How to Answer Technical Machine Questions Over WhatsApp with AI.
There's an edge bander on the factory floor, more than fifteen years old. It glues and trims the edge banding on melamine panels, and when it runs well nobody thinks about it. Its documentation is five PDFs: the machine manual, the manual for the software that drives it, the spare-parts catalogue, the pneumatic diagrams and a booklet of small parts. 1,033 pages, full of names like IN 801, J0052N00 and 2420A1073.
When something goes wrong, nobody opens those thousand pages: the operator goes looking for the guy who has been there twenty years. If he's around, it's solved in five minutes. If he's on holiday, on another machine or talking to a supplier, the line waits.
That factory uploaded the five PDFs to a VibeTasking agent and connected a WhatsApp number to it. That's the whole build. Operators now ask from their phone, standing next to the machine, and get the help they need, with screenshots of the diagrams and the exact manual page.
The questions don't arrive well written
Nobody types "query regarding the end-trimming unit". They type what they see, with typos, from their phone, with the machine stopped behind them:
- "The edge bander, the end trimmer won't cut."
- "The banding is left too long on the bottom. Tell me what parameters I have to change, where on the screen, and by how much."
- "Part 2420A1073 broke on me, can you send me an exploded view of all its components and how I'd take it off?"
- "The machine stopped on me with 3 panels inside and it gives me this error. Alarm on run 36, rear trimmer safety."
And in the same thread, mixed in with those, are the questions from someone two weeks into the job:
- "What's the premelter?"
- "What's the feeder?"
- "How do I know when to put a panel in, and how close together can I feed them?"
Both kinds are worth the same. The first avoids a breakdown. The second stops someone learning by ruining panels.
What it does before answering
The agent doesn't search a PDF for a word. Every question goes through the same route.
First it decides which document applies: safety, adjustments and faults go to the machine manual; screens, programs and parameters to the software manual; codes and exploded views to the spare-parts catalogue; air, cylinders and valves to the pneumatic diagrams. Then it searches an index of the 1,033 pages and opens the ones that come back. If the page is a drawing or a table, it renders it as an image and looks at it, because plenty of pages in the pneumatic diagrams have almost no text on them. It cross-checks sources when the answer needs it: a screen parameter usually has a physical adjustment behind it. And it cites the document and page behind every claim.
It sends the pages too. When an operator asked for the exploded view of the rear part that adjusts the blade, four images cropped from the catalogue arrived over WhatsApp: the assembly, the rear carriage, the position list with the orderable code, and the pneumatic diagram for that unit, each with one line explaining what you're looking at. The first reply landed in four minutes.
What it doesn't do matters just as much
It doesn't fill gaps. When the documentation falls short, it says so and explains how far it goes:
With the documentation available, I haven't found a table that literally confirms the "Run in use 360" alarm. But going by the text "left profiler programming", it looks like a programming conflict on that unit rather than a mechanical fault.
I can't confirm from the documentation which way that nut loosens. Don't force it with an impact wrench or by hitting it hard.
The manual doesn't give a full procedure for removing the whole motor. If you need the spindle out, have qualified maintenance or the manufacturer's service engineer do it.
A factory can work with an assistant that admits what it doesn't know. With one that improvises, something breaks.
Safety comes before the answer
Every question that involves putting a hand inside the machine starts the same way: main switch locked at 0-OFF, air valve closed and locked, wait until the tools have stopped completely, gloves on. It comes before the steps, not as a note at the end.
And there's a line it won't cross. An operator asked whether the red cadence light, the one that says when the next panel can go in, could be disabled in the settings or whether it was faulty. The answer was that it appears in no manual as an operator setting, that a light which doesn't work is a fault for maintenance, and that working without that reference means feeding panels too close together and crashing them in the trimming unit. At that point the agent is stopping someone from feeding a panel too early.
It also looks at what you send it
An operator wanted to change the glue temperature, but the glue unit wasn't the edge bander's own, it was another manufacturer's controller that appears in none of the five manuals. The agent said so and asked for a photo of the front panel.
What arrived was a blurry photo, taken from a hand's width away, of a panel with a number pad. From that it identified the model, explained which control selects the zone, where the arrows to raise and lower the value were, and which key confirms. It read the display on the way past: 137 °C, so it warned that this looks like a standby value and that the right one is whatever the glue bag specifies, because too low won't bond and too high burns the glue.
Another photo, another day: a trimmer tracer caked in dried glue. The agent connected the symptom the operator described, the next panel coming out with the banding torn off, to the tracer riding on the built-up glue instead of on the melamine. It told him to clean it with alcohol rather than a blade or sandpaper, and pulled from the manual that this copier is cleaned every 40 hours. The fault wasn't where it looked.
What changes on the floor
Eight operators have used it so far, each with their own WhatsApp thread open. Answers land in one to three minutes, with the manual pages attached when they're needed.
The deeper change is who the knowledge depends on. It used to live in two people's heads and in a folder nobody opened. Now the night shift asks what the morning shift asks, the guy two weeks in asks what the premelter is without feeling stupid, and the guy twenty years in stops being the only route to knowing which page holds the rear-overhang adjustment.
The channel also knows what isn't its job. One day a message about an Amazon order for work trousers landed there by mistake, size chart and all. The agent replied that this number is for technical questions about the edge bander and suggested checking the recipient before sending.
How it's built
Three pieces, and none of them is custom software.
- An agent with the instructions for this job: precision above everything, cite document and page on every claim, safety before the steps, say explicitly when the documentation falls short, and confirm nothing that isn't backed by it.
- The five PDFs travel with the agent as resources, alongside an index of the 1,033 pages and a search tool that understands an operator asking in Spanish about a manual written in English. When the factory adds documentation, it gets uploaded and that's it: no rewriting the instructions.
- A WhatsApp channel. Each operator has their own conversation, and the agent remembers what they discussed in the previous ones.
And behind it, a person one tap away. When a question goes beyond what the documentation can back, the agent says so and recommends escalating, rather than signing off on a procedure it made up.
This isn't about edge banders
Almost every company has a machine, a piece of equipment or a system with dense documentation that nearly nobody reads: a packaging line, a forklift, a boiler, a CNC, a refrigeration plant, a medical device, the ERP installed eight years ago. The documentation exists, it's good, and it's useless, because nobody searches 1,033 pages with the line stopped.
The pattern is always the same. Documentation you already have, people who need one specific answer right now, and a channel those people are already on.
Start with the equipment that stops you most. Upload its manuals, connect a WhatsApp number and let the first shift ask what they were going to ask anyway. You can build it yourself, or tell us which machine it is and we'll stand up the first version with you.
Keep reading
A calendar that books itself.
A firm handed client reactivation to a WhatsApp agent: 8,000 clients reached, 1,000+ appointments booked, and the whole system built in under three hours.
Invoices to Excel with AI.
Extract PDF invoice data with AI, check tax IDs and totals, and create the Excel file your accounting software imports. A real case with costs per document.
Instagram DMs, answered solo.
Learn how to build an Instagram chatbot that answers DMs on its own: handles customers, resolves common questions, books appointments, and connects to your tools (calendar, WhatsApp, email, or CRM).