AI Assistant

The CMS is in no language model’s training data. An assistant answering about it “off the top of its head” would be inventing something that looks like documentation — which is why nothing here answers without reading first: the embedded manual when the question is about the product, your installation when the question is about your installation.

Where it lives

The robot icon in the header, to the left of the bell, on every screen. It opens a side panel, not a modal: the question is almost always about the screen behind it, and a panel lets you read the answer with your hand on the setting it explains. The panel also survives navigation — opening it on Applications and walking over to Transformers before typing means asking about Transformers.

When the icon turns amber with a red counter, Standby investigated something on its own and nobody has closed the case — the click lands straight on the Investigate tab. With nothing pending, it is the same “Ask the CMS” as always.

The three tabs

A tab you have no access to stays on screen — it shows the reason, and the reason changes where it sends you:

Hiding the tabs would erase all three cases at once. The person who runs into the tab at the moment they need it is the one who best understands what it is for.

Ask — help from the manual

The Ask tab answers about the product by reading this manual, which travels embedded inside the API. It works without internet and without the paid module.

The answer always comes with Sources in the manual — the list of pages the assistant actually opened. That list does not come from the answer text (that would be the model’s own claim), it comes from the session steps: a record of what was read, not of what the model says it read. With no source, the answer is treated as “not found”.

The footer shows the cost — steps and tokens — and the date of the embedded manual.

How it uses your screen. The route you are on travels with the question, so the assistant starts at the right manual page instead of searching for it. Asking “and this field here?” while on /messages/delivery-errors works.

Suggestions and history. Before the first question the panel offers four starting questions. After that it keeps your recent questions, with a clear button.

Reused answers

In a plant, “how do I do X” gets asked by several people. When an answer came only from the manual, it is stored and reused — the session shows zero steps, zero tokens and a reused answer mark, and the sources are still listed.

The rule for what gets cached is about security, not performance: a single lookup against the installation takes the answer out of the cache. From that point on it depends on the state of this plant and on the slice that user can see — storing it would hand one person’s answer to another.

The cache invalidates itself when the documentation changes: the manual version is part of the key.

Investigate — a report backed by evidence

Paid feature. You describe what you want to understand in plain language, and the agent runs real read tools against the installation, observes the result, corrects course and delivers a report.

Ready-made examples offered by the screen:

• Interface [CODE] stopped processing. Find out since when, whether it is blocked and why.
• Analyse the delivery errors of the last [24] hours: which definitions they come from, whether there is a common cause.
• Which applications are offline right now? For each one, what to check first.
• A message sent to [CODE] on [DATE] never arrived. Trace where it stopped.

The timeline shows every step: the tool called, the arguments, the result and the duration, each one expandable. The marker is green when the tool ran, red when it failed and neutral for a plain comment — in a ten-step investigation, the colour is what lets you find where it stalled without opening them one by one.

The report

The report comes out in three blocks, not as a single ten-line paragraph: Cause, Evidence and What to do. What the person opened the panel to read is the reason, and it used to be buried in the middle of the text.

Your latest investigations

Below the goal box, the tab lists what this person has already investigated. Reopening a report from there spends no step and no token: the text is already stored. Without that list, closing the panel wiped the investigation from sight, and redoing it cost the whole session again.

Configure — a proposal that needs approval

Also paid. Here the goal is to build the integration: “in application [CODE], receive production orders from the ERP and deliver them to the MES”. The agent reads the installation, discovers the real format (probes the URL, browses the database catalogue, reads the SAP function parameters, tests the transformation against traffic that already went through) and assembles a plan.

The agent never writes to the database. It proposes; the write is a person clicking apply, through the same installer used by Packs — with idempotency, a trail and auditing.

Evidence

Every test that succeeded becomes evidence attached to the proposal, listed above the apply button. When there is none, the screen says in plain words that nothing was tested in this proposal — approving a plan that was never tested has to look different from approving one that ran against real data.

The traffic test declares its source, each with its own colour:

Approving

A proposed plan opens the full agent screen, at /agent — approval does not happen inside the 400 px side panel, because approving means writing to the database and that is not a width for reading before deciding. The screen shows the plan item by item (what will be created, what already exists), the evidence, and the apply button.

When the session ends without a plan

The state is honest about what happened:

Ending a build session without a plan is Partial, never success.

Screen shortcuts

Asking “find the message containing invoice NF-12345” does not return a message written by the AI. It returns a button, below the answer, that opens Content Search already filtered by that term. The CMS is what searches, with your permission, exactly as if you had filled in the filter by hand.

This is not navigation convenience — it is what separates an answer you can check from one you have to believe:

• Message content never reaches the AI provider. What the model assembles is the address, not the result.
• The count is exact and costs no context. A filter yielding 7,000 rows costs the same as one yielding 7 — the agent says “I found 12” without having read a single message.
• The result is live. It opens the usual screen, with current numbers, and you adjust the filter you were handed instead of asking again.
• Permission still belongs to the backend. The screen checks everything again when you click. A badly built shortcut leaks nothing: at worst, it opens empty.

A period that follows the clock

For requests relative to now — “the messages that failed in the last hour”, “what came in today” — the button carries the period shortcut, not two dates. Clicked tomorrow, it opens tomorrow’s last hour, not the one from when you asked. For a fixed window (“September 15”), it uses dates.

Exporting

“Export today’s messages from SAP” returns a button that opens the filtered screen and fires its Export to Excel. It is the screen’s own button, on the same slice: the spreadsheet is built in your browser, and no file holding production data is ever written to the server.

Where it can take you

The list is closed, and each screen declares what it actually honours. A filter the screen ignores is refused on the spot, with the list of what it accepts — accepting it silently would open the page without the slice, and the button would have promised a result you are not looking at.

When it refuses

Why the button never comes from the answer text

The agent reads messages, and the content of a message is text that came from outside. If the screen turned whatever the model wrote in its report into a link, a crafted payload could induce a phishing address rendered with the credibility of the CMS interface.

So the button is not found in the text: it comes from a field of its own on the session, written when the tool ran, and the destination comes from a closed list of screens — checked on the server and again in the browser. That is also why the answer mentions the button instead of repeating the address.

Standby

When an alert fires, the agent can investigate on its own and leave the report ready for whoever received the alert. Enabled in Settings → AI, and off by default: it is the only part of the product that spends your AI key without anyone clicking.

It sweeps fired alerts every 5 minutes and investigates the ones worth investigating: blocked queue, piled-up messages, delivery or collection errors, offline application, lost connection.

Three locks, all explicit:

1. the Agent module in the licence;
2. the switch in Settings → AI;
3. an alert recipient with access to the agent — the investigation runs with their permission, and the session shows up on their screen.

Plus: a one-hour silence window per target (a queue that stays blocked fires an alert every cycle), a ceiling of 3 sessions per sweep, and a failure on one alert does not stop the others.

The silence window is stored on the investigated session itself, not in memory. On an installation with an expired licence, which restarts every half hour, an in-memory cooldown vanished at every restart and became half an hour in practice: the same blocked queue paid for a fresh report saying what the previous one already said.

The report reaches you

Before, the agent investigated on its own and stored the result without telling anyone — whoever did not know to open /agent and scroll to the previous sessions never found out an investigation had happened. Now a Standby session is born unread and shows up in three places:

On the full /agent screen, a session opened by Standby is labelled Standby instead of by mode — it is the only row in the list the person does not recognise from having typed it.

Reading the report does not clear the header counter: opening marks it as seen, but the one who closes the case is you, by ticking resolved — the problem may well still be standing after it is read. One resolved by mistake can be reopened, and there is a mark all as resolved for the day the sweep caught a run of alerts from the same source.

What it can never do

These are not recommendations — they are properties of the design.

• The agent can never do more than whoever called it. The tools handed to the model are filtered by the session user’s permission: the same Applications and Interfaces you see on the screens, not one more.
• It never sees the message body. It sees metadata and structure; the payload is out of reach.
• Secrets are redacted before they leave. Passwords, tokens and keys are stripped from any tool result before reaching the model and before being written to the timeline.
• A session is private to whoever created it. The report carries the slice that person could see; showing it to another profile would leak exactly what the filter prevents.
• No open internet. Only the controlled URL probe the Integration Assistant already uses — GET only, with timeout, size ceiling and cloud metadata addresses blocked.
• The question is recorded, redacted and never leaves. It goes to the history the way the agent’s goal does — auditable — and the only external destination is the AI provider you configured.

Cost and limits

Every session has a budget of steps, tokens and time, and the panel shows the consumption in the footer. Running out of budget ends the session as Partial with the reason, never as Completed.

The model is the one you configured in Settings → AI — including an Ollama on your own network, for those who cannot send anything outside.

Turning it on and off

Each piece has its own switch in Settings → AI, and all of them depend on the provider master switch: Ask the CMS, Investigate, Configure and Standby. Turning one off does not affect the others; turning the master switch off turns everything off.