Using AI Agents Efficiently

The quality of the answers an AI agent can give depends heavily on how it’s asked. This section offers a few practical guidelines for getting the most out of an agent connected to your digital twin data.

Let the AI Investigate, Not Just Retrieve Data

The most valuable use of generative AI is often not simply asking it for a value. Instead, use the AI as an analyst that can investigate the digital twin environment.

For example, rather than asking:

“What is the temperature of Pump-17?”

consider asking:

“Check Pump-17 and tell me whether anything looks abnormal.”

The second question gives the AI an opportunity to:

  1. Determine which properties are relevant.

  2. Retrieve the current values.

  3. Compare those values with available limits or expected operating conditions.

  4. Identify potentially abnormal conditions.

  5. Explain its findings in natural language.

  6. Suggest what additional information should be investigated.

Similarly, a question such as:

“Are any machines showing signs that they may require attention?”

can allow the AI to examine multiple properties and identify machines that warrant further investigation.

The AI becomes a layer of reasoning between the user and the digital twin data.

Ask Questions at Increasing Levels of Detail

A useful approach is to start with a broad question and allow the AI to progressively investigate the results.

For example:

Start broad:

“Give me an overview of the current state of our production equipment.”

Then follow up:

“Which machines are currently abnormal?”

Then:

“What is abnormal about Machine-42?”

And finally:

“Look at the other properties of Machine-42 and determine whether there are other indications that could explain the problem.”

This conversational approach is one of the primary advantages of using generative AI with MCP. The user does not need to anticipate every query required to investigate a problem.

Note

This progressive style of questioning works well because the underlying models carry the semantic metadata described in Annotating Digital Twin Models for AI. Without descriptions, units, and normal ranges attached to each property, the agent would have no basis for deciding what counts as “abnormal” at any step along the way.