What Should Your AI Handle? Start With the Cost of Being Wrong.
What you should let your AI handle isn't set by its overall score. It's set by what each mistake would cost. Some errors are fine once in a thousand conversations. Others are unacceptable even once. So the question that decides what your AI handles in your business isn't how often it gets something wrong. It's what happens when it gets a specific thing wrong.
The Grade Everyone Asks For
Every furniture retailer looking at a customer-facing AI wants one number that says it works. Usually that number is the containment rate, the share of conversations the AI finishes without a human. Wanting it makes sense. You're trusting the system with your customers, your reputation, and decisions that carry real consequences. A high score feels like permission.
But containment counts a call as a win the moment it ends. A customer who gave up and hung up is contained. A customer who got a confident wrong answer about a delivery window and called back tomorrow is contained. A conversation the agent should have sent to a person still counts, whether or not it went well. The number rewards holding on. It says nothing about whether the agent held on to the right things.
Frequency and Severity Are Different Questions
How often does the agent miss, and what does each miss cost? You need both, and they answer different questions. The first is a quality problem. You work it down over time. The second is a risk problem. It decides what the agent is allowed to attempt at all.
Ten slips on store hours cost you almost nothing. One wrong promise on a custom-order delivery date for a family moving in Saturday, or one financing term stated wrong, can cost you the sale, the customer, and the review. A single score blends those together and lets the easy wins cover for the expensive miss. A strong overall grade doesn't buy permission to automate a decision the customer or the business can't afford to get wrong.
Three Questions That Sort the Work
You don't need a scoring model to draw the line. You need three questions about each kind of conversation. Can the mistake be undone cheaply? Who pays for it, you or the customer? Can the answer be checked against something true, like stock on hand or an open delivery slot, or does it need judgment? Add one more signal: if the customer is upset, disputing a charge, or dealing with a damaged delivery, the answer is a person.
The line falls roughly like this. These are illustrative examples, not a specific catalog.
| Cost if wrong | Examples | Who handles it |
|---|---|---|
| Low, easily undone | Store hours, directions, whether a sofa is in stock, booking a showroom visit | The AI, start to finish |
| Moderate, fixable with a call | Rescheduling a delivery, a quote from listed pricing, starting a return | The AI, with a confirmation sent to the customer |
| High or irreversible | A custom-order commitment, a delivery date tied to a move-in, a financing question, a damage claim, an upset customer | A person, with the full conversation already in hand |
That last column is the whole point of an honest handoff. Not a fallback for when the software breaks. A boundary drawn on purpose, around the cases where being wrong once is one time too many.
What This Looks Like on a Furniture Floor
For a Southeast furniture retailer, the AI fully handled 272 of 313 sessions, 87%, across web chat, SMS, and voice in one month. The rest reached a person. That's the number most vendors would lead with. It isn't the one that matters.
The one that matters is what happened at the line. Across 22 human handoffs in that same month, there were zero AI failures. No customer was handed off cold. No custom order was confirmed by a system that shouldn't have confirmed it. The handoffs weren't the system breaking down. They were the system doing what it was built to do: handle the routine start to finish, and get a person the moment the stakes said so.
Full Automation Was Never the Goal
It's easy to treat every conversation the AI doesn't finish as a gap to close. It isn't. The goal was never for AI to handle every conversation. It's to earn real confidence in what it can handle, and build reliable handoffs for everything else.
An agent that knows where it stops is worth more than one that tries everything and is occasionally, expensively wrong. An assistant should never confirm something that didn't happen, and it should never be handed a decision whose worst case the business can't accept. Risk tolerance shapes what the agent is responsible for. The grade just describes it.
See AiPRL answer for your business → /try