AI-Assisted Requirements Engineering and Traceability

Requirements engineering is the part of aerospace and high-end mechanical work that most students never see in school. It is also the part where bad work gets the most expensive — a poorly written requirement at the top of a program can cost tens of millions of dollars in rework downstream. AI is finally getting good enough to help here, particularly in drafting candidate requirements, finding gaps, building traceability, and flagging ambiguous wording.

This lesson teaches you how to use AI as a requirements assistant without losing the discipline that DO-178C, ARP4754A, MIL-STD-961, and similar standards require.

What You'll Learn

• What "good" engineering requirements look like and why ambiguity is so costly
• How to use AI to draft, critique, and decompose requirements
• The role of traceability and how AI helps maintain it
• A primer on DO-178C, DO-254, and ARP4754A — enough to know where AI is and is not allowed
• A workflow for using AI inside a DOORS or Jama traceability environment

What a Good Requirement Looks Like

Industry standards (especially INCOSE's Guide to Writing Requirements) converge on a set of properties. A well-formed requirement is:

• Unambiguous — one and only one interpretation.
• Singular — one requirement per statement.
• Verifiable — testable through analysis, inspection, demonstration, or test.
• Feasible — physically achievable within the budget and schedule.
• Necessary — supports a higher-level need.
• Traceable — links upward to a parent requirement and downward to design and verification.

A bad example: "The system shall be fast."

A good example: "The flight control computer shall respond to an autopilot disengage command within 50 ms, measured from command receipt to mode bit deassertion."

The good version is testable and unambiguous. The bad version is a wish.

Where AI Shines in Requirements Work

AI is genuinely useful in five places:

1. Drafting first-pass requirements from a description. You describe a feature in plain English, and the AI produces 5-15 candidate requirements in the right form.

2. Critiquing existing requirements. Paste a draft requirement and ask the AI to evaluate it against INCOSE rules. It will flag ambiguity, multiple requirements smuggled into one, and missing verifiability.

3. Decomposing high-level requirements into lower-level ones. A system-level requirement might decompose into 3-5 subsystem-level requirements. AI handles the first cut well.

4. Finding traceability gaps. Given a parent requirement and a list of child requirements, AI can suggest where coverage is missing.

5. Drafting test cases from requirements. "Given this requirement, write the verification test procedure with pass/fail criteria."

The pattern: AI is a fast junior systems engineer who can produce first drafts at scale. You still need a senior engineer to approve the final wording, especially for safety-critical items.

A Drafting Prompt You Can Use

ROLE: Aerospace systems engineering reviewer following INCOSE Guide for Writing Requirements.

CONTEXT:
\{Describe the system, subsystem, or feature in plain English. Include operating environment, expected users, performance goals, and any standards that apply.\}

TASK:
1. Produce 5-10 candidate requirements covering the described scope.
2. For each requirement:
   - Use "shall" form.
   - Make it singular, unambiguous, verifiable, feasible, and necessary.
   - Specify the verification method (analysis / inspection / demonstration / test).
3. Group requirements by category (functional, performance, interface, environmental, safety).
4. After the list, identify 3-5 areas where you believe the description has gaps and additional requirements may be needed.

I will review and refine every requirement before adding it to our authoritative requirements baseline.

The model will return a structured list. Treat it as a draft. Every requirement still needs a human owner, a verification method that you have validated is appropriate, and a traceability link upward.

DO-178C, DO-254, and ARP4754A in Plain English

If you work in civil aviation, three standards govern most safety-critical engineering:

• ARP4754A — system-level development assurance. Process for developing aircraft systems.
• DO-178C — airborne software development assurance. Five Design Assurance Levels (DAL A through E), with A being the most critical (failure causes catastrophic loss).
• DO-254 — airborne hardware development assurance for complex electronic hardware (FPGAs, PLDs).

These standards say almost nothing about the engineering content. They say a lot about process, traceability, and verification. Every requirement must be traced. Every test must show coverage. Every change must be controlled.

Where AI is allowed — drafting, summarizing, reviewing, suggesting. Anything where a human still approves the final artifact.

Where AI is not yet allowed (as of 2026) — replacing the qualified review by a human in the loop. There is no "AI signs the DAL-A requirement" path that is currently accepted. Regulators are studying this; the technical guidance is still being written.

The defensive position for your career: treat AI output as a draft, every time, and document the human approval that finalized it.

Traceability and How AI Helps

A traceability matrix links every requirement to:

• The parent requirement(s) it derives from.
• The design element(s) that implement it.
• The verification artifact(s) that show it has been met.

In a real program these matrices live in tools like IBM DOORS, Polarion, Jama Connect, or simpler systems built around spreadsheets. They contain thousands to tens of thousands of rows on a mid-sized program.

AI uses inside the traceability tool:

• "Find requirements that have no downstream verification artifact."
• "Find verification artifacts that do not trace back to a requirement."
• "Suggest links between this new design element and existing requirements."
• "Find requirements that have not been updated in the last 18 months and may be stale."

DOORS, Polarion, Jama, and Codebeamer all have AI features in 2026 — most still maturing, all worth knowing exist.

A Practical Workflow for a Student

Even if you do not have access to enterprise requirements tooling, you can practice the discipline:

Step 1. Pick a small system you understand (a quadcopter, a 3D printer, a hobbyist rocket).

Step 2. Write 10-20 system-level requirements by hand.

Step 3. Paste them into an LLM and ask it to critique them against INCOSE rules. Take the feedback seriously.

Step 4. Ask the LLM to decompose 3 of your system-level requirements into subsystem-level requirements.

Step 5. Build a simple traceability matrix in a spreadsheet linking system → subsystem requirements.

Step 6. Ask the LLM to write a verification test procedure for one subsystem-level requirement.

Step 7. Try to actually verify it on the real hardware if you can.

This workflow takes 2-3 hours and teaches you 80 percent of what an entry-level systems engineer at an aerospace company will do in their first month.

The Cultural Shift

The skill that will set you apart is not "use AI to write requirements fast". Anyone can do that. The skill is using AI to find your own blind spots — ambiguity you did not notice, gaps you did not see, verification you did not plan.

Treat the AI as a second pair of eyes for everything you write. After every draft session, paste the output back into the LLM with the prompt: "I just wrote this. Find every weak spot, ambiguity, untestable claim, and missing parent traceability."

Over a year of doing this, your unaided drafts get noticeably sharper. That is the long-term win.

Key Takeaways

• Good requirements are unambiguous, singular, verifiable, feasible, necessary, and traceable.
• AI helps with drafting first cuts, critiquing wording, decomposing high-level requirements, and drafting test cases.
• DO-178C, DO-254, and ARP4754A demand traceable, human-approved artifacts — AI assists but does not sign.
• Traceability tools (DOORS, Polarion, Jama, Codebeamer) are increasingly AI-augmented in 2026; learn the workflow.
• Use AI as a second pair of eyes on your own drafts to surface ambiguity, missing tests, and traceability gaps.