AI Assistant

The AI Assistant is built into the hub. You ask a question in your own words, for example "Which devices are not communicating, and why?", and the AI Assistant answers from the plant data that you are allowed to read. It is in Monitor > AI Assistant, and on every screen as a side panel.

The AI Assistant needs an AI service before it can answer. An administrator adds one in Platform > AI Assistant Settings. See Set up the AI Assistant.

Where to open it

The side panel knows the screen you are on. When you open it on a record, for example a device's page, "this device" in your question means that device.

1. Your question.
2. The sources the AI Assistant read for this answer.
3. The answer, with the AI service, the model and the time under it.
4. The AI service that answers. You can change it between questions.

What the AI Assistant reads

The AI Assistant answers from the hub, not from memory. For each question it reads what it needs, and the answer shows a label for every source it used.

Alarm history, alarm counts, the shift and weekly summaries and recurring problems count a period the same way: an alarm belongs to a period when it was raised in it or fired again in it. An alarm raised three days ago that fired again this morning is one of this morning's alarms, so the numbers agree whichever way you ask.

Ask for a calendar period in your own words, such as "yesterday", "last week" or "the morning shift on Monday": the AI Assistant works out its start and end in your time zone and says which period it used. "In the last 7 days" is a rolling window that ends now.

When the hub itself cannot hear the devices (it is not connected to its message bus, a live feed keeps failing, or it started a few minutes ago and devices are still reporting in), the assistant says so instead of calling the devices offline.

The AI Assistant reads with your permissions. If your role cannot see a device, the assistant cannot see it either, and it says that something may be outside your permissions rather than that it does not exist. A role limited to part of the plant (see Users and roles) limits the assistant the same way: its devices, alarms, history and records, and the logs of its own edge gateways only.

Passwords, keys, tokens and other secrets are never read, nor are the license details, raw telemetry rows or the connection settings of profiles, channels and targets.

How each answer is checked

Before an answer is final, Proxus compares every number, time and date in it with what the AI Assistant read for that answer: the sources above, your question and the earlier answers of the conversation. Times are compared on your clock. The check is plain matching, not another AI, so it costs nothing and gives the same verdict every time.

• When a value has no source, the AI Assistant is told which one and writes the answer again: it reads the value, corrects it or leaves it out. While it does, the panel shows Checking the answer against what was read, and the new answer replaces the first one with the note Checked against what was read: the AI Assistant corrected its answer.
• When a value still has no source after that, the answer stays and a warning under it lists the value: Check these values yourself.

Words such as "about two days" are not checked, nor are numbers that are part of a name (Press 4, S7-1500), list numbers, code and links. A rewrite takes a few seconds and uses tokens like a short question.

Changes you approve

The AI Assistant can prepare changes, but it never makes one on its own. Ask in plain language, for example "Disable Pump 2" or "Resolve alarm 1204, we replaced the sensor". The assistant shows a proposal card under its answer:

• what changes: each field with its value now and after the change (for a new record, its values);
• what depends on the record, when you switch it off or delete it, the same list the change guard shows;
• why it cannot be done, when your role does not allow it or a rule of the record is not met.

1. What changes: each field, now and after.
2. What depends on the record, when it is switched off or deleted.
3. Your decision.

Select Approve and apply (for a delete: Approve and delete) to make the change, or Reject. Only the person who asked can decide, once.

The AI Assistant never sets passwords, keys or other secrets, never changes an administrator, an administrator role or your own account, and never changes what the platform writes itself (deployments, statistics, logs).

When you approve, the change is made again with your permissions and the record's own rules, exactly as if you had saved it on its screen. If the record changed after the AI Assistant proposed, nothing is written and the card says so: ask again for a new proposal. The Audit Trail records the change under your name, with "AI Assistant · conversation N" as its source (changes made before the rename show "Copilot · conversation N").

Add a machine

Ask, for example, "There is an OPC UA server on my computer: add it as a device and read its simulation values". The AI Assistant uses the device wizard's own buttons, on the edge gateway that will read the machine:

1. It finds the server (Discover servers, for OPC UA and MTConnect) or, for other protocols, tries the address (Test the connection). Discovery looks at the edge gateway's own machine and the address you give; ask it to search the whole network of the gateway when the server is elsewhere (slower). These only read and need no approval.
2. For OPC UA it opens the server's folders and shows the current values of the variables you asked for.
3. It proposes the device on one card: the edge gateway, the protocol, the connection parameters and the data points. Approve and apply saves it as the wizard's Save and verify does: the profile, then the device, switched on. The device starts reading at once, and Undo removes the device and its profile. One card takes up to 50 data points; ask for the rest afterwards.

Data point names must start with a letter and use only letters, digits and _ (for example DischargePressure_raw), because pipelines, functions and rules read a value by its name. The AI Assistant keeps the unit or the scale in its answer.

On a secure OPC UA endpoint (Sign or Sign and Encrypt), the gateway and the server must trust each other first. When the gateway does not trust the server yet, the AI Assistant proposes to trust its certificate: the card shows the certificate and its SHA-1 fingerprint, and you confirm twice. This needs the Ask, propose changes and run actions level. The server must also trust the gateway's certificate in its own settings, and the assistant gives you that fingerprint. A trusted certificate lets that server talk to every edge gateway, and trusting cannot be undone from the assistant: to withdraw it, reject the certificate in the Certificate Store. A server that needs a user name and password is added in the device wizard, because the assistant never handles passwords.

The AI Assistant offers this to roles with Ask and propose changes (or more) whose role may also create devices and device profiles.

Build a pipeline

Ask, for example, "On the mixer, divide HR01 by 10 and keep it as a new tag HR01_scaled". The AI Assistant builds the pipeline the way the pipeline builder does: it picks the goal and steps from the builder's own list, fills in When, For, Steps and Then, and lets the builder check it. It tells you what the builder's preview shows on each device's latest values, then proposes the pipeline on one card.

The AI Assistant designs the alarms as an engineer would: a noisy value gets a dead-band limit, and a condition that depends on a machine running gets a lasting time, so a start-up surge raises no alarm. To change a pipeline built this way, ask the assistant to change it: it reads the pipeline and proposes the new version on one card, each part shown before and after. The change goes to the draft and can be undone; the running revision stays until the pipeline is published again.

Approve and apply saves it as a draft, with its story, so it opens in the builder as if you had built it there and you can change it. A draft does not run. The AI Assistant then offers to publish it in Shadow, and only after that Active; each is a separate card, and Active asks you to confirm twice. Undo deletes a draft that was never published.

A pipeline has up to 12 steps and results. The AI Assistant offers this to roles with Ask and propose changes (or more) whose role may also create pipelines. Publishing is an action, so it needs Ask, propose changes and run actions and the capability of each publish button: Try pipelines for Shadow, Write to devices for Active (see Actions).

Build a form

Ask, for example, "Make a form for operators to report a line stop: the reason, how many minutes, and a note when it was longer than half an hour". The AI Assistant reads the form components, the ready-made templates, the shared choice lists and the roles of this hub, writes the whole form in English and Turkish, and lets the form designer's checks look at it. The card shows how the form is laid out, its sections and fields, its rules, and who fills it in and approves; an error the designer would refuse stops the card, a warning goes with it.

Approve and apply saves the form as a draft that opens in Organize > Forms as if you had designed it there. Nobody can fill in a draft: publish it in the designer when it is right. Undo deletes the draft.

To change a form, ask the AI Assistant to change it: it reads the draft and proposes the new version on one card that lists the fields added, changed and removed and the settings that change. The change goes to the draft and can be undone; operators keep filling in the published version until you publish again.

The AI Assistant offers this to roles with Ask and propose changes (or more) whose role may also create forms. Every role can ask about the forms and records it may open.

Confirming twice

Some changes ask you to confirm a second time, after Approve and apply. The card says why:

• something is deleted (a delete cannot be undone);
• a record is switched off and something depends on it (the card lists what);
• five or more records change at once;
• the change touches who may do what, or the whole platform (users, roles, UNS permissions, certificates, settings, brand settings, AI Assistant services and their allowed roles).

Actions

Roles with Ask, propose changes and run actions (see Set up the AI Assistant) can also have the assistant press buttons for them, on the same approval card. Each needs the button's capability, exactly as on its screen:

Approve and run presses the button with your permissions, and the card shows its result. An action cannot be undone. Actions that reach the plant or the gateways (Publish Active, roll back, deploy, undeploy, templates, revoke, certificates) ask you to confirm a second time. Buttons that need a file, a designer or a person's own decision (export, import, the gateway terminal, passwords, the license) are never pressed by the AI Assistant.

Undo

An applied change that created or changed records has an Undo button on its card. Undo deletes what the change created and sets the fields it changed back to their earlier values, in one step, again with your permissions. If someone changed one of those fields after the AI Assistant did, nothing is undone and the card says so. Deletes and alarm decisions cannot be undone; the card says this before you approve.

Proposals expire

A proposal nobody approved within 24 hours expires: the plant may have changed since. Its card shows Expired. Ask again for a new proposal.

Insights

The Insights tab next to Chat lists things that need a look, found by background checks without anyone asking. Everyone who may use the AI Assistant sees the insights about what they can see: a role limited to one site sees the insights about its own devices and gateways. An insight that counts across the whole plant (open alarms without an owner) is shown to roles that see the whole plant. The checks are fixed rules: no AI service is used and no data leaves the hub until you select Ask AI Assistant on an insight.

1. The level and what was found.
2. The suggested next step.
3. When it was first and last seen, and what you can do with it.

Each insight is a warning, or critical when it is urgent: a rule raised its alarm 50 times or more in a day, one of the unowned alarms is critical, a gateway lost its link 10 times or more, a certificate has 7 days or less, or a service spent its whole budget. A notification channel can send only the critical ones (see below).

At most 15 frozen and 15 unusual values are listed at a time, the most telling first.

Alarms, gateway links and certificates are checked every 15 minutes, tag values every hour; the first check runs 3 minutes after the hub starts. Check now runs the checks at once. An insight closes by itself when its condition is gone, and closed insights are removed after 30 days.

Each insight says why it appeared and suggests a next step:

• Ask AI Assistant starts a conversation about it. The AI Assistant looks at the record and can prepare a change for you to approve.
• Snooze 1 day hides it until tomorrow. It comes back if the condition is still there.
• Dismiss hides it for as long as the condition lasts.

Snoozing and dismissing hide the insight for everyone, so only administrators have these buttons and Check now.

When insights are waiting, the AI Assistant button at the top right shows a dot, and its tooltip gives the count.

The AI Assistant going first

• Guided setup: ask to connect a first machine, create a first alarm rule or a first dashboard. The AI Assistant checks where you stand (is an edge gateway connected, is there a profile, does data arrive…), names the next step with its screen and docs page, and prepares the records for you to approve.
• Shift and week: "How was the night shift?" gives the last 8 hours (or the last week): silent devices, disconnected gateways, alarms against the period before, the noisiest rules and devices. The same summary can be e-mailed on a schedule: see Scheduled deliveries.
• Recurring problems: "What keeps going wrong?" finds the same alarm returning on the same device, edge gateways whose link keeps dropping, and insights that reopen, and suggests a lasting fix (for example a wait before the alarm, tried on past data first) instead of acknowledging the same alarm every day.
• Insights to a channel: new insights can be sent to an e-mail, Teams or other channel: see Notification channels.

What it does not do

It does not open dashboards or run reports, and it does not query the telemetry database with SQL. It reads up to 90 days back; for longer periods, use the Trend Workspace.

Conversations

• New conversation starts an empty one. The first question becomes its title.
• Select a conversation to continue it. The AI Assistant remembers the questions and answers in it, up to the last 40 messages; start a new conversation for a new subject. It reads the plant data again for every question, so answers use current values.
• The list at the top of the conversation chooses the AI service that answers, from the services the administrator enabled. You can change it between questions.
• Delete conversation removes it after you confirm.
• Stop ends an answer that is still being written.

Conversations are stored in the hub until someone deletes them. Only you can continue or delete yours. Administrators can read every conversation with Show everyone's conversations, but cannot write in someone else's. The AI Assistant shows this note next to the conversations.

What is stored: your questions, the answers, the name of each source the AI Assistant used, each proposal with your decision, and any error. The data the assistant read is not stored.

Limits

The AI Assistant keeps what it reads short, so that answers stay fast and cheap:

Language and time

The AI Assistant answers in the language of the hub's interface unless you write in another language. Times are shown in your browser's time zone.

When there is no answer

The error stays in the conversation, so you can see later why a question had no answer.

Safety

Device names, alarm texts and other plant data were typed by people or sent by machines. The AI Assistant treats them as data and does not follow instructions written inside them. Answers are shown without HTML, and only web links (http and https) stay clickable. Answers never load images: an image would open its address by itself, without a click, so an image in an answer is shown as a link you may follow or not. The same applies to summary e-mails.

When the hub cannot build its picture of how the plant is wired, the AI Assistant says so instead of reporting no problems: health checks, dependencies and data states are then not available, and a summary e-mail says the configuration was not checked.

What's next

• AI Assistant settings: add the AI service the AI Assistant answers with, and set its limits.
• MCP server: let your own AI agents read the plant model with the same permissions.