
For your team
Teams by Function
Module · When the bot answers first
The Support Team AI Check
A support bot has one honest job: resolve the request, or hand it to someone who can, fast. The failure mode is quieter than an outage. The bot closes tickets it never solved, the customer gives up, and your dashboard calls that a win. This module checks the five things that separate real automation from expensive deflection: whether you measure resolution or just containment, how fast a stuck customer reaches a human, whether the agent-assist your team leans on is worth trusting, whether the knowledge behind the answers is current, and whether satisfaction survives contact with the bot.
What the five levels look like
Every dimension in this assessment is scored 1 to 5. This is what the levels mean, dimension by dimension. The graded report diagnoses where your own answers land and what to do about it.
Resolution, not deflection
- 1Containment only
- 2Ticket volume
- 3Ended-chat rate
- 4Resolution tracked
- 5Resolution plus follow-up
At the low end: A containment number with no resolution behind it flatters the bot and hides abandoned customers. Start measuring whether the issue actually got solved, by follow-up survey or repeat-contact rate, before you trust the automation. What good looks like: Measuring true resolution plus repeat contact is what lets you scale the bot without hollowing out service. Keep watching the follow-up rate; a resolution number can drift the moment the knowledge behind it goes stale.
Escape hatch is fast
- 1No human route
- 2Buried deep
- 3After several tries
- 4One clear request
- 5Instant on intent
At the low end: A bot with no fast route to a human traps your most frustrated customers exactly when they need help most. Add a visible, one-step handoff to a person and make it work on the first ask. What good looks like: An instant handoff the moment the customer signals they are stuck keeps a bot failure from becoming a service failure. Watch the handoff volume; a rising rate is the bot telling you where it is weak.
Agent-assist is trustworthy
- 1Blindly pasted
- 2Often wrong
- 3Sometimes useful
- 4Reliable, checked
- 5Reliable, sourced
At the low end: Suggestions your agents paste without reading turn one AI error into a wrong answer sent to a customer. Require a quick check on every AI draft, and measure how often agents have to correct it. What good looks like: Reliable suggestions with a visible source let agents verify in a glance instead of retyping from scratch. Keep sampling the drafts for accuracy; assist quality tracks the knowledge base and decays with it.
Knowledge is current
- 1Never reviewed
- 2Ad hoc edits
- 3Reviewed yearly
- 4Owned and scheduled
- 5Owned, fresh, audited
At the low end: A knowledge base nobody maintains feeds your bot confident, out-of-date answers. Assign an owner and set a review cadence before the next product or price change makes the bot lie for you. What good looks like: Owned, scheduled, audited content is what makes an AI answer trustworthy at all. Tie the review cycle to your product and pricing changes so the knowledge updates the day reality does.
Satisfaction survives the bot
- 1Not measured
- 2Bot rates far worse
- 3Bot rates worse
- 4Roughly equal
- 5Bot equal or better
At the low end: If you never compare satisfaction by channel, the bot can be shedding loyal customers while the average score holds steady. Split your satisfaction metric into bot-handled and human-handled and look at the gap. What good looks like: Bot-handled satisfaction that matches or beats human contact means the automation is genuinely serving customers, not just deflecting them. Keep the comparison live; the moment knowledge or intent handling slips, this gap opens first.