AI for field service technicians: answers from your own manuals and service history

Why first-time fix depends on information

First-time fix rate is one of the numbers a service organization is judged on, and the gap between good and poor is wide. Aquant's 2025 field service benchmark, vendor research covering about 160 service organizations, put the first-time fix rate at 86% for top performers against 53% for the bottom group. It also found that a failed first visit adds two more visits on average and stretches resolution by 14 days.

Information is one part of that gap a service leader can change quickly. In The Service Council's 2022 survey of field engineers, time spent looking for information ranked second among the least favorite parts of the job, and 85% said they need more technical knowledge to do it. When McKinsey surveyed 250 senior buyers of aftermarket services in North America and Europe in May 2024, low-quality talent in the field force was among their top three concerns.

Hiring will not close the gap on its own. The US Bureau of Labor Statistics projects 14% growth for industrial machinery mechanics and maintenance workers from 2025 to 2035, with about 51,900 openings a year. The engineers you already have need faster access to what the company already knows.

Resolve the asset before answering

A good answer starts with the right machine. A question like "what is the torque for the drive coupling" can have a different answer for a 2016 build and a 2022 build, and a bulletin may have changed it again for one serial range. The assistant should first resolve the asset: serial number, configuration, as-built and as-maintained BOM, and the bulletins and engineering changes that apply to it. Only then does it search.

Your best remote support engineers already work this way. They ask for the serial before anything else, because the manual revision and the wiring diagram both depend on it. An assistant that skips that step will give confident answers for the wrong build.

Answer only from approved documents, with the section linked

Every answer should come from a controlled source and show it: the manual and its revision, and the section or page. If the approved documents do not answer the question, the assistant says so and offers to route it to second-level support. General knowledge about similar machines does not count as a source.

Past work orders are useful, but they need a label. A note that says "replaced coupling, fixed" records what one engineer did on one day, which makes it a past observation rather than a procedure. The assistant can show similar past cases next to the answer, clearly marked as history, so the engineer can weigh them, while the procedure itself always comes from the manual.

Regulation backs this up in Europe. The EU Machinery Regulation, which applies from 20 January 2027, allows instructions in digital form and requires them to stay accessible online for at least 10 years after the machine is placed on the market. Its annex on instructions also lists what maintenance instructions must contain, including the specifications of spare parts that affect operator safety. An assistant that quotes and links the controlled version supports that duty. One that rewrites instructions in its own words creates a second, uncontrolled version.

What never goes through the assistant

Some content should be shown, never summarized. Lockout/tagout is the clearest case. OSHA's standard for the control of hazardous energy, 29 CFR 1910.147, covers servicing and maintenance where unexpected start-up or stored energy could injure workers. The standard generally requires documented energy control procedures, subject to a limited exception, that specifically outline the steps for shutting the machine down and for verifying that the energy is controlled. The employer owns those procedures. The assistant displays the employer's approved procedure for that machine directly from the controlled document, with its version, outside any generated text.

The same rule applies to any step that touches guarding or a safety function. If an engineer asks how to bypass an interlock to test a drive, the assistant points to the documented test procedure and to the person who can authorize it, and adds nothing of its own.

Export control for service information

Service information can be export-controlled technology. Under the US Export Administration Regulations, technology includes information needed for the use, installation, maintenance, repair, overhaul or refurbishing of an item, and it can sit in manuals and documentation. A restricted repair procedure for a controlled machine can be controlled technology under the EAR whether it is emailed or shown in a chat window.

Showing it to the wrong person can be an export. The EAR treat a release of technology to a foreign person inside the United States as a deemed export to that person's country, and a release includes written exchanges. A knowledge assistant that shows a controlled procedure to a technician abroad, or to a foreign-national employee in the US, may be making an export that needs a license. In the EU, the Dual-Use Regulation counts making listed technology available electronically to people outside the EU as an export as well.

Each document needs its own classification under the EAR. Where an ECCN controls use technology, only the elements of use named in that entry are covered. Information that falls under no entry is EAR99, and EAR99 technology still needs a license for embargoed destinations and restricted parties. Trade compliance classifies each part and service document. The assistant reads the approved classification from each document's metadata and holds anything unclassified.

So access control has to run before retrieval. Each document carries its classification, and each user carries an authorization from trade compliance that accounts for location and, where the rules require it, citizenship or permanent residency. The assistant searches only what that user may see. A person approving the answer afterwards cannot undo a release that should never have happened. Operator manuals published for anyone to download are a different case, since information made public without restrictions falls outside the EAR.

Works councils and the EU AI Act

In Germany, a knowledge assistant whose logs show which engineer asked what, and how often, is objectively suitable for monitoring performance. Under section 87(1), point 6 of the Works Constitution Act and the Federal Labor Court's 2016 reading of it in 1 ABR 7/15, that suitability triggers co-determination whatever the employer intends. Where a competent works council exists, resolve those requirements before rollout and check whether an existing agreement already covers the use. Since 2021, section 90 has also required the employer to inform the works council in good time about planned work processes that use AI, and section 80(3) treats an AI expert for the works council as necessary.

Design the logs for audit and for improving the content, without turning them into a performance tool. Pseudonymous query logs and an agreed retention period are sensible starting points, and individual dashboards are best left out. Expect the works council to bring its own expert. Austria and the Netherlands have similar consent rules for systems suited to monitoring.

The EU AI Act draws a similar line. As amended by Regulation (EU) 2026/1744, AI used solely for non-safety user assistance or service efficiency does not count as a safety component, but a system whose failure would endanger health and safety does. Classification therefore depends on the intended use and on what a wrong answer could cause. Have counsel assess whether an incorrect maintenance answer could endanger health or safety before treating the assistant as outside the safety-component rules.

A tool that allocates jobs based on individual technicians' traits, or scores their performance, falls in the employment category of Annex III, where high-risk obligations apply from 2 December 2027. Obligations for products under Annex I follow from 2 August 2028. Regulation (EU) 2026/1744 also amends the Machinery Regulation so that delegated acts adding AI health and safety requirements for machinery apply by that date. Article 4 has applied since 2 February 2025, and as amended it asks providers and deployers to take measures to support the AI literacy of their staff. Keep the assistant on the work item and away from rating people.

This is general information, not legal advice. Have your legal and trade compliance teams confirm the requirements that apply to your documents and sites.

Worked example: what the hours look like

This is an illustrative example with assumed numbers, not a client result. Take 60 field engineers who each look something up six times a week, such as a manual revision, a bulletin, a wiring diagram or the last visit on this serial. At 20 minutes per lookup, that is 360 lookups and 7,200 minutes, or 120 engineer hours a week.

Now assume 80% of lookups, 288 a week, can be answered from indexed and approved documents in about 5 minutes, including the time to open the source and check it. The other 72 still need a call to second-level support and take the full 20 minutes. Total time falls to 2,880 minutes, or 48 hours, which gives back 72 engineer hours a week.

Revisits are the bigger prize, and they need careful labeling. Aquant's benchmark found that a failed first visit adds two more visits on average. If first-time fix moved from 75% to 78% on 500 visits a month, which is an assumption and not a promise, that would mean 15 fewer failed first visits and, at Aquant's average, about 30 fewer revisits a month.

Starting small with technician knowledge

Pick one equipment family and its documents: the manuals by revision, the bulletins, the wiring diagrams and two years of work orders. Have trade compliance mark which documents are restricted, and inform the works council if you operate in Germany. Then, as an illustrative starting sample, ask your second-level support team for the 50 questions they answered most often last quarter, and test the assistant against their answers with nothing going to the field.

The nearest work we can show you is about findability. For Rockwell Automation we rebuilt the intranet and its search for 28,000 employees, and finding information went from 12 minutes to under 4. That was intranet and search work rather than a service desk tool. What carries over is the work of making the right document easy to find.

Those 50 questions also give you the acceptance test, such as the share answered correctly from the right revision with the source linked. With us, that test is signed before the build and the assistant is built in your environment in two weeks. You pay $10,000 only after every criterion passes, and running it afterwards starts at $1,200 a month.

Common questions