See what the assistant actually did查看助手到底做了什么
A record of every message the assistant handled — what it looked up, what came back, and what it replied.助手处理过的每一则讯息的记录——它查了什么、查到什么，以及回了什么。
这篇指南的正文目前只有英文版。标题、摘要与导览已翻译；内文仍在翻译中。
Letting something answer your customers unsupervised is a leap of faith. This is the page that turns it into a checkable one.
Every time the assistant handles a message it records what it did: the question that came in, which of your records it opened, what those records said, and the reply it sent. One entry per message.
Who can see it
Your super user, and nobody else. It contains whatever the assistant read, which means real families and real money.
1. Open it
CRM → AI Activity. The list is newest first, with a line per message handled.
The tiles at the top count how many messages the assistant handled over the period you pick.
2. Open one and read it as a sequence
Each entry expands into the steps in the order they happened, which is how a person works and how this reads:
- What came in — the customer's message.
- What it looked up — each record it opened, and why. "Checked the Tuesday 5pm class for space."
- What came back — the actual answer from your data. Use Show what came back on any step.
- What it replied — the message the customer received.
- Where it stopped — if it went silent and flagged for a human, the reason is here.
That third step is the one worth your attention. If a reply looks wrong, it is almost always because the record it read was wrong, not because the assistant invented something — and this is where you see which.
3. Check the instructions it was working from
An entry also shows the instructions in force for that message, split by where each part came from:
- Yours — the personality, the background about your academy, the playbooks you wrote, and your never-discuss list.
- Built in — Hull's own guardrails, which you cannot switch off.
Worth reading when the assistant sounds wrong rather than says something wrong. Tone almost always traces back to a sentence in your own personality or background, and this shows you which sentence to edit. See Set up the AI agent.
What is kept, and for how long
The full record of what the assistant saw is encrypted at rest, the same as parent data, and deleted after 30 days. The list view is built from a summary and never opens the encrypted part, so browsing the page does not decrypt anything.
Thirty days is deliberate: long enough to investigate a complaint, short enough that a copy of your customer conversations is not accumulating indefinitely.
What to do with it
- A parent says the assistant told them something wrong. Find the message, read what came back at step 3. Either your data was wrong — fix the data — or your instructions permitted it — fix the instructions.
- It keeps handing conversations to staff. The stop reason on each entry tells you why. A pattern here is usually a playbook that does not cover something people keep asking.
- Before you widen what it may answer. Read a week of entries first. It is the cheapest possible way to find out whether you trust it with more.
Related: Set up the AI agent · Approve what the agent proposes · Work the WhatsApp inbox.