Give every agent a job description

Most companies still file AI under “tools.” It shows up in the budget next to software licenses, and in the org chart nowhere at all. On September 11, Salesforce started selling agents the other way: as staff.

What shipped

Salesforce announced seven agents with human names and job titles. Six are generally available, covering customer service, IT and HR, shopping, supply chain, inbound leads and customer experience. A seventh, Hunter, handles outbound sales and is in pilot, with general availability planned for November. Casey answers customers. Piper works inbound leads.

Salesforce also added support for agents that pursue a goal over days or weeks rather than finishing one request and stopping. It says its agents have delivered 7 billion “agentic work units” to date. That number is company-reported, and “work unit” is Salesforce’s own measure, so treat it as a sign of intent rather than a benchmark. The announcement came four days before Salesforce’s Dreamforce conference, which ran September 15 to 17.

Why a CEO should care

The product news is less important than the packaging. A major software vendor has decided that customers understand agents best as roles: a name, a remit, a place in the company. Whether or not you ever buy Agentforce, your team is going to start thinking about agents that way.

That is useful, because a role comes with questions every CEO already knows how to ask. Who does this report to? What does good look like? What does it cost? What is it never allowed to do? Most small companies run agents without answering any of them. The agent lives in one person’s browser, nobody else knows what it touches, and nobody can say whether it is working.

The fix does not need a platform. It needs a page.

The agent role card

Before any agent does recurring work for the company, fill in one card. If a line stays blank, the agent is not ready for unattended work.

Role. One sentence, written the way you would write it for a person. “Drafts first replies to support tickets tagged billing.”
Output. What it hands over, to whom, and in what form. A draft, a ticket update, a row in a sheet.
Measure. One or two numbers, each with a baseline from before the agent started. Time to first reply. Share of drafts sent without edits.
Supervisor. A named person, not a team. This person reads a sample of the agent’s work every week and owns its mistakes.
Budget. A monthly ceiling for model and tool spend, and what happens when it is reached.
Access. The systems it can read, the systems it can change, and the ones it must never touch.
Stop rule. The situations where it hands the work to a person instead of guessing. Angry customer. Refund above a set amount. Anything legal.
Review date. When the supervisor decides to keep, change or retire it.

Line 4 matters most. An agent without a named supervisor is not a teammate. It is an unattended process with a friendly name.

A worked example

Take a hypothetical 40-person software company with three people in support. The CEO wants to “use AI in support” and has been told to buy an agent.

Filling in the card changes the conversation. The role narrows from “support” to “first replies on billing tickets,” because that is the work with clear rules. The measure becomes time to first reply and the share of drafts the team sends unchanged, both measured for four weeks before launch. The support lead becomes the supervisor and gets two hours a week to do it. The access line rules out issuing refunds. The stop rule sends anything that mentions cancellation to a person.

None of that required a new vendor. It required the CEO to treat the agent the way the company treats a hire: with a scope, a manager and a way to judge the work.

It also changed what the team heard. “We are buying AI for support” sounds like a headcount plan. “We are giving the support lead an assistant for billing replies, and they will judge whether it earns its place” sounds like help. How you describe changes to people’s work carries real weight, and decisions about roles and employment carry legal and HR duties. This is education, not HR or legal advice; take those questions to qualified advisers.

A decision rule for the org chart

Put an agent on the org chart when it has a filled-in card and a named supervisor. Leave it off when it does not. The chart then shows what it should: who is accountable for every piece of recurring work, human or not.

What to do this week

List every agent or automation that already does recurring work in the company. Ask each team; some will surprise you.
Pick the one that touches customers or money, and fill in its role card with its owner.
Where no supervisor can be named, pause it or assign one before Friday.
Add one line to your next leadership meeting: which agents exist, who supervises each, and what each one costs.

The vendors will keep naming their agents. The CEO’s job is to make sure each one has someone it answers to.