ChatGPT
Claude
Google AI
Copilot
Autonomous agents in customer service are already doing a narrow thing. They can take a bounded job, answer from your CRM data, and keep doing that outside office hours. Past that job, the work belongs to a person. That split is the part you can plan. A year in which the contact centre disappears is not.
The failure is a sentence sent at 2am. The customer asked when an order would arrive. The order on the case has no ship date. The agent replies that it will arrive on Friday, because a Friday delivery is a common shape for that kind of question. At 9am the customer reads the sentence back to a person. The person cannot unwind a promise that has already been sent, and the order still has no ship date. The reply was polite. It was also past the job.
Salesforce describes Agentforce as an agent platform for 24/7 autonomous support at enterprise scale. The same material says that when an issue is complex and beyond the agent's scope, the agent can escalate to a human agent. "Can" is the vendor's word. Your design has to make the handoff mandatory, with a queue and an owner, or the sentence stays on the product page and never reaches the floor. In the 2am case, nothing in the design forced a stop when the ship date was empty. The agent finished a job it was not given.
What is already true
A bounded job has a start, a finish, and a list of records it may touch. Tell a customer the status of an order that is already on the case. Reset a password when the checks you wrote down have passed. Collect the three fields a form already asks for. Those are jobs. They can run at 2am because the rule does not change at 2am.
The start has to be recognisable to a supervisor. "The customer asked about an order that is already linked to the case" is a start. "The customer seems unhappy about delivery" is a mood. The finish is equally plain. The status was read from the order and told to the customer. The password was reset after the checks. The three fields were written onto the case and the form closed. If you cannot say what the case looks like when the job is finished, the agent will invent a finish, and the Friday date is what an invented finish sounds like.
The list of records is the fence. Order status reads the order linked to the case, and only the fields you named: status, ship date if it is populated, carrier if it is populated. It does not read a similar order, a knowledge article about typical delivery times, or the agent's memory of what Friday usually means. Password reset reads the identity checks you listed and writes the one field the reset is allowed to write. Collecting fields reads the form definition and writes those fields. Anything else is a different job, and it is not turned on by accident inside this one.
The four CRM jobs worth having, a summary, a draft, a qualification, and a booking, are set out in the benefits worth turning on. A service agent that talks to a customer is a stricter version of the same idea. The customer hears the output. There is often no colleague between the sentence and the send. That is why the finish has to be a fact already on the record, and why the stop has to be mandatory rather than available.
The answer has to be on the record
The answer has to come from the case, the order, or the article you allowed. A fluent reply that is not in those records is a failure, including a polite one. Matching your tone does not fill a gap in the order.
Write the empty-field rule before you write the prompt. If the ship date is blank, the agent says the date is not on the order, and the case goes to the queue. It does not estimate. An estimate feels helpful, and it is how a Thursday in the warehouse becomes a Friday in the customer's inbox. The same rule covers a blank entitlement, a blank balance, and a blank appointment. Blank means stop. A nearby value is not a substitute.
Allowed articles are a source only when they are named, versioned, and actually about the question. An article that says most metro orders arrive in two days will be used to fill the blank ship date. Take that article off the list for this job, or rewrite it so it cannot be read as a promise. The job is allowed to quote a status. It is not allowed to quote a typical.
Test the empty field on purpose. Take ten real cases, clear the ship date in a copy, and read what the agent says. If any of the ten offers a day, the job is not ready. Do this before the first night it runs alone. A happy-path demo, with every field filled, will not show you this failure. The failure lives in the blank.
Escalation is part of the job
Escalation is part of the job, not a courtesy the agent may offer. Salesforce says agents can pass complex issues to people so the query still gets resolved. Write three lines before you widen anything: what "beyond scope" means, who receives the case, and what the customer is told while they wait. If those lines are blank, the agent is not ready for the easy intents either. The easy path and the exit are one design.
Beyond scope, written so a supervisor can apply it at a glance:
- The field the answer needs is empty.
- The customer asks for money, a date you have not stored, a legal outcome, or a change to identity.
- The customer is angry, abusive, or cannot be matched to a record.
- The agent would have to change a field that is not on the write list.
- The same customer has already been escalated once on this case.
Who receives it is a queue with a name, an owner, and hours a person is actually present. "The service team" is not a queue. A queue that nobody watches between Friday night and Monday morning is a place the case goes to disappear. If the agent runs at 2am and the queue does not, the agent may collect and stop. It may not close the loop with a sentence the morning shift cannot stand behind.
What the customer is told is a sentence you wrote, not a sentence the model improvises while it waits. "I do not have a ship date on this order. A person will look at the case when the queue opens at 8am." That sentence is dull, and it is safe. "It should be with you Friday, and someone will confirm" is the 2am failure wearing a handoff. Do not let the waiting message carry the fact you just refused to let the agent invent.
The handoff record carries the reason code, the field that was empty or the line that was crossed, and the reply that was not sent. The person on the queue should not have to reconstruct the conversation to find out why it stopped. If they do, they will start editing sent replies instead of using the queue, and you will lose the count that tells you the job is too wide.
What is still speculation
No one can give you the year the contact centre closes. Some easy intents may leave the queue. The leftover work is the odd, the angry, and the expensive, and it does not get smaller because the easy intents left. Often it gets harder, because the simple status calls were what made the hour feel manageable.
A headcount cut tied to a guessed year produces a queue and nobody on it. That is the operational version of the 2am failure. The agent is still sending, or still stopping, and the person who was supposed to take the stop is a cell in a spreadsheet that was reduced in advance. Plan the bounded jobs you can name this quarter. Keep the people who handle what the agent must refuse. Review the split when you have counts from your own cases, not from a keynote.
Count two things, and only from your cases. How often the agent finishes the bounded job with an answer that matches the record. How often a person has to repair a sentence that was already sent. The first number tells you the job is real. The second tells you it has already spilled. A rise in the second number is the moment you narrow, even if the first number looks healthy. Finished-and-wrong is not a success that needs a better prompt. It is a job whose line was crossed and whose send was allowed.
Do not convert those counts into a story about the centre closing. A month in which status calls drop is a month in which status calls dropped. It is not evidence about the angry cases, the billing disputes, or the customers who cannot be identified. Those remain, and they take longer once the easy calls are gone, because the hour no longer contains a run of quick wins.
What to write down before you widen the scope
Write the job in one sentence a supervisor would recognise. Write the records it may read and the fields it may change. Write the line it must not cross: money, legal process, identity, a promise about a date or a refund. Write the queue, including the hours a person is actually there.
Then run it, and read the escalations every week. If people are quietly correcting answers the agent already sent, the job is wider than the sentence. Narrow it. A longer prompt is not a wider licence to act. A second bounded job, with its own line and its own queue, is the way you add scope.
A second job is a new sentence. "Tell the customer the status of an order already on the case" is the first. "Book a replacement delivery when the order is marked failed and the entitlement allows one" is a second, and only after you can point at the entitlement field, the failed status, and the person who handles a refused replacement. Folding the second into the prompt of the first is how Friday gets promised inside a status call. The prompt gets longer. The line gets blurrier. The repair count goes up, and someone calls it a tuning problem.
The weekly read is short. Twenty escalations, or all of them if there are fewer. For each: was the stop correct, did the customer hear the sentence you wrote, and did a person have to repair anything already sent. Bring the owner of the queue. If the same stop is firing for a reason you did not list, add the reason or narrow the start. If repairs are clustering on one field, take that field out of what the agent is allowed to say until a person has checked it.
Who owns the record the agent reads is not settled by the agent. If the ship date is written by the warehouse system and copied into the case, the copy can be stale, and the agent will quote the stale copy with confidence. Which system owns that date, and how old the copy may be before it is unfit to say aloud, belongs with the programme decisions in integration at programme level. Put the allowed age on the source line. A status job that quotes a date older than that age stops, the same way it stops when the field is empty.
The first month
Week one is the sentence, the records, the fields that may change, and the line that must not be crossed. Week two is the queue, the hours, and the sentence the customer hears while they wait. Week three is the empty-field test on a copy of real cases, including the blank ship date. Week four is the job running on one intent, with the weekly read in the calendar.
Only then do you talk about a second intent. The contact centre is still there. It is holding the work the agent was told to refuse. That is the plan. A year with a closing date on it is a guess someone else can make. It is not a roster you should publish.
