ChatGPT
Claude
Google AI
Copilot
The benefit of generative AI in enterprise CRM is a finished job. Four jobs cover almost everything worth turning on: a case summary, a draft reply, qualification, and booking a meeting. Name the job. Staff the exception. Leave the rest off until those four are dull and reliable.
The failure that shows the difference is ordinary. A customer writes three sentences about a late delivery. The summary the next agent reads adds a fourth: that the customer asked for a refund and was promised a call by Friday. Neither sentence is in the case. The agent calls. The customer says that never happened. The case now contains a false step that looks like history, because someone acted on it. Clean sentences hid it until a person repeated it out loud.
Salesforce describes Agentforce as the agent platform that delivers 24/7 autonomous support at enterprise scale. The page on how agents work says agents generate responses from trusted business data, including Salesforce CRM data and external data from Data 360, and that those responses stay inside your brand rules. Data Cloud was renamed Data 360 on 14 October 2025. The practical reading is short. The model has to see the record you mean. Tone comes after that. A fluent sentence with no home on the record is still a false step. A chat window that can talk about customers in general can be polite at 2am and still have nothing you can hand to the next person.
What a finished job is
A finished job has a start, an output, and a stop. The start is a record that already exists. The output is something a colleague can use without redoing the reading. The stop is a rule that says the job is done, or that it must go to a person. A summary on the case, a draft in the reply box, a qualification note on the lead, a meeting on a calendar: those are outputs. Each one can be wrong in a way a supervisor can see by reading the source beside it.
The four jobs rise in consequence, and that is the order to turn them on. A bad summary misleads the next person, who can still check the thread. A bad draft can be sent. A bad qualification can send a rep after the wrong account for a week. A bad booking puts two people in a conversation who should not be there. Turn on the next job only when the previous one is dull. Dull means a supervisor can sample ten and not find a sentence that is absent from the source.
Staff in that same order. The summary needs a person for the odd case. The draft needs a person who sends. Qualification needs a person who decides to work the record or to close it. Booking needs a person, or a rule written in advance, who accepts the time. If that person is not named, the job is a demo, and the odd case has nowhere to go.
The case summary
A case summary is for the next person who opens the case. It says the status, the last thing that happened, and the question still open. If it adds a step the case does not contain, throw it out.
Write the contents as a list the summary is allowed to fill. Status, as the case field has it. The last customer message, meaning what they asked, not what a helpful reading thinks they meant. The last action a person took, if one is on the case. The question still open, which may be that none was stated. If the thread does not contain a next step, the summary says the next step is not on the case. It does not invent a callback, a refund, a part number, or a date.
Close the record and ask whether you could still have written the summary. If you could, it is not grounded. Check each sentence against the case, the comments, and the latest email. A sentence with no home is deleted. Wrapping a second sentence around it, so the first sounds more careful, keeps the false step and makes it look reviewed.
Two failures show up in the first week if anyone looks. The first is the added step, as in the late delivery that grew a refund. The second is the dropped constraint. The customer said they are travelling until Thursday and asked that nobody call before then. The summary reports the late delivery and omits the instruction, because it sat in the third paragraph and a tidy plot was easier. The next agent calls on Tuesday. The summary was short, and the wrongness was an omission a skim will not catch.
The draft
A draft is a reply someone can send. It follows the order, the entitlement, or the complaint on the record. Brand voice keeps the wording inside your rules. It does not prove a refund is allowed, a price is current, or a date was promised.
Salesforce says the responses stay consistent with brand voice and guidelines, using trusted business data. Read that as two fences, in that order. The first is whether the fact is on the record. The second is whether the wording sits inside the rules you gave it. A polite sentence that offers a refund the entitlement does not allow is a worse failure than a blunt sentence that refuses it. People send polite sentences. They trust the manners.
The person who sends the draft is the control. The model prepares. The person sends. If the send happens with no person, the job has changed, and the handoff rules in autonomous agents in customer service have to be in place first. A draft that leaves on its own is a message to a customer.
Delete a draft that states a price, a discount, a refund, a replacement, or a date nobody authorised and that the record does not contain. Do not polish it. Polishing keeps the false fact and improves the manners around it. Discard the draft and send the case to the queue, with the discarded sentence attached.
Articles you have allowed are part of the source. An article that describes last year's returns policy will be quoted as if it were this customer's entitlement. Put the allowed articles on the source line, with a date or a version. An article off that list is not a source.
Qualification
Qualification lines up fit, timing, and the missing fact on a lead or an opportunity. The model can prepare that from fields you trust. A person decides to work the record or to close it. A score with no line for the missing fact is a decoration.
The output is three lines, not a number. Fit: which fields say this account matches the business you sell to. Timing: which field says there is a reason to talk this month. Missing: the field that is empty, named, so the rep knows what they still have to learn. If every gap was filled with a plausible guess, the note is fiction, and it will be believed because it reads better than the truth.
The failure here is confidence. The model does not convert the lead because the note sounded sure. Conversion changes what the pipeline claims, and that decision sits outside this job. The lead has a company name and a generic inbox. The note says the account is a strong fit and ready this quarter. None of those words are on the lead. The rep believes the note, spends a week, and then finds the domain belongs to a student group. The note looked like work already done. A missing-fact line that names the empty employee count, the empty industry, and the absence of any meeting takes ten seconds to read. The score, if you still want one, sits under those three lines. It does not replace them.
Booking a meeting
Booking a meeting writes to a calendar. The model may propose a time from the thread and the free slots. The booking stands when a person accepts it, or when a rule you wrote in advance accepts it. A proposed time is not a meeting.
Write the rule on the same screen as the button. One sound rule is that the rep confirms every booking. Another is that the model may book only with a named contact, only inside hours you have listed, and only when the opportunity already has an owner. A button with no rule fills calendars with the wrong people, and the reps stop offering it.
The loud failure is the wrong person. The thread mentioned a colleague in procurement, copied for information. The model booked that colleague. The buyer, who was never asked, hears about it afterwards. The source line is the contact you are allowed to invite, and the calendar connection. Someone mentioned on the thread is not a contact. A free slot is not consent to be invited.
A hold is not a meeting. If the model may place a hold and the rep must confirm, write when the hold expires, tell the rep, and leave the opportunity untouched.
Where the words have to come from
Write one line for each job: which records it may read. Summary reads the case, the comments, and the latest email. The draft reads those and the articles you have allowed. Qualification reads the lead or the opportunity, and the account. Booking reads the contact and the calendar connection.
If a field comes from Data 360 rather than the CRM record, name it on that line. An answer you could have written with the record closed is not grounded. Do not turn the job on until the line is true in the org, not only on a diagram. A diagram can show a connector that the permission set does not actually grant. The test is to open the record as the user the job runs as, and to see the field.
The line is an artefact a person can read, not only a sentence inside a prompt. Put it above the summary, in the header of the draft, as the first line of the qualification note, and next to the booking rule. A plausible detail the model supplies from habit, absent from that line, is deleted. The people who can see the line are the ones who can take the job off.
Data 360 is a source only when the field is named. A connection on its own is not a field. A consent flag, a product holding, an external order status: each is on the line, or the job may not use it. Write the staleness that job accepts. A summary may quote yesterday's order status. A draft may not quote yesterday's price. Who owns that field, and which copy wins, is the programme decision in integration at programme level. The job inherits it.
The exception is part of the design
Each job needs a person, and the person needs a queue, before the model needs a prompt.
Legal language, abuse, or a caller you cannot identify stops the summary and goes to that queue. A draft that states a price, a refund, or a date nobody authorised is discarded, and the case goes with it. Qualification that sounds sure while a required field is empty goes back to the owner of the record. A meeting with the wrong contact is not booked. Those four stops are the design. A prompt that says to be careful is not a substitute for them.
Name the queue before go-live, and name the hours a person is actually there. A queue that is empty from Friday night until Monday morning sends the odd case into a void. The next agent then reads an unchecked summary, or a customer receives a draft that was allowed to send itself. If you do not have the hours, leave the draft in the box.
The handoff is a record. The case arrives with the reason it stopped, the source line, and the sentence the model was about to use. If the handoff is only a transcript, people correct answers quietly and skip the queue. Quiet corrections mean the job is wider than the sentence you wrote. Narrow the job. A longer prompt is not a wider licence to act.
The figure you are allowed to keep
Time the job against a person doing the same record. Use your own cases. Keep two numbers together: how often the output matched the source on the sample you read, and how the remaining person-time compared with the same records done without the job. A summary that saves a few minutes, measured on a week of your cases, with the false-step count beside it, is a figure you can defend. If you only keep the time, you will scale a job that is fast and wrong. A percentage copied from a vendor story is not your result, and it will be used to cut a queue you still need.
If the four jobs are funded and nothing is on, use a fortnight. In the first week, write the four sentences and the four source lines with the people who do the work. Include every Data 360 field by name, or leave it off. In the second week, name the queue, the hours, and the stops. Turn on the summary only. Read twenty against the thread. The other three jobs wait until that one is dull. The connector, the prompt, and the brand voice are the smaller pieces. They are worth doing after the source line is true.
