Industrial AI Risk: Benson Chan's Automation–Impact Matrix

On 21 October 2026, IIoT World published the response of Benson Chan, COO of Strategy of Things, to an audience question raised during the session "From Industrial AI Ambition to Operational Reality: Choosing the Right Use Cases" at the Industrial AI Summit 2026. Chan proposed a simple way to classify the risk of AI use cases before they go into production.

The session also featured Jonathan Weiss (Cumulocity), Flemming Kongsberg (Siemens) and Maddie Zheng (Nyotta AI), moderated by Yash Titus of Strategy of Things.

The Automation–Impact Matrix

The model, called the AI Automation–Impact Matrix, relates decision severity to level of automation. Every use case falls into one of four quadrants. Chan stresses that this is one possible way to classify risk, not the only one.

The article describes only the two extreme cases:

• High severity, high automation: the AI autonomously controls temperature, pressure and flow in a safety-critical process, with no human approval. An error could cause equipment damage, hazardous releases, injuries or a plant shutdown.
• Low severity, low automation: the AI analyzes machine data and suggests when to perform routine maintenance, while a person reviews and approves. A wrong recommendation would most likely result only in inefficiency, unnecessary maintenance or avoidable downtime.

No examples are given for the other two quadrants, and the image of the matrix was not legible, so its exact layout and labels are unverified.

The key message: classify in order to mitigate

The model does not say to avoid high-risk use cases. Once the use case has been located on the matrix, mitigations are planned. Chan identifies three typical causes of incorrect AI decisions:

• improper training, or use outside the conditions the model was trained for;
• late, missing or poor-quality data;
• *model drift* when the model is not monitored and retrained periodically.

Suggested mitigations

• verify that models are properly trained and used within their intended scope;
• periodically test for drift and bias;
• make sure data sources are reliable, consistent and fully available.

If the adopting organization cannot guarantee all of this, Chan says the use case may not be suitable for a risk-averse organization.

Why it matters to automation professionals

For automation engineers and integrators, the matrix offers a simple, defensible way to scope an AI project. It helps decide where to insert a human approval step and how much effort to allocate to data quality and model monitoring. It can also serve as a basis for dialogue with the customer: before discussing technology, you discuss what happens if the decision is wrong.

First place the use case, then plan the mitigations.

Limitations to keep in mind

This is a short answer, about three minutes long, given at a conference by an executive of a consulting firm. It includes no statistics, case studies or references to standards. It is therefore a methodological prompt, not a validated procedure: it does not replace the risk and functional safety analyses already required for critical processes.

A practical takeaway

During scoping, it is worth asking two questions for each use case: how severe is a wrong decision, and how much autonomy does the AI have? The answer guides the choice between a recommendation with human approval and autonomous control, and makes explicit the data and monitoring work to include in the project from the start.