Home/Technical Resources/AI-assisted engineering example

A practical AI assurance example

AI can suggest a test link.
What makes it evidence?

Follow a requirements review from a plausible AI draft to a traceable engineering decision.

The useful question is not whether the output looks convincing. It is whether the review can find an omitted requirement, incomplete test coverage or a result that does not apply to the current design.

Original illustrative example. The requirements, test extracts and simulated AI draft below are invented for this guide. Numerical values are not recommended design values. This is not a client case, model benchmark, qualified-tool demonstration or approval.

The agreed use of the tool

Suggest relationships.
Keep decisions accountable.

The AI assists the engineering process. It is not installed in the aircraft.

Permitted task

Propose candidate links between a fixed source packet of requirements and tests. Identify supporting source locations and possible gaps.

Release condition

A defined engineering review checks the complete requirement inventory and every disposition before the traceability record is released. The AI cannot mark a requirement verified.

One source packet. Three views.

Look at what the evidence actually says.

A hypothetical flap-monitor controller. Current baseline: FM-REQ Rev B and FM-CFG Rev B. A real review would use the complete applicable specifications and evidence.

Four requirements

FM-101 / Rev B

Position discrepancy

With both position inputs valid, if the absolute left/right position difference exceeds 5 degrees continuously for 200 ms, the controller shall set MOTOR_ENABLE to FALSE within 50 ms after that 200 ms interval ends.

FM-102 / Rev B

Invalid position input

If either position input remains invalid continuously for 100 ms, the controller shall set MOTOR_ENABLE to FALSE within 50 ms after that 100 ms interval ends.

FM-103 / Rev B

Inhibit retention

After MOTOR_ENABLE has been set to FALSE by FM-101 or FM-102, the controller shall retain that inhibit when the initiating condition clears, until a maintenance reset is accepted.

FM-104 / Rev B

Airborne reset

The controller shall reject a maintenance-reset request while the aircraft ground-state input indicates airborne, and shall retain the inhibit.

Three test extracts

TC-11 / Rev AProcedure only

Set both positions valid. Introduce a 6-degree difference for 500 ms. At 500 ms, check that MOTOR_ENABLE is FALSE.

No execution result or output-transition timestamp is supplied.

TC-12 / Rev AExecuted on FM-CFG Rev A

Set the left position input invalid for 200 ms. Check MOTOR_ENABLE at the end of that interval. Recorded outcome: FALSE.

No right-input case or timing trace is supplied. The report identifies the earlier configuration.

TC-13 / Rev BProcedure only

With the ground-state input indicating on ground, create a position discrepancy, confirm inhibit, clear the discrepancy and issue a maintenance reset. Check that the inhibit clears.

No observation between fault clearance and reset, no airborne-reset case and no execution result are supplied.

Start with the inventory.

There are four input requirements. A review of only the rows returned by the AI could miss an entire obligation.

Make the checking process specific

“A person will review it”
needs a method.

  1. Reconcile the complete inventory. Start from the input requirements, including anything the AI omitted.
  2. Read the original source. Check conditions, timing, exceptions and the exact procedure or result being credited.
  3. Separate plans from results. A proposed test, an execution record and a reviewed verification conclusion are different things.
  4. Check the baseline. Record which documents and configuration support each claim and assess changes.
  5. Give each open item an owner. Keep unsupported dispositions open and retain the rationale for decisions.

Evidence for the engineering process

Retain enough to understand
and challenge the decision.

Intended use

Task, users, allowed output, reliance limits and the defined checking process.

Input manifest

Document revisions, source locations, requirement count and unreadable or omitted material.

Run record

Available model identifier, date, prompt revision, settings, supplied sources and raw output.

Review record

Clause coverage, exact evidence locations, accepted and rejected links, rationale, reviewer and date.

Action register

Gaps, owners, resolution records and restrictions on releasing or relying on the output.

Change assessment

Reassessment after changes to the sources, model, prompts, extraction, retrieval or review method.

Take the method into a review.

Download an editable text worksheet containing the example dispositions and a blank review record.

FAA context

Define the use.
Establish the assurance needed.

The FAA AI roadmap identifies potential uses in requirements traceability while recognising the risk of inaccurate or incomplete generated material. Responsibility for the product and its compliance evidence remains with the applicant. This example is a proposed bounded workflow, not an FAA-approved procedure.

For software-life-cycle use, AC 20-115D recognises DO-178C and DO-330. Determine any qualification need from the actual use, possible errors, reliance and process credit. A generic promise of human review does not establish an exemption. This example makes no TQL determination.

Before routine use, evaluate representative cases with known omissions, incorrect links, stale revisions and partial coverage. Check both tool behaviour and whether the specified review detects those errors. A small demonstration does not establish a general error rate.

Using AI in your engineering workflow?

We can help define the task, the controls and the evidence needed before your team relies on the output.