AI has a narrow working range in manual production. In derivative production, only one of the eight process steps centers on writing. Most technical documentation work inside a manufacturer is derivative production.
This article draws on operating data from a large manufacturer’s multilingual manual development program. It separates new production from derivative production and rates all eight process steps for AI applicability. It also marks the two places where AI earns its keep.
Documentation managers at manufacturers field a version of this question every quarter, usually the way an executive frames it. Tools draft reports and marketing copy in seconds, so why does the operator’s manual still take a team of writers and a localization vendor?
The question is fair. It carries an assumption worth surfacing, that producing a manual is mostly a writing job. Open the process step by step and the picture changes.
New production and derivative production are different jobs
Manual production splits into two kinds of work. The split is the strongest predictor of where AI can contribute.
| New production | Derivative production | |
|---|---|---|
| Starting point | No base document. Begins from product source material | An existing base manual that has already been validated |
| Nature of the work | Builds the document from scratch | Analyzes differences against the derivative model and reflects only those |
| Share of workload | A small number of jobs, such as the first manual for a new product family | Most of the annual volume |
Published AI success stories come almost entirely from new production conditions. With no base to respect, the sentence generation capability of an LLM is genuinely useful.
Manufacturing reality runs the other way. In a large manufacturer’s multilingual manual program, with a stable product lineup and derivation repeated over several cycles, reuse of existing content runs 70 to 80 percent. Newly written text often stays near 10 percent of the body. The remainder is partial revision of sentences that already exist.
That 10 percent is not free-form writing either. New sentences have to sit correctly against the terminology and style of the base manual, and against the multilingual translation data already banked for it. When a writer has to re-edit AI output back into the house standard, throughput and quality both tend to move the wrong way.
Where the AI conversation goes off track
Tool performance is seldom the issue. The mismatch begins when success cases drawn from new production get applied to a workload that is mostly derivative.
Six conditions that set manuals apart from ordinary content
AI practices proven on blogs and marketing copy do not transfer to an operator’s manual, and the reason sits in the conditions this document type carries.
| Condition | What it means in practice |
|---|---|
| Strict constraints | Terminology and phrasing consistency, safety messaging under standards such as ANSI Z535.6, and region-specific regulatory and certification content govern how the text can be written. |
| Continuous derivation and change | Manuals derive from existing documents and get revised repeatedly between program start and final pre-launch delivery, driven by software updates, design changes, UI changes, and certification changes. The timing and content of those changes are hard to predict. |
| Hard deadline | The manual ships with the product. The date is closed to negotiation and governs the entire process. |
| Zero-defect requirement | Every error has to be cleared before launch. Prose errors matter, and a missing regional or certification requirement carries liability exposure. |
| Multilingual fan-out | One source manual reaches markets in anywhere from one to several dozen languages. |
| Low latitude | Standards compliance and terminological consistency are absolute, which leaves almost no room for creative phrasing. |
The last item carries the most weight. Variety of expression is a headline strength of generative AI, and in this document type it functions as a defect. A manual that describes the same operation differently each time is a worse manual.
AI applicability across the eight steps of derivative manual production
Derivative manual production generally runs through eight steps. Rating each one by the nature of the work and the room available for AI gives the following picture.
| Process step | Nature of the work | AI applicability | Basis |
|---|---|---|---|
| Specification analysis | Secure the base model, review target model specifications, identify differences, plan content reuse | Low | Judgment work that pulls only the relevant items out of unstructured source material |
| Derivative content authoring | Call and adapt existing content per the reuse plan, write new text for the differences | Partial | New text stays near 10 percent of the body, and generated sentences still need re-editing to terminology and style standards |
| Visual asset production | Product screen captures, derivative versions of illustrations | Not applicable | A factual record that has to match the physical product and its screens exactly. Generated imagery cannot substitute |
| DTP and output generation | Layout and style editing, output file generation | Low | Typesetting work of a different nature from text generation. Rule-based automation territory |
| Review and change handling | Draft submission and feedback, incorporation of pre-production changes | Low | Judgment and coordination around changes whose timing and content cannot be predicted |
| Multilingual translation | Translation memory driven translation, translation of new sentences | New sentences only | Most of the body reuses validated existing translation. The AI-relevant slice sits near 10 percent and translator editing stays mandatory |
| Multilingual DTP | Per-language layout editing, text expansion and contraction, right-to-left and other scripts | Not applicable | The work handles pages rather than text. Per-language typesetting rules and visual verification carry it |
| Publishing, delivery, version control | Multi-channel publishing, final delivery, version and revision history | Low | Rule-based conversion territory. Approval and coordination still need people |
Why only one of the eight steps is writing
Read the applicability column top to bottom and the pattern surfaces. Of eight steps, one centers on text generation, derivative content authoring. One more touches it partially, multilingual translation. Both stay confined to the slice of new sentences that sits near 10 percent of the body.
Group the remaining six by what the work actually consists of and four categories appear.
| Category | Steps |
|---|---|
| Judgment | Specification analysis, review and change handling |
| Typesetting | DTP and output generation, multilingual DTP |
| Visual record | Visual asset production |
| Coordination and control | Publishing, delivery, version control |
Why multilingual DTP production does not go to zero with AI
Automation hopes land hardest on multilingual DTP production. One manual fans out to dozens of languages and cost scales with language count and page count, which makes it look like the highest-yield target on the board.
The cost in this step does not come from writing. It comes from per-language layout adjustment, page reflow after translated text changes length, handling of right-to-left scripts, and verification that regional regulatory and certification marks landed correctly in every language.
Aiming at zero for this step is a reasonable goal. The lever sits elsewhere. Structuring content for single-source management and generating per-language output through automated composition is a content management and publishing question, and a CCMS is one route to it. That work reduces the typesetting step itself, which puts it beyond the reach of text generation. The subject falls outside this article, so we note the direction and leave it there.
The tool and the problem are misaligned
An LLM generates text. Point a text generation tool at a step whose cost is typesetting and verification, and the cost structure stays where it was. Shrinking this step belongs to content management and publishing.
The two places AI does pay off
AI produces reliable returns in two places in manual production, document analysis and reuse of finished documents.
Analysis is the surer of the two. An LLM reads structure across large document sets and catches differences faster than a person can. Benchmarking competitor manuals, diagnosing document structure against international standards such as IEC/IEEE 82079-1, assessing the quality of current documentation, and source research ahead of new production all sit here. This is the one use that holds without conditions, and it applies to derivative and new production alike. The role is analytical support for documentation strategy.
The second is reuse of what has already shipped. A validated manual is a content asset with accuracy guaranteed, and AI converts that asset into other formats well. FAQ content, chatbot knowledge bases, voice guidance text, and training material are the usual destinations.
Three questions to settle before evaluating tools
For a documentation organization looking at AI, these three set expectations before any tool comparison starts.
- Does a base document exist for this work? If it does, the job is derivative and generation capability has little room to operate.
- Does the cost of this step come from text generation? Where the cost sits in typesetting, verification, and judgment, a text generation tool will not move it.
- Does this deliverable tolerate errors? Where zero defects are required, AI output always carries a verification stage behind it, and that cost belongs in the calculation.
The workable operating model is a division of labor. AI handles analysis and drafting within a bounded scope, and a standardized content reuse system holds consistency across variants. Final quality stays with professional technical writers. As long as technical documentation carries the conditions it carries, that division protects cost and quality together.
Frequently asked questions
- Can manual production be fully automated with AI? Not realistically. Most of the working volume is derivative production, where reuse of validated existing content is the center of the job. The reuse rate varies with document type, how similar the product lineup is, and how many times derivation has repeated. With a stable lineup and several cycles behind it, 70 to 80 percent is typical, and heavily repeated programs run past 90 percent. In every case, one of the eight process steps centers on text generation. The rest are judgment, typesetting, visual record, and coordination.
- How much new text does a derivative manual actually need? Around 10 percent of the body, varying with document type and how often derivation has repeated. Partial revision of existing sentences sits on top of that. New sentences also have to match the terminology and style of the base manual and the multilingual translation data banked for it, so human editing follows.
- Can AI reduce multilingual DTP cost for manuals? The cost in that step comes from per-language typesetting, page reflow after translation, and verification of regional regulatory and certification marks. Adding a text generation tool leaves the cost structure intact. Structured content and automated composition are the levers that move it.
- So where should a documentation team use AI? Document analysis and reuse of finished documents. Benchmarking competitor manuals, standards-based structure diagnosis, quality assessment of current documentation, and source research for new production are all effective, as is converting validated manuals into FAQ, chatbot, and training content.
Read the full analysis
The complete comparison of the eight derivative production steps against six representative LLM capabilities, built on operating data, is available in the white paper “Expectations and Reality in Manual Production Automation.“