Help Centre › Building Processes
The supervisor watches your processes run. When a step keeps failing, or a run wastes time, it works out a small change and tries it on the next runs. It keeps the change only if the numbers get better, and puts it back if they do not.
It is off until a company admin switches it on.
1. Switch it on
- Go to Settings › Automation › Supervisor. Only company admins see this page.
- Switch on Let the supervisor propose and try changes to your processes.
- For each kind of change, choose Try it without asking, Ask the process owner first or Never.
- Click Save.
2. What it can change
| Kind of change | What it does | Starts as |
|---|---|---|
| Wait longer for a step | When a step keeps failing because the screen was not ready yet, it waits longer before giving up. A longer wait costs nothing on runs where the screen is ready. | Without asking |
| Wait less on a step that never succeeds | When a step times out on every run, even the ones that work, its wait is shortened so every run finishes sooner. | Without asking |
| Replace a fixed pause with "wait until ready" | Carry on as soon as the screen is ready, and wait longer when it is slow. | Ask first |
| Stop runs in hours they always fail | Stop starting runs at the times of day they nearly always fail, for example when the other system is down overnight. | Ask first |
| Answer a pop-up the bot has not seen before | When runs stop at an unknown pop-up, the owner chooses which button to press, or to stop the run. | Always asks |
| Use the backup selector that worked | When a step's main selector stops finding its button or field but a backup finds it, the backup becomes the main one. The old one stays as a backup. | Ask first |
A pop-up answer is always the owner's choice; no setting makes it automatic. A backup selector is only tried without asking when the bot was at least 80% sure it found the right control.
The supervisor has no way to change business values (amounts, accounts, addresses), credentials, environments or promotion, checks and verify steps, the order of steps, or to delete anything. It also never retries work after a step that can't be undone.
3. Where you see it
Open the process and choose the Runs tab. Why it keeps failing is at the top, one row per step, in three groups:
- Worth fixing: ranked by failed runs in the last 7 days, then by time lost. Three show; the rest are folded under more.
- Being tried: changes on trial now, and who made the call.
- No harm this week: folded away.
Click a row to see its step on the map. Why? lists everything the supervisor found about that step. A row can also show the bot's own reading of the cause and what to do about it (see Why did my run fail?), and fixes that worked before for the same kind of failure.
Look into it compares the process's recent failed runs with the ones that worked, step by step. That is the only way to find a wait that is wasted on every run, because those runs succeed.
If nothing shows, it says Nothing I recognise. That does not mean the process is healthy; it only knows the patterns it has been taught.
4. Accept or turn down a change
Only one person gets the buttons: the person named in Who gets called when it breaks under Edit › Care level on the process, or, when nobody is named there, the person who built the process. They are told about every change.
- Try it: the change is tried on the next runs. It is kept only if it helps.
- No thanks: the step is left alone. It comes back only if the step starts failing clearly more often.
- Hide until it gets worse: not worth fixing for now.
- For a pop-up, one button per answer, such as Press "OK" or Stop the run, plus No thanks. The bot then gives that answer whenever the pop-up appears.
- Ask AI what to change: shown when the supervisor has no change it can make. It opens the chat, which reads the step and its recent failed runs. Nothing changes until you do it.
A suggestion nobody answers is withdrawn after 14 days.
5. How a trial is judged
A change is only ever tried. The supervisor compares the runs after the change with the four weeks before it, not just the bad spell that prompted it, since a bad spell often eases by itself.
- It waits for 20 runs, or 10 runs after 3 days for a process that runs rarely.
- If the step clearly gets worse, the trial can stop early.
- If there are still too few runs after four weeks, the trial ends.
- Runs that failed on an earlier step, that a person stopped, or that failed because the application was down or the bot was logged out do not count against the change.
At the end the change is Kept or Put back. After a change has been put back twice, it is not tried again.
6. Undo a change
Every change is in the process's Changes tab. To reverse one, click Undo on its row in Why it keeps failing. The step goes back to how it was before the change.
7. Limits
- Up to 3 changes are tried at once on one process, each on a different step. A change to the whole process, such as its hours, is tried alone.
- At most 6 changes a week on one process, and 10 a day across your company.
- A wait is never set longer than 10 minutes, or more than 3 times what it was.
- On or after a step that can't be undone, the supervisor only ever makes a step wait longer by itself. Anything else there needs the owner's yes.
8. Make it stricter for one process
On the process, open Edit › Care level and set What the supervisor may change here:
- Follow the company rules
- Ask me before every change
- Do not propose changes
A process can only be made stricter than the company, never looser.