Human-aware AI: designing workplace assistance around people

Start with context that people can understand

A useful assistant knows which workflow is open, what information the user has supplied and which actions require review. It can respond to a failed operation, a changed task or an explicit request for help. None of these functions requires a claim that the system understands a person’s inner state.

This article sets out a design approach for workplace assistance in SupraWorx. A named event or trigger is a software interface: its presence does not demonstrate that an accurate detection model or a validated safety function exists behind it.

Interaction signals are observations, not diagnoses

The previously described „disruptive behaviour“ events include inactivity, repeated clicks and unusual interaction patterns. These labels should be interpreted narrowly. Repeated clicks can indicate a slow interface, a faulty device or an accessibility need. Inactivity may simply mean someone is reading, thinking or working elsewhere.

A proportionate response is to offer optional help with the current task or report a technical error. It would be inappropriate to treat these signals alone as proof of frustration, disengagement, cognitive overload or poor performance.

Three practical patterns

Optional help during a workflow

After repeated validation errors, an assistant could explain the required input or offer a worked example. The user can dismiss it. The trigger, the data used and the response should be documented.

Recommendations with explicit feedback

A manager reviewing a schedule can accept, edit or reject a proposal. Record that choice with an optional explanation. A rejection does not automatically reveal a hidden constraint, and a logged event does not itself mean that the model learns from it.

Context changes and handover

When an approved workflow changes from planning to incident handling, the assistant can prioritise the relevant procedures. If an action exceeds its permission or available information, it can route the matter to an accountable person.

Emotion inference requires a different assessment

The EU AI Act, Article 5(1)(f), prohibits the specified use of AI to infer emotions in workplaces and education institutions, with exceptions for intended medical or safety uses. The regulation’s definition of emotion recognition concerns inference based on biometric data. An opt-in setting alone does not remove a prohibited-use restriction.

Consequently, camera-, voice- or wearable-based emotion inference cannot be advertised as an ordinary workplace personalisation feature on the assumption that consent makes it acceptable. A safety label also does not establish that a particular system qualifies for an exception or meets other applicable requirements.

Machine safety needs demonstrated engineering

Detecting a person near a robot and safely stopping machinery is a distinct engineering problem. It needs an appropriately designed and validated safety system. A chat assistant, an inactivity trigger or a speculative interpretation of heart-rate data should not be described as preventing accidents.

In a learning article, factory and warehouse examples should be treated as scenarios for requirements analysis. They are not accounts of proven SupraTix deployments.

Trust depends on more than a helpful tone

Lee & See (2004) connect trust with appropriate reliance on imperfect automation. Hoff & Bashir (2015) describe how individual differences, context and experience influence trust. These reviews support testing whether people understand the system’s limits; they do not establish that a human override makes every automated process safe.

What to evaluate before deployment

• Can users explain why help appeared and control whether it appears again?
• Does the feature reduce a defined task difficulty, including for users with different access needs?
• How often does a trigger fire unnecessarily, and what does that cost the user?
• Are interaction data limited to an agreed purpose, with clear access and retention rules?
• Can consequential actions be reviewed by someone with enough time, information and authority to intervene?

For SupraTix, human-aware design is most useful when it makes people’s work easier without guessing who they are. Clear evidence, limited permissions and direct feedback provide a sound starting point.