On-Premise AI AOI: Why Board Images Should Stay in Your Plant

On-premise AI AOI means image capture, inspection, model training and data storage all happen on hardware inside your plant, with no connection to a vendor's cloud. Board images are your customers' design data, often covered by NDAs and, for defense work, by export-control rules, so the safest place for them is the machine that took them. DaoAI machines work this way by default: no cloud, no account, no internet connection required, and model training runs offline on the machine.

Why are AOI board images sensitive data?

A high-resolution image of an assembled board shows the layout, the component choices, the part markings, the revision label and often the serial number. At 10 to 15 µm per pixel, the fine detail is readable. Across a whole program, that is close to a visual bill of materials for the product.

The inspection data around those images is sensitive too. Defect history, first pass yield and false call rates describe how well your process runs, which is commercially sensitive for any EMS provider.

For a contract manufacturer, most of that information does not belong to you. It belongs to the OEM whose board is on the conveyor, and you hold it as a custodian under their terms.

What do customer NDAs and quality agreements usually require?

Terms vary, but many NDAs and supplier quality agreements share a few patterns. They limit use of design data to manufacturing the customer's product, restrict disclosure to third parties, and require notice or approval before data goes to a subcontractor or outside service.

Supplier security questionnaires then ask where data is stored, who can access it and in which country. A cloud-processed AOI turns the AOI vendor, and that vendor's cloud provider, into a party that handles customer images. That has to be disclosed, and some customers will say no.

An AOI that never sends images out of the plant removes that question from the questionnaire. The data stays inside the boundary you already manage for your CAD files, travelers and test logs.

How do ITAR and defense-style rules affect AOI data?

In the United States, ITAR controls technical data related to defense articles, and making that data accessible to a foreign person, including through servers or cloud storage, can count as an export unless specific conditions are met. Images and inspection programs for a defense assembly may qualify as technical data. Whether they do is a decision for your export compliance officer.

Defense supply chains add further rules for controlled unclassified information, such as NIST SP 800-171 and CMMC requirements, which govern where that information is stored and processed. In Canada, the Controlled Goods Program sets similar expectations for who may examine controlled goods and related data.

DaoAI does not claim any ITAR, CMMC or Controlled Goods certification, and nothing here is legal advice. What an on-premise architecture does is keep images and models inside the facility you already control, so your existing compliance program covers the AOI instead of needing a new exception for it.

What should "on-premise" actually mean for an AI AOI?

Some systems describe themselves as on-premise while still depending on a vendor connection for licensing, training or updates. Use this checklist when you evaluate any AI AOI, including ours.

How does AI model training work without the cloud?

Many AI inspection products improve in one of two ways: the vendor retrains a shared model and ships it, or the customer uploads images so the vendor can retrain for them. Either way, your images or your improvements live in a dataset you do not hold, and fixes arrive on the vendor's release cycle.

DaoAI starts from a visual foundation model trained on over 1,000,000 real production images, then adapts it to your board locally. When an operator marks a false call at the review screen, that example goes into the product's dataset on the machine, and an AI Update retrains on it there.

For defects specific to your products, model training lets your team collect images, label them OK or NG, train, evaluate on held-out images and apply the model, all inside the inspection software without an internet connection, cloud service or online account.

How do you run AI AOI on a network-isolated line?

An isolated line still needs programs going in and results coming out. The practical pattern is to keep every one of those flows on your own network or on media you already control.

• Programming: DaoAI supports offline programming. Program a product on a machine away from the line, export it, and import it on the production machine.
• Training: collect and label images, train on a separate machine, then transfer the finished model to the production machine under your media-handling procedure.
• Results: send CSV results and images to your MES over the plant network. Nothing extra has to be installed.
• Access: use operator and admin roles so tuned settings are not changed by mistake, and keep the permission list in your access-control records.
• Support: leave remote support off to keep the line fully isolated, or enable it for a specific session under your own change control.
• Retention: decide how long inspection images are kept, and make that match what your customer agreements say.

Because every DaoAI machine runs the same core software, these rules apply the same way on an offline P1 beside the line and on an inline 3D P5 in it.

Is cloud-based AI AOI ever the better choice?

Cloud processing has real advantages: the vendor handles model updates, and a fleet dashboard across several plants is easy to build. For a single-product OEM that owns its own designs, that trade can make sense.

For an EMS provider holding many customers' designs, the calculation usually runs the other way. Every customer's NDA, questionnaire and export rule applies to wherever the images go. For a broader comparison of approaches, see AI AOI vs. traditional rule-based AOI for EMS.

A system that learns on the machine also keeps the improvement where it belongs. DaoAI's feedback learning puts each correction into a labeled example in your dataset, with the image attached, rather than into a request to a vendor.

FAQ

Does DaoAI AOI need an internet connection?

No. DaoAI machines run on-premise with no cloud, no account and no internet connection required; inspection, feedback learning and model training all run on the machine. Remote support is optional and off unless the customer enables it; the system has no backdoor access.

Do board images ever leave the plant?

Not for inspection or training. Board images stay on your machine and in the systems you export them to, such as your MES.

Who owns an AI model trained on our boards?

You do. A model your team trains on your production images belongs to you and is not shared with other DaoAI customers.

Is DaoAI ITAR or CMMC certified?

DaoAI does not claim ITAR, CMMC or other defense certifications. Its on-premise design keeps images and models inside your facility, so your own compliance program can cover it.

Can we program and train away from the production line?

Yes. Programs can be created offline and imported to the production machine, and models can be trained on a separate machine and transferred.

Want to check how it fits your security requirements? Book a demo with one of our engineers, or send us one of your own boards and we will test it for you, free of charge.