Table of Contents
An operator assistant is an AI layer on top of your approved digital work instructions. It finds the right instruction and step for the equipment in front of the operator, answers questions in Arabic or French, and quotes the controlled text, warnings and values word for word. It explains; it never writes or changes the procedure.
Key takeaways
- The released work instruction stays the reference. The assistant retrieves, quotes and explains it, and every answer cites the document, revision and step.
- Guidance helps new workers most. In one field study, an AI assistant raised productivity by 14% on average and by 34% for novice and low-skilled workers, with minimal effect on the most experienced (Brynjolfsson et al., 2023).
- Constant instructions can slow experienced workers down (Kosch et al., 2017), so the assistant should answer on demand.
- Retrieval reduces but does not remove invented content: commercial RAG tools for lawyers still hallucinated 17% to 33% of the time (Magesh et al., 2024). Safety warnings, limits and values must be shown verbatim.
- Arabic and French need explicit handling: answer language, original-language quotes, tags and units kept intact, and testing per language pair.
What is the difference between digital work instructions and an operator assistant?
Digital work instructions are controlled procedures displayed at the workstation, on a screen or tablet, step by step, with the drawings, values and checks for that product or piece of equipment. They are documents: written by a procedure owner, approved, released under a revision number and withdrawn when superseded.
An operator assistant adds a question-and-answer layer. The operator asks "what torque for the flange on P-201?" or "ما هي خطوة التنظيف بعد تغيير الدفعة؟" and the assistant retrieves the relevant step, quotes it, and explains it in the language of the question. Technically it is retrieval-augmented generation (RAG), as in our guide to RAG for maintenance manuals, with stricter rules because operators act on the answer at once.
The division of roles matters. Process safety rules treat procedures as controlled material. The US process safety management rule, for example, requires written operating procedures that "provide clear instructions", requires them to be "readily accessible to employees who work in or maintain a process", and requires the employer to certify annually that they are current and accurate (29 CFR 1910.119(f)). It is not Algerian law, but it is a useful reference; confirm the requirements that apply to your site with your HSE department and legal counsel. An assistant makes procedures easier to reach. It does not take over their ownership.
Why are paper instructions not enough?
The UK HSE names "a failure to follow safe working procedures" and "the use of inadequate procedures" among the main causes of procedure-related problems, alongside too much reliance on procedures to control risk, and warns that procedures "do not replace competency" (HSE). Its advice is to involve users in writing procedures and to use "a format, style and level of detail appropriate to the user, task and consequences of failure", whether step-by-step instructions, checklists, decision aids or flowcharts.
A paper binder serves one format to everyone. Researchers studying assembly workplaces note that experienced colleagues are not always available to guide new workers, and that paper instructions are easily outdated and must be swapped at every product change (Kosch et al., 2017). In a bilingual plant there is a third problem: the instruction may be written in French while the operator thinks and asks questions in Arabic.
What does the evidence say?
Published evidence on LLM assistants for plant operators is still thin. The closest studies:
| Study | Setting | Finding | Caveat |
|---|---|---|---|
| Tang et al., 2003 | Assembly task: printed manual, monitor, head-mounted display, or 3D instructions overlaid on the work pieces | The overlay cut the error rate by 82% and lowered mental effort | Experiment with augmented reality, not a language model |
| Kosch et al., 2017 | Projected step-by-step instructions at manual assembly workplaces | New workers benefited while learning; experienced workers found them distracting, with longer task times and higher workload | Projection, not a language model; workshop paper summarising a research project |
| Kernan Freire et al., 2024 | LLM tool over manuals and issue reports, tried by 4 managers and 5 operators from a detergent factory | Users valued quicker access to information; one manager warned "if the responses are not adequate, you risk safety"; participants preferred a human expert when one was available | Nine participants; not tested during live production |
| Brynjolfsson et al., 2023 | AI assistant for 5,179 customer support agents | 14% more issues resolved per hour on average, 34% for novice and low-skilled workers | Customer support, not a plant |
Two lessons carry over. First, assistance pays off most for people still learning, and the factory study's authors see the tool as something to use when no expert is at hand, such as on a night shift run by a single operator. Second, help that guides a new worker can slow down an experienced one, so the assistant should answer on demand, not push every step.
How should the assistant handle Arabic and French?
Language is where a generic chatbot fails first. The rules below build on our guide to documents that mix Arabic, French and English.
- Answer in the language of the question, and quote the source in its own language. An Arabic question about a French instruction gets an Arabic explanation and the French step text beside it, so the operator can check values and terms against the original.
- Check the model card. At the time of writing, the Llama 4 model card lists Arabic, English and French among its supported languages (Meta). Supported languages differ between models and versions, so check every candidate.
- Test answer language explicitly. Language models sometimes answer in the wrong language, and this "language confusion" grows with complex prompts and high sampling temperatures (Marchisio et al., 2024). State the answer language in the instructions given to the model and keep the temperature low.
- Keep tags, values and units intact. Equipment tags such as P-201 and units such as N·m run left to right inside Arabic text. The Unicode Bidirectional Algorithm acknowledges that implicit ordering is sometimes "not sufficient to produce comprehensible text" and defines isolate characters for such runs (Unicode UAX #9). Test that ranges and units display in the right order.
- Treat any translated instruction as a new controlled document. Machine translation can produce the first draft; a fluent engineer who knows the process approves it.
- Test voice input on your own operators. Speech recognition helps when hands are busy or gloved, but Whisper's own documentation says its "performance varies widely depending on the language" (OpenAI Whisper). Record samples with real accents, Darija and French mixed in the same sentence, and shop floor noise.
What should an operator assistant never do?
| Risk | Plant example | Control |
|---|---|---|
| Invented or altered value | Answer gives 45 N·m where the instruction says 54 N·m | Values, limits and warnings displayed verbatim from the released document, never paraphrased |
| Outdated revision | Answer cites a procedure that was superseded last month | Index only released revisions; remove superseded ones automatically |
| Dropped warning | A summary of a cleaning step leaves out the lockout requirement | Warnings attached to their step and always shown with it |
| Deviation advice | "Can I skip step 4 today?" | No permission to approve deviations; refer to the supervisor and the change procedure |
| Injected instructions | A contractor document in the library contains text addressed to the AI | Restrict who can add documents; OWASP lists a poisoned RAG document as an indirect prompt injection scenario (OWASP LLM01:2025) |
| Answer outside the library | Question about equipment with no instruction indexed | Say that no instruction was found and who to ask |
Retrieval does not make a model reliable on its own. When researchers tested commercial legal research tools built on retrieval, which some providers had promoted as eliminating or avoiding hallucinations, the tools still hallucinated 17% to 33% of the time, giving false statements or claiming that a source supported a statement when it did not (Magesh et al., 2024). That is why the step text is quoted, not regenerated. Our guide to measuring hallucination in an industrial assistant covers testing.
Deviation questions deserve a firm rule. The assistant explains what the instruction says and why; whether a step can be skipped is a decision for a named person under the site's existing procedures, as described in human-in-the-loop for AI agents.
How do you build an operator assistant, step by step?
- Choose one line or unit and the 10 to 20 instructions operators consult most, including at least one changeover and one cleaning or start-up procedure.
- Fix the source documents first. Each needs an owner, a revision number, a language and a clear step structure, with warnings and values in identifiable fields. Scanned documents need OCR and a manual check, as described in our OCR guide.
- Index by step, not by page, with the equipment, product, revision and language attached to every chunk.
- Set the answer rules: cite document, revision and step; quote warnings and values verbatim; answer in the language of the question; say so when nothing relevant is found.
- Build a test set with operators, covering every question language and document language combination, deviation questions, questions with no answer, and questions that mix Arabic and French.
- Choose the display with the users. A tablet at the station, a screen on the line or voice input, with a short checklist view for experienced operators.
- Run a pilot alongside the existing instructions on all shifts, including nights, and collect feedback separately from new and experienced operators.
- Route failures to the procedure owner. Questions the assistant cannot answer, or answers operators flag, often reveal a gap or ambiguity in the instruction itself.
How do you know it works?
Track a small set of measures from the pilot onwards:
- the share of answers that cite the correct document, revision and step;
- answers in the wrong language, and values that do not match the source;
- questions with no answer, by topic;
- instruction revisions triggered by operator questions;
- feedback by experience level and by shift.
Training people to work with the tool matters as much as the tool; see our article on training engineers in AI skills.
What we have seen at KABAS AI
DALILE, our AI assistant, runs on the client's own servers without internet access, organises documents into searchable knowledge libraries and answers questions with exact sources. Those are the properties an operator assistant needs: procedures stay on site, and every answer can be traced to the released text. Our manufacturing solution puts precise work instructions and instant access to equipment documentation at the execution stage of production.
We have not published an operator assistant project. Our practical advice for one is to start with a language inventory of the instructions on one line: which documents exist in which language, who owns them, and in which language operators ask questions. That inventory decides the test set and which translations need a reviewer. Linking the assistant to the shift log, as in our guide to automating shift reports and handover notes, lets issues raised at handover point back to the instruction concerned.
Our agentic AI team deploys DALILE and connects it to your document management systems.
Sources
- Brynjolfsson, Li and Raymond, Generative AI at Work, NBER Working Paper 31161, 2023
- Kosch, Funk, Abdelrahman and Schmidt, One Size does not Fit All: Challenges of Providing Interactive Worker Assistance in Industrial Settings, UbiComp/ISWC Adjunct, 2017
- Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, arXiv 2405.20362, 2024
- 29 CFR 1910.119, Process safety management of highly hazardous chemicals (Cornell LII)
- UK HSE, Human factors: Procedures
- Tang, Owen, Biocca and Mou, Comparative effectiveness of augmented reality in object assembly, CHI 2003
- Kernan Freire et al., Knowledge sharing in manufacturing using LLM-powered tools: user study and model benchmarking, Frontiers in Artificial Intelligence, 2024
- Meta, Llama 4 Scout 17B-16E Instruct model card (Hugging Face)
- Marchisio et al., Understanding and Mitigating Language Confusion in LLMs, arXiv 2406.20052, 2024
- Unicode Standard Annex #9, Unicode Bidirectional Algorithm
- OpenAI, Whisper (GitHub repository)
- OWASP GenAI Security Project, LLM01:2025 Prompt Injection
Frequently asked questions
Can an LLM write our work instructions?
It can draft or restructure them, for example turning a long French procedure into numbered steps with warnings and values in separate fields. The draft then goes through your normal document control: the procedure owner checks it against the process, it is approved and released, and only the released revision is given to the assistant.
Should we translate all our instructions into Arabic?
Not necessarily. The assistant can answer an Arabic question from a French instruction while showing the original passage. Where a written Arabic version is needed, treat it as a controlled document in its own right: a machine translation can be the first draft, but a fluent engineer who knows the process must review it before release.
Does an operator assistant need internet access?
No. Open-weight language models, search indexes and speech recognition models can all run on servers inside the plant. This keeps procedures, which often contain process know-how, on site, and keeps the assistant available when the external link is down.
What happens if the assistant's answer contradicts the instruction?
The released instruction always prevails. Design the assistant to show the step text, warnings and values verbatim next to any explanation, so the operator can see a contradiction at once, and give operators a one-click way to report it. Each report should reach the procedure owner.
Will experienced operators use it?
Only if it stays out of their way. Research on projected step-by-step assembly instructions found that experienced workers found them distracting and were slower with them, while new workers benefited while learning. Make the assistant available on demand rather than pushing every step, and let experienced operators switch to a short checklist view.



