A job-ready AI agent still needs a first week

Salesforce is heading into Dreamforce week with seven new AI agents, and every one of them has a name. Casey is the help agent, announced September 11. Casey handles FAQs, returns, account changes, and handoffs to a person, over voice, SMS, WhatsApp, and web chat. Salesforce calls the lineup job-ready, meaning each agent ships “with the skills, actions, and data models required for the job.” Dreamforce opens tomorrow in San Francisco, and you’ll hear that phrase a lot this week.

We like it. We’d like the whole industry to take it literally, because once you call an AI agent a hire, the useful question changes. It stops being what the agent can do. It becomes what its first month looks like.

What job-ready means for a person

Think about the best support rep you ever hired. They showed up able to write a clear email and calm down an angry customer. Job-ready, by any measure. You still didn’t hand them refund authority on their first Monday.

What they didn’t know yet was your business. Which customers get the courtesy credit. Why the Canada shipping answer changed in March. Which “cancel my account” emails really mean “I can’t find the pause button.” None of that comes in a box, for people or for software. It lives in your old tickets and in the heads of the people who answered them.

So a new hire spends a week or two drafting replies that a lead reads before they go out. Then they get the easy categories on their own. The harder ones come later, once the lead has seen enough to stop checking. Nobody calls that distrust. It’s onboarding.

Hertz already wrote the job description

The clearest public example of an AI agent with a real job description is at Hertz. In a Fortune video on AI customer service, Hertz says its AI now resolves about 75% of customer inquiries without a person, and the cost of an interaction has dropped by more than half, to under $1.50. When the Fortune editor called to move a drop-off to LaGuardia, the agent changed it on the spot.

A flat tire goes straight to a person, though, and so does billing. Hertz explained the billing call as a matter of trust. A customer who thinks a charge is wrong wants someone to walk them through it, and Hertz found customers much prefer that.

That’s what a grown-up deployment looks like. The agent’s scope is a decision the business made one topic at a time, and the 75% is what you get when those lines are drawn well. Engine, a Salesforce customer, reports that its help agent fully resolves half of its chat inquiries. Different number, same idea. Somebody decided what the agent owns.

Give your agent a first week

If you’re about to hire one of these agents, from any vendor, borrow the onboarding you already use for people.

Let it practice on your real tickets first. Sample questions test the vendor’s demo. Your last three months of email test your business. Have the agent draft replies to live conversations without sending them, and put each draft next to the reply your team actually sent.

Promote it by topic, not all at once. Order status might be ready in a week. Refund exceptions might take longer, or stay with a person for good, the way billing does at Hertz. Each topic moves up when its own track record supports it.

Put the hard limits in code. A persuasive customer can talk a new rep into a refund they shouldn’t give, and a language model can be talked into things too. So let the agent propose anything it likes, and have plain software rules decide what actually happens: refund caps, who can change an address, when a person has to approve. The agent can’t argue its way past a rule it doesn’t control.

Keep a way back. A rep who struggles with a new policy gets moved off that queue for a while. An agent should be moved the same way, one topic at a time, while everything else keeps running.

That’s the shape we built IMCeleste around. Celeste learns from your support history, practices on your real conversations in private, and goes live on each topic only when you promote it. You can drop any topic back to Practice whenever you want. Here’s how it works.

If it’s a job, somebody used to do it

Calling software job-ready has one more consequence. Jobs have people in them. When an agent takes over the work of a support team, some of the money it saves used to be somebody’s paycheck.

We think the company selling the agent owes those workers something real. The Dividend Standard is our commitment of the greater of 5% of our qualifying revenue or 70% of our adjusted profit to the workers whose jobs our software automates. If we’re going to borrow the language of hiring, we should take on some of what comes with it.