A factory judges an idea by whether it survives the line. Software sold into this sector arrives promising throughput and often leaves having produced a slide, because what decides the outcome was never the demonstration — it is the drawing, the tolerance and the machine bolted down years before anyone said the word AI. Plant and engineering leaders open with the work: the manual nobody can search, the quote that eats half a day, the defect that has to be caught before the pallet ships. Four of those jobs came back often enough from manufacturers, industrial operators and aerospace suppliers that we treat them as this sector's own. For the view across every sector, visit the cross-industry catalog.
AI Use Cases in Manufacturing, Industrial and Aerospace:
The Four Jobs This Sector Actually Described
What plant, engineering and quoting teams asked us for in their own words, the gate standing in front of each request, and the point at which our software runs out.
The AI use cases in manufacturing this sector described come to four jobs, and three of them are document work:
- Get the one applicable step out of the technical documentation.
- Build the bill of materials, price it and get the quote out in the mandated layout.
- Detect the defect or the event at the edge.
- Work a supply chain disruption through to its downstream impact.
Where Iternal stops. The third of those four is not document work at all, and it is where our own limit starts. Iternal stays away from computer vision, Blockify has nothing to do with computer vision and will not work with computer vision use cases, and AirgapAI has no vision ability, so only the text of an attached document is used. Detecting the defect at the edge is therefore a job this sector described to us, not a capability we deliver. The document jobs carry their own recorded limits: AirgapAI does not currently work with images or PowerPoint files and does not handle scanned PDFs of old documents, which is exactly what an engineering drawing usually turns out to be, and the blueprint builder does not build the whole bill of materials because there is too much variability.
Three things decide whether any of it works on your floor. Which document class each job runs on, because a typed specification and a scanned print take different routes through a pipeline. What accuracy bar the output clears before it replaces a manual check, because this sector told us its expectation is one hundred percent and that ninety-nine is nowhere near it. And which system holds the answer, because AirgapAI works with files and cannot query databases or other systems.
Read the list as demand, not as shipped software. A job on this list means somebody in a plant asked for the work; it does not mean anything was built, run or measured. These are jobs rather than a tool shortlist, and for most of these four our record carries the request and no published result. For more information on which formats a pipeline takes, visit the file-format page.
Who Described This Work
Nobody drew this sector on a whiteboard. It assembled itself out of separate organizations that each sat down with us and walked through their own operation: precision job shops, a regulated aerospace manufacturer, an electronics assembler tracking every component on every board, an industrial equipment distributor, and a manufacturer supplying the defense industry. Regulated aerospace manufacturing is the segment Iternal works in by preference, because its compliance requirements are the ones a local deployment was built to answer.
What they brought was specific. First-article inspection forms that have to be right before a prototype ships. One quote taking half a day to build in a regulated aerospace plant. Part numbers and firmware pulled out of an engineering corpus. Vials scanned on the line for defects. Certificates of compliance mapped by hand to every component on a board. Put a factory photograph on a generic productivity pitch and you still would not have any of it.
The Four Jobs This Sector Described
Read the last two columns as things said to us across sales and customer meetings, not as anything an outside body measured about this industry. Where the record is silent, the cell says so:
| The job | Who described it | The gate it has to clear | Where our own software stops on this job |
|---|---|---|---|
| Get the one applicable step out of the technical documentation | Design, applications and program engineer | No registered gate. What we heard instead was the difficulty: the answer is one paragraph inside thousands of pages. | AirgapAI has no vision ability, so only the text of an attached document is used. It does not currently work with images or PowerPoint files and does not handle scanned PDFs of old documents. |
| Build the bill of materials, price it and get the quote out | Finance leader, accounts payable and procurement | Output must match a mandated form, template or standard | The blueprint builder does not build the whole bill of materials, because there is too much variability. AirgapAI works with files and cannot query databases, and small models still make mistakes on numbers and calculations. |
| Detect the defect or the event at the edge | Operations, quality and plant leadership | Latency or data volume rules out sending it anywhere | Iternal stays away from computer vision. Blockify has nothing to do with computer vision and will not work with computer vision use cases, and AirgapAI has no vision ability. Described demand, not a delivered capability. |
| Work a supply chain disruption through to its downstream impact | Operations, quality and plant leadership | Not recorded. Nobody described a binding constraint on this job. | Not recorded. Nothing in our conversations covers where our software stops on this job, either way. Left explicitly empty rather than guessed. |
Two cells are blank on purpose, and they may be the most useful thing in the table. A confident sentence about your supply chain that nobody in a plant ever said to us would be worth exactly nothing to you. Read the second column too — four jobs, three roles, no single approver.
What Makes Each Job Specific to This Sector
Industries look alike on a slide and diverge on one thing: which artifact is on the desk, and what happens downstream when that artifact is wrong.
Get the one applicable step out of the technical documentation
The document is the engineering drawing, the spec sheet and the part and firmware record; the consequence is a maintainer applying the wrong procedure. Buyers described engineers guessing exactly how content was worded in order to find it, and internal search returning nothing on a straight part lookup. Our limit attaches to the document class rather than to the question: AirgapAI reads the text of an attached document, so an image, a PowerPoint file or a scan of a decades-old print does not carry. Iternal routes heavy engineering-drawing work to a server rather than to the laptop for that reason. What a pipeline will take is settled on the file-format page.
Build the bill of materials, price it and get the quote out
The document is a first-article inspection form and a quote in a fixed layout; the consequence is a prototype that cannot ship. The AS9102 form has to be right before the prototype moves, and the quote has to return in the shape the customer dictates. Buyers described complex quotes built by hand at an hour and a half each, alternates arriving by email that all have to flow into the quote, and every line validated manually because the environment is regulated. The useful lesson is where the labor sits: the burden is reconciliation rather than keystrokes, so automating the typing alone leaves most of the cost where it was. Our limits bite mid-job — the blueprint builder does not build the whole bill of materials, and AirgapAI cannot query the system holding the pricing.
Set that against the bar this sector states for itself: an accuracy expectation of one hundred percent, with ninety-nine nowhere near it and the criterion on the floor measured in parts per million. The answer is not a higher benchmark. It is a boundary. The software assembles and arranges, a qualified person accepts, and the acceptance is recorded — which is how first-article paperwork already works. Two further recorded properties fix where that boundary sits: small models on a laptop still make mistakes on numbers and calculations, and the least predictable AirgapAI results we have seen ourselves involved spreadsheet content. Visit the back-office and quoting page and the traceable answers page.
Detect the defect or the event at the edge
The subject is the production line itself; the consequence is a defect that leaves the plant. The gate is physics rather than policy: latency or data volume rules out sending it anywhere, so inference happens at the machine. Buyers described vials scanned on the line, a point solution that will not generalise across every site and model type, and a two percent false-positive rate on a high-volume line making more scrap than the saving is worth in a plant with nobody spare to chase alarms. We list this job and we do not sell into it. A headline accuracy figure attached to a language pipeline is not a claim about a camera watching a conveyor.
Work a supply chain disruption through to its downstream impact
The document is the supply chain review report; the consequence is a downstream commitment nobody re-planned. A supplier pushes a delivery out. The dock date moves, and behind it the build slot, the customer promise and the cash forecast move too — each by hand, or not at all. Buyers described the same failure in inventory: one plant buying parts the plant two minutes up the road already holds as surplus. Iternal does have a manufacturing solution tying into ERP and supply chain for quality checks and certificates of compliance. What we hold no statement about is where our software stops, and the way to fill that cell is a real disruption from last quarter.
Agentic AI Use Cases in Manufacturing
Agentic AI use cases in manufacturing are these same four jobs run as a sequence rather than a single question: several steps, several documents, one output at the end. What this sector described is the sequence. What our record supports is the retrieval and the drafting inside it, with a qualified person accepting the result.
- Carry a technical documentation lookup end to end: Find the revision that is current, pull the paragraph that applies to the machine in front of the technician, and return the document it came from instead of a paraphrase. The sequence runs on text: a drawing image, a PowerPoint file or a scan of a decades-old print does not carry into it.
- Assemble a quote up to the numbers: Gather the specification, the alternates that arrived by email and the options into the layout the customer mandates, then hand the draft to the person who accepts the pricing. The bill of materials is not built end to end because there is too much variability, and the system holding the prices is one AirgapAI cannot query.
- Watch the line for the defect: The one sequence on this list a document agent cannot take. Catching a defect or an event at the machine belongs to a vision system, and Iternal stays away from computer vision. Listed here because the sector described it, not because we deliver it.
- Walk a supply chain disruption to its commitments: Take a supplier delay through the dock date, the build slot and the customer promise, and put the affected commitments in front of a planner in one pass instead of four. Where our own software stops on this job is not recorded, so read the sequence as demand.
Read this list the way you read the four jobs above: a sequence somebody in a plant asked for, not a shipped capability with a published result. Every step that our software does take ends with a person accepting the output, which is the same boundary this sector's own accuracy expectation already implies. For the sector overview across plant, quality and engineering operations, visit the page on AI for manufacturing.
What to Settle Before a Pilot
One word does two jobs here, and the gap between its meanings is where evaluations go wrong. Reading text off a page with a vision model is extraction, and Iternal runs that server-side and in the cloud, where extracting data from images at scale is possible today. Watching a line in real time for a defect is computer vision as a plant means it, and Iternal does not build that. Both are true at once, they describe different machines, and a buyer who hears only the first scopes a pilot that cannot land:
-
Which of our document classes runs on a laptop and which has to run on a server, and who makes that call at ingestion time?Where each engineering document actually gets processed, before anyone sizes anything.
-
Run our own engineering drawing, spec sheet and scanned print through your ingestion path, then tell us what came out and what was skipped.What an image-bearing page contributes in practice, on your material rather than on a sample.
-
For the quote, which parts does the software assemble and which parts stay with a person?The scope of the bill-of-materials job against the recorded limit, in writing.
-
What accuracy bar must this output clear before it replaces a manual check, and whose signature accepts it?The acceptance boundary, agreed before the pilot rather than argued after it.
The Work Here That Is Not Sector-Specific
Plants also run work that would look identical in a bank or a hospital. It is real here, and it is left off the sector list for a reason: swap the industry and the job does not change:
- Pull the same structured fields out of every document in the pile: the file-format page.
- Produce the document in the exact format the rule demands: the template-filling page.
- Resolve the service-desk ticket from our own runbooks: the IT operations page.
- Work the site where the signal keeps dropping out: the disconnected operation page.
The distinction pays for itself at budget time: work on that list competes with every other department for one platform, and the four above it compete with nobody.
What Would Have to Be True Here
What follows only points. Each gate gets a name and a destination, and nothing from the destination is summarized on the way past.
- Output must match a mandated form, template or standard — the template-filling page.
- Latency or data volume rules out sending it anywhere — we have no dedicated page for this one. It is the only gate on this list without a page of its own, so this is the only place it is covered.
- No registered gate — the technical-documentation lookup and the supply chain disruption job both run without one. Read that as freedom to start rather than as a guarantee: an ungated job clears no rule, and no rule protects your assumptions.
Where This Evidence Runs Out
A catalog is not a demonstration. Seeing your industry named on a list establishes one thing only: somebody in it brought this work to us. Nothing on the list demonstrates that a system was built, run and measured on a plant floor, and for most of the four jobs above our record carries the request and no published result.
Our own shortfalls are recorded alongside the buyer evidence, so they belong here too. We hold no material translating what we do into the language of a large aerospace manufacturer's production output, which is the first translation this sector asks for. We have not been able to offer genuinely different demonstrations by sector, only different wording around the same one; a long internal list of use cases and the attempts to filter it by sector did not work for the partners it was built for; and more than once we have had no demonstration of a prospect's exact use case and shown a prior one instead. One guided blueprint session here ran well past its advertised length purely because the manufacturing process was complex, which may be the most representative fact in the record. The repair does not change: run the job you need, on material you already hold.
Where to Go Next
Two of the four have a deep workflow page underneath them, and it is the same one: the quoting job and the supply chain job both sit on the back-office and quoting page, because our record treats a supply chain review as reporting work. The other two have none of their own — the documentation job routes through our manual document-work pages, and the edge-detection job has none of its own.
- Choosing which of these jobs to pilot first, and scoring the candidates against each other — see the use-case selection method.
- Comparing the cost shapes of the two ways an organization pays for AI at scale — see the cost comparison page.
- Wiring an assistant into the ERP and the systems that already hold your part data — see the enterprise integration page.
- Which machine the software installs on, once the job is chosen — see the placement decision.
- Proving any of this on documents you already hold, before a purchase order exists — see the proof-on-your-own-data page.
FAQ: AI in Manufacturing, Industrial and Aerospace Work
Four recur across this sector: getting the one applicable step out of the technical documentation, building the bill of materials, pricing it and getting the quote out, detecting the defect or the event at the edge, and working a supply chain disruption through to its downstream impact. Three are document work. The third is not, and we do not deliver it.
Detecting a defect or an event at the edge does — vials on a line, a leak, an event caught at the machine before it leaves the plant. That job is here because the sector described it to us, not because we sell it. Iternal stays away from computer vision, Blockify has nothing to do with computer vision and will not work with computer vision use cases, and AirgapAI has no vision ability.
The same four jobs run as a sequence rather than a single question: pulling the current revision and the one applicable step out of the technical documentation, assembling a quote in the mandated layout up to the numbers, catching a defect at the edge, and walking a supply chain disruption through to the commitments it moves. The third needs a vision system, which Iternal does not build, and every sequence we do support ends with a qualified person accepting the output.
Not on the laptop today. AirgapAI has no vision ability, so only the text of an attached document is used; it does not currently work with images or PowerPoint files, and it does not handle scanned PDFs of old documents. Iternal routes heavy engineering-drawing work to a server instead.
Part of it, and which part matters. The blueprint builder does not build the whole bill of materials because there is too much variability, and AirgapAI works with files and cannot query the system holding the pricing. Buyers also made a deeper point: the labor in quote-to-order is reconciliation rather than keystrokes, so automating the typing alone leaves most of the cost in place.
Buyers here told us their expectation is one hundred percent and that ninety-nine is nowhere near it, with the criterion on the floor measured in parts per million. Treat that as a boundary rather than a benchmark: the software assembles and arranges, a qualified person accepts, and the acceptance is recorded. Small models on a laptop still make mistakes on numbers and calculations, which is where the boundary has to sit.
In most plants the state lives there, and AirgapAI works with files and cannot query databases or other systems. Iternal does run a manufacturing solution tying into ERP and supply chain for quality checks and certificates of compliance. What we hold no statement about is where our software stops on this job, so that cell above is left explicitly empty.
Run One Job on Your Own Documents
Take the one job here that would give a week back to somebody in your building, name the gate it has to clear, and run it on material already on your own drives. That sequence answers what no sector list can: whether the work holds up in your formats, at your tolerance.