AEISFrame the engineering question
Provide the source, objective, constraints, and expected evidence. A useful prompt identifies what the engineer wants prepared or checked—not merely the desired conclusion.
Anchor the question in the approved BIM or IFC revision, the active product, and the acceptance target. Vague prompts produce vague plans; precise briefs produce reviewable work.
- Name the source revision and the product or delivery stage in scope.
- State constraints, exclusions, and the code or analysis basis you expect.
- Define the artifact or check result that would count as useful assistance.
Inspect the proposed plan
Before authorizing execution, review the Agent’s proposed plan: source scope, assumptions, tool path, implemented code or analysis basis, and exclusions. Plans are proposals, not approvals.
Reject or revise plans that skip validation, invent inputs, or jump to conclusions without a tool path the engineer can audit.
- Check that the plan uses the approved source, not an implied substitute.
- Confirm assumptions are explicit and acceptable for this run.
- Verify the tool path matches a supported AEIS product or worker stage.
Demand tool and artifact evidence
Inspect inputs, validation messages, calculations, run state, and generated artifacts. Treat a confident response without traceable evidence as incomplete.
Prefer linked tool activity and package identity over narrative summary alone. If you cannot restate which source, which run, and which file produced a claim, the claim is not ready for engineering use.
- Trace each material claim to an input, tool step, or generated artifact.
- Review warnings and partial failures, not only successful summaries.
- Keep run identity and package identity visible through the review.
What Engineering AI Assistant can and cannot do
Engineering AI Assistant can structure context, surface missing inputs, draft scoped plans, organize tool activity, and connect outputs to evidence in the active product. That assistance is valuable when it shortens setup and review time without hiding decisions.
The Agent cannot autonomously design, approve, or certify engineering outcomes. It does not silently execute protected desktop actions, and it does not replace the qualified engineer’s judgment on modelling, code interpretation, or issue readiness.
- Can: plan, question, organize context, and point to tool evidence.
- Cannot: approve, certify, or own the engineering decision.
- Cannot: invent an unsupported code scope or bypass worker approval gates.
Engineer control checklist
Qualified engineers remain responsible for acceptance, revision, rerun decisions, and final use. Build a habit of explicit control points so assistance never becomes silent authority.
Use the same control sequence for AI-assisted planning and product execution: frame, execute, verify, decide.
- Confirm source of truth and revision before any assisted step.
- Approve or reject the plan before protected execution.
- Review evidence before accepting results into design issue.
- Record who accepted the outcome and under which assumptions.
- Rerun from the corrected stage when source or configuration changes.
Common review failures to avoid
Teams lose control when they accept narrative answers without artifacts, allow unstated assumptions, or treat a green status as certification. Another failure mode is expanding scope mid-run without re-approving the source and plan.
Stop those patterns early. Assisted speed is only useful when the review trail remains strong enough for a later engineer to reconstruct the decision.
- No artifact trail for a material claim.
- Assumptions changed after plan approval without re-review.
- Worker or license failures treated as engineering success.
- Results issued from a superseded run identity.
Cloud data and model improvement
After the published effective date, content handled through AEIS cloud services may be used for AEIS model improvement by default unless an organization administrator opts out.
Eligible material is screened and held in quarantine before dataset use. Credentials, secrets, authentication and session records, payment data, security and audit records, malware, and restricted content are excluded before any dataset inclusion.
- Cloud service content is governed by the organization's model-improvement setting.
- An organization administrator can opt the organization out.
- Exclusions, sensitivity review, and quarantine apply before dataset inclusion.
Retain engineer control through issue
Close the loop by deciding accept, revise, or rerun with clear ownership. AI assistance can prepare and organize work; a qualified engineer approves the engineering outcome.
When the assisted path is complete, the delivery record should still answer the same questions as a manual path: what source, what assumptions, what tools, what evidence, and who decided.
- Keep final issue authority with the qualified engineer.
- Preserve the evidence package with the decision record.
- Use the same standard for assisted and unassisted engineering work.