Torek AI maintenance assistant
Updated
The Torek AI assistant answers questions about your own equipment: which defects are open, which machine is due for service, how many jobs are in progress, what is running low in stores. The question is asked in ordinary words, and the answer is assembled from your company's records — not from the internet and not from anyone else's data.
It can also propose an action — raise a work order, change its status, log a defect, reserve a part — but every proposal stays a draft until a person confirms it. It is an assistant sitting next to the data, not an automatic manager.
The questions that pay off are the ones that would otherwise mean opening three or four lists at once:
Before asking, the user can select the assets in question — the whole fleet or a few units. In the answer, machines are named by their display name and registration number rather than internal identifiers, and any defect mentioned carries a link that opens that defect.
| Kind of question | What the assistant reads |
|---|---|
| Assets | The asset list, the record itself and its history |
| Defects | Open defects, their severity and the detail of one |
| Work orders | The job list and aggregate statistics |
| Maintenance | Service intervals and what falls due next |
| Stock | The balance overview and parts running low |
| Inspections | Completed checks and overdue ones |
| People | Who is working and how many open jobs they carry |
Counts are calculated, not estimated: when the question is “how many”, the number comes from a counting tool, not from counting the rows on screen — a list may only be the first page. If no such number exists, the assistant says so.
Four write actions are available, and each one stops at a draft:
Proposals are listed with the fields they would fill in: asset, description, severity, priority, assignee, part, quantity. The user ticks what to confirm, or rejects the lot. The assistant never claims to have done something it only proposed.
How personal data is handled: before any text reaches the language model, employee names are replaced with impersonal tokens and mapped back in the reply — the user sees a real name, the model sees a token. The assistant reads only your company’s records and only the kinds of data the company has allowed in its settings. Torek stores company records in the EU.
What it does not do: it does not search the internet, does not email customers or suppliers, does not act without approval, and does not make the repair-or-replace call for you — it brings the numbers that call rests on, faster. Longer breakdowns and exports belong in the maintenance reports.
A foreman working out what tomorrow looks like without opening five lists. A manager checking whether one machine keeps producing the same fault. A storekeeper seeing what the next repair will need. The assistant is worth exactly as much as the records behind it: it reads the same log the workshop fills in on a phone and at the kiosk.
Only from your company's records: asset cards, defects, work orders, stock balances, inspections and maintenance schedules. To answer, the assistant calls the system's own tools, which only ever read your company's data, and only the kinds of data the company has allowed — data belonging to another company is structurally out of reach.
When the tools cannot produce an exact number, the assistant says so rather than presenting a guess as a fact.
No. Write actions — creating a work order, changing its status, logging a defect, reserving a part — are prepared as drafts and wait for a person. The screen shows exactly what is proposed, and the user confirms the selected ones or rejects them.
Confirming is an ordinary system action: it goes through the same permission and completion rules as a record created by hand.
At two levels. The AI module is enabled for the company, and then access is granted to individual users — it is off by default for everyone. Nobody gets an assistant they were not meant to have.
In settings the company chooses which data it may read (assets, defects, work orders, stock, employees, maintenance schedules, inspections) and which actions it is allowed to propose.
No. Before any text is sent to the model, people's names are replaced with impersonal tokens, and the reply is mapped back — the user sees real names, the model never does. If a token cannot be mapped back, the placeholder stays visible: the system fails safe rather than fails convenient.
That is the company's choice. Settings switch conversation history on or off, set how long it is kept (30 days, 90 days, a year, or indefinitely) and allow the whole history to be deleted.
Yes. The assistant works in the browser and in the Android app, for users who have been granted access. The same rules apply on both: answers come only from your company's records, and every proposed action waits for a person to confirm it.
A walkthrough: questions, record links and action approval.
Get in touch