Personal data in the log

Tick Personal data on a field and its value is redacted everywhere the provenance log is shown:

• an email becomes s…@acme.com — you keep the domain, which is what tells you whether it's a work address;
• a phone becomes its last four digits;
• free text becomes a character count.

The real value still goes to the contact record, where it belongs and where your deletion tooling can find it. The log keeps enough to recognise a value you already know, and not enough to learn one you don't — and deliberately isn't a hash, because a hash of an email address is still a stable identifier for that person.

You can also set a retention window on an agenda. After it passes, the stored values in the log are cleared while the trail of what was written, when, and by which model is kept — the audit question is about the write, not the value.

Testing an agenda before it goes live

On an agenda, use Test and type a few messages a visitor might send. You'll see, turn by turn:

• what the agenda would extract from each message;
• which single question it would append;
• where each value would be written;
• how the visitor would qualify, and their score;
• which turns would have needed an AI call — so you can see what the agenda costs before it's live.

Nothing is written, no model is called and no credits are spent. If a field is never picked up in your test, it won't be picked up in reality either — usually a sign the question needs rewording or the field needs a type.

Stats

Each agenda has a funnel: per field, how often it was asked, answered and written, plus completion rate, qualification split and median time to complete. The number worth watching is the share of values that came from AI rather than a free format match. A low share means a cheap, predictable agenda; a high one is worth a look — often a field that should be a choice or a number is sitting as free text.

Reporting on agendas

Agenda runs are a report source, so qualification rate by agenda, completions per week or outcome by channel are ordinary reports you can filter, chart, pin to a dashboard and export — not numbers trapped in a statistics panel.