Build vs. buy: industrial AI vision system
By Tommy Jonsson, CEO, Byte Motion · Updated 2026-08-26
Build vs. buy means choosing between hiring and training an internal team to develop AI vision and robotics software, or engaging a vendor that delivers a proven platform plus engineering and support. The right choice depends on how variable your product is, how fast you need a reliable system running, and whether vision expertise already exists on your team.
What "build" actually costs
Building an in-house AI vision system is rarely "hire one machine learning engineer." A working production system needs several specialisms that seldom sit on one person's desk: camera and lighting selection (the image-quality problem that decides whether any model can succeed at all), a data collection and labeling pipeline for the product it's meant to inspect or handle, model training and evaluation, and integration with the PLCs, robots, and conveyors already running the line. Each is its own discipline, and most factories do not have all of them in-house on day one.
Time-to-first-working-system is the real cost driver, more than any single line item. Camera and lighting setup alone often takes several iterations before images are clean enough to train on. Collecting a dataset that actually covers the variation the model needs to see — seasonal change, natural product variation, rare defects — takes longer than a first estimate usually assumes. Getting a model from "works on a laptop" to "runs reliably at line speed" is a separate project again, and it includes a question many first attempts skip: what should the system do when it isn't sure? Deciding whether an uncertain item gets rejected, flagged for a person, or passed through — and wiring that decision into the PLC or robot logic — is engineering work in its own right, not an afterthought once the model works.
The burden does not end at go-live. Products drift, packaging changes, and lighting in the hall shifts with the seasons — a model trained on last year's images degrades quietly unless someone is watching for it and retraining on new examples. That ongoing ownership is the part most often missing from a first build estimate.
None of this makes an internal build the wrong choice — plenty of vision projects are built successfully in-house. But an honest accounting includes the specialist hiring, the iteration time before a first working system, and the standing maintenance commitment, not just the cost of training one model. The real risk of the wrong choice is a stalled or quietly bypassed system and a second attempt with a year already spent — which is exactly the outcome a feasibility step is designed to prevent.
What "buy" actually costs and de-risks
Buying shifts the cost profile: instead of assembling the specialisms internally, you pay for a vendor's platform and engineering time, and the biggest risk — does this even work on our product? — is retired early, before a production-scale commitment.
Byte Motion structures this as a pre-study first. It is a fixed-scope feasibility phase, typically one to three weeks, in which images are collected from your real line, a limited dataset is annotated, a simplified AI model is trained, and detection performance is measured against your acceptance criteria. Cost is contained — around €10,000, excluding VAT, travel, special hardware, and third-party costs — precisely because its job is to answer one question with evidence before anything larger is committed.
If the pre-study shows the task is feasible, the next step is a Production Cell: a full Optimus installation with the camera, AI model, an industrial PC, PLC or robot communication, and a production dashboard. Installation typically runs €10,000–€30,000 depending on hardware, integration scope, and any custom development, and includes 15 months of warranty and license. After that period, the license renews from €500 to €1,500 per month depending on deployment scale and support needs — the ongoing cost of the model retraining, monitoring, and camera/PLC support an internal team would otherwise carry itself. See the full pricing breakdown for what each stage includes.
That structure de-risks two things at once: the capital decision (you see feasibility evidence on your own images before committing to a full installation) and the maintenance burden (support and model updates are part of the license rather than an open-ended internal responsibility).
When building in-house makes sense
Building in-house is a genuinely good choice under a specific set of conditions, and it would not be honest to argue otherwise. It fits large organizations that already run a dedicated AI vision or robotics engineering team and want that capability permanently in-house rather than as a project cost. It fits a single, very specific, largely unchanging use case run at a scale where amortizing a small internal team over several years costs less than an equivalent license over the same period. It also fits companies that want full ownership of the trained models and the underlying training data, with no vendor relationship in the decision chain at all.
The trade-off does not disappear, though. An internal team still has to solve camera and lighting selection, data pipelines, PLC and robot integration, and long-term model maintenance itself — building removes the vendor relationship, not the underlying engineering work. If the use case is genuinely narrow and stable, and the team hiring for it will stay on staff for years rather than one project, that ongoing internal cost can be the more economical path over the long run.
When buying makes sense
Buying makes sense when the product is variable or the environment is unstructured — natural products, deformable packaging, mixed waste — where the pattern that separates good from bad only becomes clear from real production images, and working that out in-house takes exactly the specialist time and calendar months discussed above. It also fits companies that need production reliability quickly and do not currently run a vision team, and situations where keeping existing cameras, robots, and PLCs in place matters more than owning the model-training pipeline outright.
Byte Motion's deployments show what a bought outcome looks like at production scale. At MOWI, one of the world's largest salmon producers, Optimus inspects 100% of outbound pallets (MOWI ASA, deployed 2021) for overhang, broken boxes, and box count, with every pallet photographed for a distribution network spanning more than 25 countries. At IVAR's waste facility in Norway, an AI vision and robot system detects batteries of varying shapes, sizes, and colors in mixed waste with over 95% accuracy, and fire incidents linked to battery contamination dropped to zero within the first months of operation — a task whose material variability makes an internal build materially harder to reach reliably (battery-sorting case study). And in the sausage-picking deployment, a food producer had already bought a vision system from another vendor — it struggled with accuracy, speed, and rotation, and product was sometimes destroyed during picking. Byte Motion replaced only the cameras and the AI model; the six existing ABB IRB 360 FlexPicker robots and three conveyors stayed exactly in place. That is the buying argument in miniature: the fix was a different decision mechanism, not a rebuilt line.
Camera- and robot-agnostic delivery is what makes that kind of upgrade possible rather than a rip-and-replace project. Optimus is built to connect to the equipment a factory already runs, add the intelligence it lacks, and leave the rest of the line alone.
How long does a pre-study take?
A Byte Motion pre-study typically takes one to three weeks. It covers image collection from your real production line, annotation of a limited dataset, training of a simplified AI model or prototype pipeline, and a feasibility report with a recommendation on whether to proceed to production.
The pre-study is deliberately scoped small — around €10,000, excluding VAT, travel, special hardware, and third-party costs — because its purpose is evidence, not a finished system. It ends with a camera and lighting recommendation, a technical feasibility assessment, and a budget estimate for the next phase, so the decision to invest in a full Production Cell is made against your own data rather than a vendor's sales pitch.
What happens if we already have cameras or robots?
In most cases, existing cameras, robots, and PLCs stay exactly where they are. Optimus, the platform behind Byte Motion's deployments, is camera-agnostic and robot-agnostic, integrating over ROS2 and standard industrial protocols rather than requiring a full line replacement.
The clearest example is the sausage-picking deployment: the six existing ABB IRB 360 FlexPicker robots and three conveyors were kept exactly as they were, and only the cameras and the AI model changed. At MOWI, the vision system integrates with existing AGVs, a labeling robot, and the ERP system rather than replacing any of them. Buying an AI vision system, in practice, usually means upgrading the decision layer on top of equipment you already own — a smaller and faster project than most build-vs-buy comparisons assume before the equipment inventory is actually checked.
If your equipment inventory and product variability are the open questions, the next step is a pre-study on your own images. Contact Byte Motion to scope one — we reply within one business day — and review pricing for each stage first.