Support chat
In support chat you describe a problem in your own words, or add a screenshot of the error, and the Support Assistant looks for similar solved and open incidents and for runbook steps, then walks you through a fix. Your recent chats are listed.
Chat is for people who may run agents: platform administrators, agent builders and operators. Otherwise the chat page says it needs run access.
Start a chat
Open Chat in the top bar (in the menu on a phone), type the problem and press Enter. Shift+Enter starts a new line. Each message can be up to about 4,000 characters.
While an answer is on its way, the chat lists what it is doing: searching your documents, checking past incidents, asking an advisor model when it consults one, and writing the answer, each with its time. When the answer starts, the list folds into one summary line with the time it took. The answer is then written word by word, each word fading in. If your device asks for reduced motion, nothing moves, pulses or fades.
Try an example fills in one of six synthetic error screens and a question. Nothing is sent until you press Send.
Screenshots
You can add up to 5 screenshots to one message, as PNG, JPEG or WebP files of up to 5 MB each: use the screenshot button, paste or drop them.
Use synthetic screens only. Never upload screens from company systems.
A first agent reads the screenshot before the answer is written: the message text and ID, the transaction, the program or job and the screen title. It is told to leave out personal names, user IDs and email addresses, but a screenshot can still show them, and nothing removes them automatically. Check a screenshot before you add it.
Above each answer, What I read from your screenshot shows what was taken from the image, or that it could not be read. A later message can reuse the last screenshot that was read.
What an answer contains
New problems get five parts:
- What's going on: what the message or symptom means, in one to three sentences.
- Seen before: past incidents and runbook passages that match.
- Already reported?: open incidents that may be the same problem.
- Steps: what to do next, numbered.
- If it still fails: what to collect when the steps do not help.
Sources sit under each answer in a closed list; select a marker such as [1] in the answer to open its source.
"Possibly already reported" means an open incident that may describe the same problem. It is a suggestion, not a match: read the incident and decide yourself.
The answer is a starting point for your own checks. Read each step before you act on it.
Hard questions may take longer, because the assistant can ask an advisor model for a second opinion before it answers.
Checked answers
An administrator can turn on a check of real model answers. The demo site runs with the check off, so its answers show none of what follows.
While a checked answer is written, the chat lists each step as it starts, such as Checking the answer, then its time. No draft is shown.
A checked answer ends with Confidence and a percentage: how much of the answer comes from your documents, and how much of what they say it covers. Select it for this explanation. Under 90%, an advisor model reviews it first, and Reviewed by an advisor model is shown.
A dotted underline tagged general SAP knowledge marks a point with no document behind it.
Plain words or analyst
Answer depth at the top of a chat sets how the next answer is written. Plain words uses everyday language, at most five steps, and names a transaction only when you need to open it. Analyst names transactions, programs, jobs and tables, ties each step to a runbook, and labels causes as hypotheses. A change applies to the next answer only. Explain that again repeats the last answer at the new depth.
Answer mode
People with real model answers can choose Auto, Fast, Balanced or Thorough under Answer mode in the message box. Auto picks for you and uses the same fast model as Fast, which is "Quickest". Balanced is "More careful" and Thorough is "Most careful; slower". The selected option is the dark one. The choice is remembered on that browser.
Follow-up messages
Under the latest answer, quick replies fill the message box: That fixed it, A step failed, I got a different error and Explain that again. None of them sends by itself.
Each answer uses the last three exchanges of the chat as context. New chat starts fresh. A chat that grows too long has to be started again.
Message actions
Under each message you sent are its date and three buttons. Retry sends it again, except a message with screenshots, which cannot be resent. Edit puts it in the message box; nothing is sent until you press Send. Copy copies it.
Under an answer, Copy copies it with its markers and a list of its sources. Read aloud reads it in your browser's own voice; nothing is sent anywhere. Retry under the latest answer sends your message again. Retry and Edit always add a new message; they never replace or remove one.
Rating an answer and opening its sources
Under a real answer, Helpful and Not helpful rate it. After Not helpful, pick what was wrong, add a note if you like, and press Send. Your rating goes to the team that maintains the knowledge packs. A note that looks like it contains sensitive data is not saved.
In the sources list, the title of a source the answer cites opens that document, read-only, with the passage the answer used. Only you can open it.
Simulated answers
Real model answers are open to accounts on the real-model list. On other accounts the chat says that answers are simulated: screenshots are not read, and answers are placeholders.
Raising a ticket
After an answer, Raise a ticket opens an editable draft built from the chat. Nothing is sent until you choose Create ticket. The ticket is created in ServiceNow with urgency 3 and impact 3, and your confirmation is recorded as the approval. When the ticket is created, the chat shows its number; when it is still being created, the chat says so. Text that looks like an email address or phone number is refused; remove it and try again.