AI software integrates with existing systems by reading from them through their native interfaces, such as database views, OPC UA, REST APIs, or file drops, running the model alongside rather than inside them, and writing results back to where people already work. No system needs to be replaced. The AI layer sits next to the current stack and is switched off as safely as it is switched on.
The most common objection to AI in an operating business is not accuracy or cost. It is the fear that adopting it means replacing the ERP, rewiring the line, or retraining everyone on a new interface. That fear is reasonable, because many vendors do sell it that way.
It is also unnecessary. Well-designed AI software is a layer, not a replacement. This guide covers the integration patterns that work across office and plant systems, how to choose between them, and where integrations typically go wrong.
What does AI integration actually mean?
An integrated AI system has three connections: an input path that brings data from the source system, a processing step where the model runs, and an output path that delivers the result to the place where a decision or action happens. Everything else, including the model itself, is an implementation detail.
The mistake is to treat the model as the project and the connections as plumbing. In practice the connections take most of the effort and decide whether the system gets used.
Integration patterns by system type
| System | Typical input path | Typical output path |
|---|---|---|
| ERP (SAP, Dynamics, Netsis, Logo) | Read-only database views, scheduled exports, or vendor APIs | Write-back through the API, or a recommendation the planner confirms |
| MES and SCADA | Historian queries, OPC UA subscriptions | Dashboard tile, alarm, or a tag the line reads |
| PLCs and controllers | OPC UA, Modbus, or a gateway that mirrors tags | A tag the PLC polls, never direct control without a safety review |
| Cameras and vision hardware | GigE or USB stream, or the vision system output | Pass/fail signal, reject trigger, image with overlay |
| CRM and sales tools | Native API or webhook | Enriched record, task, or draft reply |
| Spreadsheets and file shares | Watched folder or scheduled read | A new file in the same folder, or a filled template |
The pattern is the same every time: read through the interface the system already exposes, and write back in a form the existing workflow already understands.
Where should the AI run: on-premise, at the edge, or in the cloud?
- At the edge, on a device next to the line, when the decision must happen in milliseconds or the network cannot be trusted. Visual inspection is the standard example.
- On-premise, on a server inside the plant or office network, when data cannot leave the site or the systems it connects to are internal only.
- In the cloud when the data is already there, the latency requirement is loose, and the workload varies enough that elastic compute saves money.
- A mix is normal: an edge device makes the fast decision, an on-premise service aggregates results, and a cloud dashboard shows management the trend.
Five integration mistakes that stall AI projects
- Building the model on an export and discovering at rollout that live access was never arranged.
- Giving the AI direct write access to a control system or master data without a human confirmation step for the first months.
- Putting the output in a new dashboard nobody opens, instead of the screen people already watch.
- Ignoring failure modes: what the system does when the source is down, the data is late, or the model is unsure.
- Treating IT and OT as one team. Plant networks, office networks, and their owners have different rules, and both need to be in the room.
Frequently asked questions
- Do we need to replace our ERP or MES to use AI?
- No. AI software reads from and writes to the systems you already run through their existing interfaces. Replacing core systems for the sake of AI is almost always the wrong order of operations.
- Can AI connect to old PLCs and machines?
- Usually, yes. Most PLCs from the last twenty years expose tags through OPC UA, Modbus, or a vendor protocol, and older ones can be mirrored through an inexpensive gateway. The AI reads the tags and writes recommendations or a result tag, with direct control only after a safety review.
- Does our data have to go to the cloud?
- No. Edge and on-premise deployment keep data inside the plant or company network. Cloud is one option among three, chosen on latency, data policy, and cost rather than by default.
- How long does integration take compared to building the model?
- Often as long or longer, which is why it should start in the first weeks of a pilot rather than after the model works. Data access, network rules, and output placement are the usual sources of delay.
Have systems you cannot replace?
Neither can most of our clients. We integrate with the cameras, machines, and software you already run and stay involved until the results hold in production.
Discuss your stack