Manufacturing copilots built on controlled content: current revisions only, every answer cited to a document somebody can open, and a refusal when the revision cannot be established.
A copilot over a manufacturing document estate will find the wrong thing. Not because the model is poor, but because the estate contains every revision of every drawing, specification and work instruction ever issued, and most of them are superseded.
The design of a manufacturing copilot is mostly a set of decisions about which content it may use. These are the six categories, and two of them must be excluded entirely.
Technical documentation nobody can locate
Six repositories, a shared drive, and no agreed authority.
The drawing is current, the copy is not
Because somebody saved it locally in 2021.
Troubleshooting depends on three people
And two of them are near retirement.
A fault that was solved last year
Solved again from scratch, at a different plant.
Reports assembled by hand each shift
From data that already exists in systems.
Plants that do not share what they learn
Because there is no route and no habit.
Instructions in a language the operator reads second
Which is a comprehension risk, not an inconvenience.
A pilot that impressed and then stalled
Usually on content, not on technology.
Ask your document control function how many superseded revisions sit in the repository the copilot would index. In most plants the answer is most of it. That single number decides whether the project is a retrieval project or a content project, and it is available in an afternoon. Every manufacturing copilot that produced a confidently wrong answer in its first week skipped this question.
A manufacturing copilot programme is a content and retrieval programme with a model attached. The model choice matters considerably less than which documents are in the index, what revision state they carry, and whether the retrieval respects it.
So the first phases are classification and retrieval design, done with the document control function, and the evaluation set is built before the pilot rather than after it.
Step 1
Assess readiness: content, systems, access and the real questions
Step 2
Discover every knowledge source, including the shared drives
Step 3
Classify content by control state with document control
Step 4
Identify the copilot use cases and the questions they answer
Step 5
Design the knowledge architecture and the index boundaries
Step 6
Design retrieval to carry revision and effective date
Step 7
Build the evaluation set with the business function
Step 8
Integrate with the systems that hold operational context
Step 9
Design the two answer shapes: shop floor and engineering
Step 10
Pilot with one group, with feedback captured in the tool
Step 11
Evaluate against the set, not against impressions
Step 12
Roll out by group, with content ownership agreed per area
Collect fifty real questions from the people who would use the copilot, with their correct answers. That set is the evaluation suite, the content gap analysis and the scope definition all at once, and it takes a week. Programmes that skip it demonstrate the copilot on questions chosen because they work, which is how a pilot passes and then disappoints in its first week on the floor.
Grouped by who uses them. Every copilot below answers from controlled content with the source cited, and refuses where the revision cannot be established.
Shop floor and maintenance
Shop-floor operator assistants: the next action, cited, in the operator’s working language
Maintenance troubleshooting assistants, answering from manuals, procedures and the maintenance history together
Equipment-specific guidance that resolves to the machine and its revision rather than to the model family
Explicit handover points where the next step needs authorisation, a permit or a quality decision
Engineering and quality
Engineering and technical support copilots over specifications, drawings and change records
Engineering specification search, including metadata search across drawing sets
Quality and compliance knowledge assistants over procedures, standards and your own quality documentation
Quality investigation support: assembling history, previous occurrences and related records, with the findings left to the investigator
Operations and commercial
Manufacturing knowledge copilots for supervisors and planners, integrated with the operational context in ERP and MES
Production reporting copilots that assemble the shift report from systems rather than from recollection
Procurement and supply chain copilots over supplier agreements, terms and specifications
Employee onboarding and training support, which is the use case with the clearest payback and the least risk
Platform
Retrieval architecture with revision and effective date enforced, and exclusions maintained rather than configured once
Enterprise document search and summarisation across repositories, with permissions respected per user
Evaluation harness and prompt management, so a change can be tested rather than judged
Multilingual support using approved translations for anything operational
Feedback capture in the tool, routed to the content owner rather than to a project inbox
These are the real questions. What differs between them is which content has to be right, and that determines how much content work precedes each one.
Equipment troubleshooting guidance
A machine has stopped or is producing out of tolerance and the operator or technician needs the next step.
Standard operating procedure assistance
What is the correct procedure for this operation, on this product, on this line.
Engineering specification search
What does the specification say, which revision applies, and what changed.
Maintenance history summarisation
What has happened to this asset, what keeps happening, and what was done about it.
Quality investigation support
Assembling what is known: previous occurrences, related nonconformances, relevant specifications.
Supplier and procurement knowledge
What did we agree with this supplier, on what terms, to what specification.
Production shift report generation
Assembling what happened in the shift from the systems that already hold it.
Onboarding and training support
A new starter asking the questions they would otherwise ask a colleague.
Ask the three people everybody goes to what they get asked most often. Those questions are the scope, in priority order, from the only people who know. It takes an hour, it is more accurate than any workshop, and it tends to reveal that the most frequent questions are answerable from content that already exists, while the hardest ones are not written down anywhere.
We work within your existing platform decisions and data boundaries. In a manufacturing copilot the retrieval layer determines quality and the model determines fluency, and effort is better spent on the first.
Content pipeline
Ingestion with control state, revision and effective date carried through.
Index and exclusions
Current content in, superseded and draft out, maintained not configured once.
Retrieval
Search that resolves to the right machine, product and revision.
Model layer
Selected for fit, cost and where it may process your content.
Grounding and citation
Source citations and response traceability on every answer. No citation, no answer.
Multimodal input
A photograph of a nameplate, a fault or a part, where that is how the question arrives on a shop floor.
Permissions
Each user sees what they are entitled to see, enforced at retrieval.
Evaluation harness
The question set from section 3, run on every change.
Feedback loop
Captured in the tool, routed to the content owner.
Document repositories and the PLM system, which is where the controlled engineering content lives
ERP and MES for operational context: which product, which order, which line, which machine
The CMMS for maintenance history, including the free-text fields that hold most of the useful detail
Quality management systems for nonconformances, investigations and procedures
Microsoft 365 and your existing collaboration tools, so the copilot appears where people already work
And the shop-floor interface, which is a tablet or a terminal in difficult conditions rather than a desktop browser
In most industries the governance concern with a copilot is personal data. In manufacturing the primary asset is the company’s process and product knowledge, and the controls follow from that.
Intellectual property
What may leave the boundary, what a provider may retain, and under what terms.
Access control
Each user sees only what they are entitled to, enforced at retrieval.
Factual accuracy
Grounding, citation, and refusal where the revision is unresolved.
Traceability
Every answer carries its source. Questions and answers retained per your policy.
Human verification
Required for anything that affects product, process or safety.
Model and vendor risk
Version changes, provider terms, exit, per section 6.
Responsible AI policies and guardrails
What the copilot must not be used for, stated, enforced where possible and known to users.
Content ownership
Named owner per content area, since answer quality is now their output.
Copilot reporting usually leads with usage, which rises during novelty and falls afterwards regardless of quality. Three other measures tell you what is actually happening.
ANSWERED CORRECTLY, against the evaluation set, re-run on every change to content, retrieval or model. Response accuracy and relevance assessed together, since a correct answer to a different question is not useful. This is the only objective quality measure available
REFUSAL CLUSTERS. Where the copilot repeatedly cannot answer, which is a precise content gap report by area and the most actionable output of the whole system
REPEAT USE BY THE SAME PEOPLE after the first month, which is the honest adoption measure. Novelty produces first use; only usefulness produces second use
Retrieval quality: whether the right document was found, measured separately from whether the answer was well written, because the two have different fixes
Citation rate, which should be effectively total, and any answer without one investigated as a defect
Query resolution: the proportion of questions that ended without the person asking a colleague afterwards, which can be asked directly in the tool
Content coverage and staleness per area, which is the content owners’ dashboard
Time saved per task type, measured on a sample rather than estimated across the population
Feedback volume and what happened to it, since feedback-driven optimisation only works while people can see their feedback changing something
Language distribution of queries, which frequently shows that the shop floor need is larger than anybody estimated
Look at the questions the copilot could not answer last month, grouped by area. That list is the content plan, written by the people who needed the information, prioritised by how often they needed it. It is the single most valuable artifact a manufacturing copilot produces, and in most deployments nobody is looking at it.
Outcomes depend on the state of your content and how much knowledge is undocumented. These are the changes the work is aimed at.
Answers come from the current revision, with the source openable by the person who asked
The copilot refuses rather than guesses, which is why people trust what it does answer
Troubleshooting time falls, measured on a sample rather than claimed
The same fault is not solved twice at two plants, because the first solution is findable
New starters reach competence faster and interrupt experienced people less
Knowledge from experienced people is captured as controlled content with their name on it
Shop-floor staff get answers in their working language from approved translations
Content owners get a precise list of what people needed and could not find
And nothing operational is answered from superseded or unapproved material
Bring us one knowledge area and the people who use it, and we will classify the content, build the fifty-question evaluation set, tell you which questions your estate cannot currently answer, and deploy on the part where the content is already good.
CaliberFocus builds manufacturing copilots on controlled content. We classify the document estate with your document control function, exclude superseded and draft material at the index, cite every answer to a document somebody can open, and design the copilot to refuse when the current revision cannot be established. Answers are shaped differently for the shop floor and the engineering office.
- Current revisions only, with superseded content excluded at the index
- Every answer cited, and a refusal when the revision is unresolved
- Separate answer shapes for operators and engineers
- An evaluation set of real questions, re-run on every change
Security & Compliance
