Where AI agents actually save time

Every week someone asks us whether an AI agent could “handle support”. The honest answer is: part of it, and the part is usually smaller and more boring than people expect. That’s good news. Boring is where the hours are.

Start with the inbox, not the model

Before we talk about models, we export the last three months of conversations and sort them by what the customer actually wanted. In almost every support inbox we’ve looked at, five or six intents make up more than half the volume: where is my order, how do I return this, can I change my address, when will it be back in stock, what’s your opening time.

These questions share three traits:

• They repeat. The same question, worded differently, hundreds of times a month.

• The answer lives in a system. An order record, a carrier scan, a policy page.

• A wrong answer is cheap to fix. If the agent misreads a tracking status, a person corrects it in the next message.

When all three are true, an agent can close most of those conversations without a person, and the time saved is easy to measure: count the conversations, multiply by the minutes each one took.

Where agents don’t save time (yet)

The opposite traits tell you where to wait:

• Low volume. Fifty unique questions a week are cheaper to answer by hand than to build, test and monitor an agent for.

• The answer lives in someone’s head. If your best agent “just knows” which customers get an exception, write that down first. An agent can’t read minds, and neither can a new hire.

• Mistakes are expensive. Refunds, cancellations, anything legal or medical. Agents can prepare these — gather the facts, draft the reply — but a person should press the button.

The 60/40 pattern

In our projects the pattern is remarkably stable: once the agent is tuned, it closes somewhere between 50% and 70% of conversations on its own. The rest go to a person, with the conversation, what the agent looked up and why it stopped.

That second part matters as much as the first. A good hand-off saves the person two or three minutes of re-reading and re-checking. On a team handling a thousand messages a week, that alone is more than a day of work.

How to estimate it for your team

You don’t need a vendor to get a first number:

1. Export a month of conversations and tag the top ten intents by hand. It takes an afternoon.

2. For each intent, note where the answer lives and what a mistake would cost.

3. Multiply the volume of the safe, system-answerable intents by the average handling time.

4. Assume the agent closes about half of those in the first month, and that a person still spends a little time reviewing.

That’s the same arithmetic as our ROI calculator — and the same conservative assumptions. If the number doesn’t justify a build, don’t build one. A better FAQ page and a few saved replies might be the right first automation.

What to measure after launch

Once an agent is live, three numbers tell you whether it’s saving time or just moving work around:

• Closed without hand-off, by intent, every week.

• Hand-offs that needed no change — if a person approves the agent’s draft as-is most of the time, the approval step can probably be relaxed for that intent.

• Re-opened conversations — the honest measure of whether customers got what they needed.

If those move in the right direction for eight weeks, widen the scope. If they don’t, the agent is telling you something about your process, and that’s worth fixing first.