“Can We Mute the Light Curtain to Clear a Jam?” Answer From That Serial's Own Manual, or Not at All

"Can we mute the light curtain to clear the jam?" That question arrived at 6:50 in the morning, Central time, in a two-line email from a maintenance supervisor at a customer plant, with the serial number in the subject line and nothing else. The second shift had been clearing an infeed jam twice a night by defeating the guard, and somebody finally wondered whether that was allowed. Whoever on your service desk answers that email first is writing a document that could be read aloud in a deposition three years from now.

That is the shape of the risk for an industrial equipment builder. Your service desk answers technical questions about machines you shipped ten and fifteen years ago, from memory, from a shared drive with three revisions of the same manual in it, and from two people who know. Most of the answers are fine. The expensive ones are not the questions the desk gets wrong loudly. They are the ones it gets wrong confidently.

What does the desk actually do with that email today?

The service coordinator searches a shared drive organized by customer name, not by serial number. She finds an installation, operation and maintenance manual — probably not the revision that shipped with that machine. She finds a folder of field service reports, some scanned as photographs. She asks the two guys. If they are on a plane or on a site, she either waits or she answers from the manual she found, because the customer's line is down and waiting has a cost too.

Now consider the specific failure modes. The manual at revision C says the knife bolts torque to 42 foot-pounds. A change order in 2019 moved that assembly to a different hub and a higher-grade fastener, and revision F says 65. Give a revision C answer to a revision F machine and you get a knife coming loose at speed. Or the lubricant question: a machine in a food plant takes a food-grade lubricant, and the wrong one means the customer is holding product, which is a claim rather than a service call. Or the electrical one: a customer modified the control panel in the field, which quietly ended the listing on that panel, and now they are asking you to bless it in writing.

What changed in 2026 that makes a wrong answer catchable?

Two years ago an assistant would answer that torque question instantly and fluently from a general reading of the world, and there was no way to tell the good answer from the confident one. What changed is that answers can now be restricted to a set of documents you approved, and the answer has to show which document and which page it came from.

Grounding means the assistant may only answer from material you have approved, and must show you the source — so a wrong answer looks wrong instead of looking plausible. It does not make the assistant always right. It makes it checkable in ten seconds by a coordinator who is not an engineer, which is a completely different property, and it is the one that lets you put this in front of a service desk at all.

For a machine builder the approved set is specific: the as-built manual for that serial and its current revision, the change order log, the safety circuit calculation on file, the electrical drawings, the parts list, and the field service reports for that machine. Not the shared drive. The shared drive contains a salesman's marked-up copy from 2011 and two superseded revisions, and if you point an assistant at all of it you have automated the exact mistake you are trying to stop.

flowchart TD
  A["Maintenance supervisor emails the service desk at 6:50am"] --> B["Agent reads the serial and opens that machine's as-built file"]
  B --> C{"Is the answer in an approved document for THIS serial?"}
  C -->|"Yes, current revision"| D["Answer, naming manual, revision and page"]
  C -->|"Only in a superseded revision"| E["Say which revision, route to the service manager"]
  C -->|"Not in any approved document"| F["Refuse and open a service ticket"]
  D --> G["Coordinator sends and files the answer to that serial"]

Which questions in this shop are expensive to get wrong?

Sort your service email for a month and the costly categories separate themselves out. They are not evenly distributed, and they are worth writing on a whiteboard:

• Anything inside the safety circuit. Substituting a light curtain, a safety relay or an interlock switch changes the rating of that circuit, which means your risk assessment and conformity file no longer describe the machine in the plant.
• Guarding and defeat questions. The muting question above has exactly one correct handling: it does not get answered by an assistant, it gets routed to a person, and the answer is a documented procedure rather than an email.
• Lockout energy sources. Stored pneumatic and gravity energy on a machine with a lifting frame is where the manual and reality most often diverge, especially after a customer added a conveyor.
• Torque, fastener grade and setpoints on rotating assemblies. Revision-sensitive, and failure is immediate and violent.
• Lubricants and materials in contact with product. Wrong food-grade class means held product.
• Interchange on pressure-retaining or code-stamped components. "Can I use this other valve" is a code question wearing a parts question's clothes.

What does a grounded answer actually read like?

Like this, which a coordinator can send in under a minute:

For serial 4471 the infeed guard door interlock must not be defeated during production. Jam clearing is covered in the manual for this machine, revision F, section 6.4, page 6-12, which requires the machine be placed in a de-energized state before the infeed guard is opened. Note: this machine shipped with the light curtain listed on drawing E-4471-03, and the safety circuit calculation on file rates that channel at the level stated in the customer's specification. Substituting a different device changes that calculation. Routing to the service manager for a written response.

Two features make that answer safe. It names the revision, so a coordinator can see instantly whether it read the right book. And it stops rather than blessing a modification. The single most valuable sentence a grounded assistant produces is "I do not have an approved document that answers this" — because today the desk's alternative to that sentence is a guess in an email.

What does one bad spec answer cost, and what does the approved set cost?

Assumptions, illustrative: a single wrong torque answer leads to a rotating assembly coming apart, twenty-two hours of unplanned downtime at the customer, and a response you pay for whether or not you are contractually liable, because you want to keep the account.

One avoided event pays for it twice, and the document work is worth doing even if you never turn the assistant on, because right now your defense against that fourteen thousand dollars is one coordinator's search skills at 6:50 in the morning.

Where does a human have to stay in the loop?

Anything that changes a safety function, permanently and without exception. Anything that would amount to written permission to run with a guard defeated — that is a policy answer, not a document lookup, and the policy answer is no. Anything about a machine the customer has modified without you, which the field service reports usually reveal and which invalidates the file you would be answering from.

Keep a person on any statement that touches your conformity file. If you ship into the European Union, also keep in mind that the EU AI Act's transparency and high-risk obligations carry a 2 August 2026 compliance date and reach US companies whose systems affect users there — a service assistant that answers customer questions is something you should be able to describe, log and point to a human owner for. That is not a reason to avoid it. It is a reason to write down what it does before somebody asks you.

Finally, keep the refusal rate visible. If the assistant is refusing 30% of questions in month one, that is the correct behavior and it is also a work list: every refusal names a document you have not written yet.

Frequently asked questions

Our older manuals are scanned drawings from the nineties. Does this work on those?

Partly. Text — procedures, torque tables, lubricant charts — comes through well enough to answer from. Dimensions read off a scanned drawing do not, and you should tell it plainly that drawings are for reference and route dimensional questions to engineering.

Can we just point it at the whole shared drive?

No, and this is the single most common way this fails. Somebody has to decide what is approved. Current revision manuals, the change order log, drawings of record, and safety calculations are in; marked-up copies, draft revisions and old proposals are out. It is a week of unglamorous sorting and it is the entire value of the project.

Should our customers get direct access, or only our service desk?

Internal only for the first ninety days. Watch what it gets wrong when the stakes are a coordinator's time rather than a customer's line. Most builders find two or three answer categories they would never have let out of the building, and they only find them by watching.

What about questions where the customer is asking about a competitor's part on our machine?

That is a routing case, not an answer case. Any third-party substitution on a machine you built goes to your service manager, because it is simultaneously a technical question, a warranty question and a commercial one.

The same question arrives by phone about as often as it arrives by email, usually before your desk is staffed. CallSphere builds voice and chat agents that answer the service line day or night, capture the serial number, the symptom and the site, and book the callback or open the ticket. Hold it to the same rule as everything else here: it answers from your approved material, it names what it read, and anything touching a safety circuit goes straight to a person.