AI in engineering: evidence, review and deployment

Overview

Choose an engineering task with a defined input, output and review owner before adopting AI. Document summarisation, retrieval and draft preparation can help teams find information, but generated text must remain distinguishable from accepted engineering evidence.

Limit retrieval to approved documents and the user’s authorised project scope. Preserve source names and revision information. Treat instructions embedded in uploaded documents as untrusted content. Require explicit approval before sensitive project material is sent to an external model provider. Keep provider credentials outside browser code and public downloads.

Evaluate outputs with representative project examples and record failure modes, missing evidence and review effort. Engineering calculations need stated units and assumptions and independent verification. Establish who can publish an AI draft, how corrections are stored and when the system should decline a question because evidence is missing.

An evidence test set should contain routine questions, questions with conflicting documents and questions for which the approved material contains no answer. Check whether the system finds the right revision, identifies uncertainty and refuses to invent missing information. Include documents containing misleading instructions to verify that document text is handled as evidence rather than as an instruction to the assistant.

For draft preparation, separate extraction from interpretation. An extracted quantity should point to the relevant document or worksheet; a suggested conclusion should state its assumptions and identify the reviewer. A BOQ draft needs checks of units, duplication, scope boundaries and excluded items. An image or drawing needs a supported interpretation workflow; text retrieval alone does not establish that geometry has been read correctly.

NIST provides a voluntary AI Risk Management Framework for considering risks during AI design and use. Apply a project-specific evaluation rather than presenting the framework as a certification. Define when to require human review, who can release a result and how an employee reports a wrong answer. Retain the approved source revision with the final reviewed output.

Further reading

NIST AI Risk Management Framework

Key considerations

• Define the operating question and asset boundary.
• Record input quality, units, ownership and revision.
• Agree review responsibilities and acceptance criteria.
• Verify project-specific requirements before implementation.

Related solutions

Predictive Asset Intelligence

Infrastructure Condition Analytics

Computer Vision & Video Analytics

Product capabilities

• Define the operating question and asset boundary.
• Record input quality, units, ownership and revision.
• Agree review responsibilities and acceptance criteria.
• Verify project-specific requirements before implementation.