AI deflection vs. deflection theatre
In short
Real deflection means the customer got a correct answer and did not need a human within 24 hours. Deflection theatre means the AI closed its turn — usually with "please contact our support team" — and a human redid the whole exchange later with an irritated customer. The difference is whether the agent is grounded in live commerce data and escalates honestly, not how confident its writing sounds.
"Deflection rate" is an easy number to inflate and a hard one to inflate honestly. A bot that answers "I don't understand, please contact our support team" technically deflects — it closes the AI turn of the conversation — while quietly guaranteeing a human has to redo the whole exchange later, now with an irritated customer.
We call this deflection theatre: metrics that look good in a weekly report and feel terrible in the actual chat thread.
What real deflection requires
Grounded answers. An agent that can see your actual SKUs, live stock counts, and order status doesn't have to guess — it can tell a customer their order shipped yesterday instead of reciting a generic "please allow 3-5 business days" that may already be wrong.
Honest escalation. A well-built agent knows what it doesn't know — a price dispute, a damaged-item claim, an angry repeat complaint — and hands those to a human immediately, with full context attached, instead of stalling with reassurances.
Measured by outcome, not by turn count. The metric that matters is: did this conversation need a human within 24 hours after the AI's reply? Not: did the AI send a reply at all.
The tell
If your deflection rate is high but your negative reviews mention "the bot didn't understand me" or "I had to explain everything again," you're running deflection theatre. The fix isn't a better script — it's giving the AI access to the same commerce context a good human agent would use: product catalogue, order history, shipping status, and the actual conversation history, not just the current message.
Frequently asked
- How do I measure real deflection?
- Take the conversations your AI handled end to end, then check how many were reopened by the customer or touched by a human within the next 24 hours. The share that were not is your real deflection. Anything measured at the moment the bot replies is measuring effort, not outcome.
- Should the AI always try to answer?
- No. Price disputes, damaged goods, refunds outside policy, and a customer who is already angry should go to a human on the first message, with the full thread attached. An agent that stalls those with reassurances converts a solvable complaint into a public review.
- What makes an AI answer 'grounded'?
- It cites facts pulled from your live systems — this SKU, this stock count, this order's shipping status — at the moment of replying. An agent working from a static FAQ or a copied product description is generating plausible text, which is exactly how a bot promises a size that sold out last week.
Read in中文Bahasa Malaysia