A press-brake manufacturer in Tashkent sells machines that cost more than an apartment. The buying committee has never seen one run. The sales engineer has a forty-second phone video, shot vertically in a badly lit hall, of a machine that has since been repainted and moved to a second site. Every tender deck in the sector looks the same: a specification table, three photographs of a unit still on its pallet, and a paragraph about being a reliable partner since 2011.
That is not a marketing problem. It is a production problem, and it is the kind AI video is genuinely good at: not the corporate film, but the library of explainers a machine sale quietly depends on. What the machine does. How a changeover works. What the service interval looks like. What an operator must never do with their hands.
Manufacturing is the hardest vertical to generate for honestly, because a frame that looks plausible and is mechanically wrong is worse than no frame at all. Having built these jobs, here is where the technique earns its place and where it will get a manufacturer into trouble.
The films a factory actually needs
Four, in the order they are requested and the order they return value.
A capability film, sixty to ninety seconds, answering what a procurement committee asks first: what range, what materials, what cycle time, what footprint, what certifications, what service response. Not a montage over music. A montage tells a buyer nothing, and industrial buyers watch with the sound off.
Machine explainers, forty-five to ninety seconds each, one per model or one per function. This is where generation beats a camera: a cutaway through a housing, a slowed cycle, an exploded view of a spindle assembly, a section showing the coolant path. You cannot film the inside of a gearbox that is bolted shut, and a traditional animation budget can take eight weeks.
Service and changeover procedures, chaptered, three to eight minutes. A die change, a tool change, a calibration, the cleaning sequence before a product switch. These are watched by a technician standing at the machine in gloves, which dictates the whole design: big close-ups, no elegant camera moves, a visible count of steps.
Operator and safety induction, four to ten minutes. This is the highest-volume asset in the library and the one remade most often, because a guard design changes, a lockout procedure is revised, or a new line is commissioned. It is also where multilingual delivery matters most, and where production budgets quietly collapse.
Showing a machine that has not shipped yet
The most valuable output is not footage of a machine that exists. It is a visual of a machine that exists only as a drawing: the unit on the stand at a trade fair, the line being installed next quarter, the configuration a customer is being asked to sign for.
Generated context around an accurate machine model works well. Gallery, concrete floor, extraction ducting, a technician in the customer's colours, daylight from the loading door. What must not be generated is the machine itself. The honest split is engineering animation for the mechanism, generative tools for the world around it, and photography for anything you may be asked to prove later.
The trade-fair loop is the clearest example. Ninety seconds on a screen behind the stand, three configurations and the changeover between them, running silently, subtitled, in the language of the fair. It costs a fraction of shipping a second machine, and it answers the question a visitor asks while waiting for a colleague to arrive.
Technical accuracy is the entire job
A generator will happily produce a lathe turning the wrong way, a guard that is not there, a robot arm passing through a fixture, or a welding torch held at an angle no welder would use. To a buyer who runs that machine every day, one wrong frame is an instant credibility loss, and it travels: a bad shot in a tender deck circulates through an industry faster than any case study.
So the input is not a prompt, it is a document set: general arrangement drawings, the data sheet, the parts book, photographs of the actual unit from four sides, and a note on the one detail the sales team always points at. Before a frame is generated we fix a glossary of thirty to sixty terms in the customer's own words rather than the catalogue's, and we get a sign-off from the engineer who will have to stand behind the film.
Then the review rule: every frame showing a mechanism is checked against the drawing, not against taste. The list of mechanical facts that must be true in every shot is short and boring. Rotation direction. Feed direction. Guard presence. Number of stations. Colour of the safety circuit. Sequence of the cycle. Put it in a checklist and run it before the client sees anything.
Where a render becomes a liability
Three places, and none of them is a matter of style.
Certification. If a machine carries a CE mark, an ISO certificate or a local conformity document, the film must not imply a rating, a capacity or a safety feature the certificate does not cover. A generated shot of an operator reaching into an unguarded area, on a page making a compliance claim, is a document you produced and can be asked to defend.
Tender and warranty evidence. Procurement asks for evidence of facility and installed base. A render presented as a photograph of your plant is a disqualification risk, and in some markets a misrepresentation one. Label illustrative footage as illustrative inside the frame, not in a footnote. Keep real photography in the same deck. Never let a render stand where a site visit is expected.
Customer confidentiality. Machine builders routinely work under NDA on a client's product or process. A generated frame can leak a fixture, a part geometry or a customer logo that was never cleared, and generation makes that easy to do by accident, because the model will fill a background with something it has seen before.
None of this argues against generation. It argues for a scoped brief: this film shows the machine and the process, it makes no performance claims, and the compliance language is written by your quality manager rather than by the studio.
The floor, in three languages
A plant in Uzbekistan typically runs induction in Uzbek and Russian, with English for the OEM's service engineers and often a fourth language for a joint venture partner. Filming that three times is a budget nobody approves. Recording it three times is a day of studio work and a marginal cost once the visual master exists.
The practical approach: one visual master, no on-camera presenter, voice-over in each language, burned subtitles, and a chapter menu that matches the printed procedure. Lip-sync is worth the extra work only where a person is explaining rather than narrating. An induction video does not need a face, and a synthetic face speaking a language it does not know is a distraction on a shop floor.
Keep the master editable chapter by chapter. Procedures change more often than campaigns, and a library you cannot update cheaply becomes the part of the system that stops being true.
Timing, formats and what to send
A realistic schedule for a sixty to ninety second machine explainer, once drawings and photographs are in hand: two to three days for the scene list and script, four to seven working days for generation and assembly, two rounds of revision inside the original estimate. A four-language induction with chapters runs longer, and the review is the slow part, not the rendering.
Deliver both orientations. Vertical for Telegram, WhatsApp and shop-floor phones. Horizontal for tender screens, trade fairs and a laptop in a meeting room. Burn subtitles into the vertical cut, because sound is off most of the time in a factory.
To start, send one machine, the drawing set, four photographs of the real unit, your brand book, the certificate list, and the date the film has to be finished. That is enough to return a scene list and a fixed price.
KAMOD is an AI video studio in Tashkent working with manufacturers and exporters worldwide. Send the machine and the drawings, and we will tell you honestly which shots should be filmed instead of generated.