The short answer
It does three kinds of work: work that fires on a signal, work that runs on a schedule, and work that stops and waits for a person. In a sales context most of that is reading and drafting, with a human gate in front of anything a customer will see. Most products sold as agents only do the scheduled part, which is automation with a new label.
What are the three kinds of work?
A scheduled task runs at a fixed time whether or not anything happened. An inbox sweep at six in the morning runs on an empty inbox.
A signal-triggered task runs because something changed. A form was submitted, a reply landed, a meeting was booked, a deal crossed a threshold of silence. Nothing fires when nothing happens.
A human-gated task is one the agent prepares and then refuses to complete. It writes the email and stops. It proposes the stage change and waits.
A schedule fires whether or not anything happened. A signal fires because something did.
What does the day look like hour by hour?
| Time | Trigger type | What runs | Where it lands | Who releases it |
|---|---|---|---|---|
| 06:00 | Scheduled | Overnight replies read and classified, threaded to the right opportunity | CRM activity records | Nobody. These are facts |
| 06:20 | Scheduled | Structured fields written from what was actually said. Interpretive fields proposed only | CRM, plus a proposals queue | Human on the proposals |
| 07:30 | Scheduled | Coverage pass over every open thread. Anything gone quiet past threshold gets a drafted next message | Your drafts folder | You, at 08:45 |
| 08:45 | Human gate | You open the queue. Send four, delete one, rewrite one | Sent items | You. Four minutes |
| Any time | Signal | Inbound form submitted. Company researched, prior contact history pulled, reply drafted | Held draft, flagged urgent | You, within the hour |
| Any time | Signal | Meeting booked. Pre-call brief written from the thread, attendees resolved against records | Calendar invite body | Nobody. Internal only |
| 11:00 | Scheduled | Target list construction and research on named accounts | A reviewable document | You, at your convenience |
| 14:00 | Signal | Reply received on a live deal. Response drafted against what was promised | Drafts folder | You |
| 16:00 | Scheduled | Reconciliation pass. Contacts on threads that have no CRM record get created. Conflicts flagged | CRM, plus an exceptions list | Nobody on creates. You on conflicts |
| 17:30 | Scheduled | End-of-day summary: what was sent, what is waiting, what it declined to touch and why | One message to you | You read it |
| Any time | Escalation | A deal quiet past the second nudge. No third draft written. A decision is raised instead | Your queue, marked decision | You |
Add up the human column. Well under an hour across the day, against coverage of every open conversation rather than the ones you remembered. Your own number will depend on how many drafts you rewrite rather than send, which is the figure worth tracking.
Notice what is absent from that table. No autonomous sending, and no stage advanced on the agent's own authority. No customer is ever addressed by the machine directly.
Why does the human stay on anything customer-facing?
Two findings, and together they close the argument.
Dell'Acqua and colleagues, in a field experiment with 758 consultants, found that on tasks requiring judgment AI users were 19 percentage points less likely to reach the correct answer, while producing work that was better presented than the control group's. Not worse-looking and wrong. Better-looking and wrong.
Luo et al., in Marketing Science 2019, ran a field experiment on outbound calls and found that disclosing AI identity before the call cut purchase rates by 79.7%. Buyers penalise the machine heavily once they know it is one.
Dell'Acqua is why the output gets checked. Luo is why a person is the one who sends it. Both findings are about a moment in time and a particular arrangement, and both are conditions you design around rather than properties you are stuck with.
Concealment is what you get if you read Luo alone, and it misleads the buyer and breaks the moment they ask. The working arrangement is different. A person is the sender, and that person is accountable for what arrives. The agent wrote a draft. A human sent an email. Both statements are true, which is the only arrangement that survives being asked about.
Karlinsky-Shichor and Netzer measured the value of that arrangement directly. Across 17 reps and 67,851 quotes in Marketing Science 2024, human-machine collaboration returned 7.8% profit improvement against 4.9% for full automation. The hybrid beat the machine alone by a wide margin, in a study that ran both.
How do you tell a real agent from renamed automation?
Gartner's June 2025 assessment is the most useful number published on this. Of the thousands of vendors describing themselves as agentic AI, roughly 130 were judged genuine. Gartner also expects more than 40% of agentic AI projects to be cancelled before the end of 2027.
Automation does what you scripted. Autonomy decides what to do next.
| Question | Renamed automation | Actual agent |
|---|---|---|
| Can it decide to do nothing? | No. The trigger fires, the action runs | Yes. Judging that no message is the right move is a valid outcome |
| Who sets the order of work? | You did, in the builder | It ranks what it found this morning |
| What happens on an unanticipated case? | Falls through, or errors silently | Escalates it to a person with the reason attached |
| Can it use a tool you did not name in advance? | No | Yes, within the boundary you defined |
| Is there a record of why it acted? | A run log of steps executed | A decision log with the reason for each choice |
That last row is the one I would not compromise on. Without the audit log, autonomy is just hope. With it, autonomy is a contract.
What does it not do all day?
It is idle for most of the day. A well-scoped sales agent is doing real work for perhaps ninety minutes out of twenty-four hours, and anyone describing a machine that works continuously is describing a scheduler running on an empty queue.
It does not learn your business by watching. Positioning and the reasons you lost the last three deals get written down by a person or they do not exist.
It does not improve a process that is broken. A bad list researched perfectly is still a bad list, delivered faster.
And it does not run itself. Gartner's survey of 227 chief sales officers found 72% reporting low reinvestment of AI-saved time. The time comes back and then goes nowhere. If nobody opens the 08:45 queue, the entire day above produces one thing: drafts, unread, in a folder.
The whole design rests on a person spending fifteen minutes. That is the actual dependency, and it is worth checking you will spend them before anyone builds you anything.