What AI Should Know Before an Aviation Draft

"Prepare a compliance note for this requirement."

An AI tool will answer that in seconds, and the answer will read well. It will also have guessed at least six things: which authority applies, what kind of organisation this is, which source and revision to use, what evidence exists, who is asking, and what the note is for. A sentence that fits a Part-145 quality procedure may be wrong for a CAMO process. A study note is not an OM amendment. A finding sounds strong until the sampled record fails to support it.

The fix is not a cleverer prompt. It is the same intake a competent colleague would ask for before starting.

The six things AI needs before it drafts

1. Authority and framework. EASA, another authority, or a customer or contractual standard. Say which rule set applies.
2. Organisation type. Part-145, CAMO, operator, Part-147, or an independent professional working for a client. The same topic has a different answer in each.
3. Source and revision. Regulation, AMC, GM, the approved exposition or a local procedure, and which revision. Say whether the tool may use anything beyond what you supply.
4. Evidence. What was actually sampled or recorded, and what is only reported or assumed.
5. Your role. Preparing, reviewing or approving; for your own notes or for a company-controlled record.
6. The output. The artifact you need (finding, procedure section, matrix, CAP, reply) and who will review it.

Most weak AI drafts in aviation come from one of these being missing. The blocks below turn them into intake you can copy, fill in and paste before the request.

Intake: audit finding

TASK         Draft an audit finding for my review
AUTHORITY    EASA / other:
ORG TYPE     Part-145 / CAMO / operator / other:
REQUIREMENT  Source, section, revision (quote if you can)
AUDIT SCOPE  Area, activity, date
SAMPLE       What was sampled, how many
OBSERVED     What was seen or not found
EVIDENCE     Record refs, photos, interview notes
CLASSIFY BY  Which procedure defines finding levels
STATUS       Draft observation or intended finding?
REVIEWER     Who checks before issue
RULES        Do not classify. Mark assumptions.

Without the requirement and the sample, the right output is a list of questions, not a finding. Can AI draft aviation audit findings safely? covers what the draft should look like once the intake is there.

Intake: procedure section

TASK         Draft a procedure section for review
DOCUMENT     MOE / CAME / OM / local procedure, revision
PURPOSE      Why this section exists or is changing
SCOPE        Who and what it applies to
ROLES        Who does what, by title
INPUTS       Documents, data, forms needed
STEPS        Current practice, in rough order
RECORDS      What is recorded, where, by whom
INTERFACES   Other procedures or departments
REFERENCES   Requirements the section must meet
ROUTE        Document control and approval route
RULES        Flag assumptions and open questions

The route matters even at draft stage: it tells the tool the output is a proposal for document control, not text for the manual. More in how aviation AI prepares procedure drafts.

Intake: compliance matrix

TASK         Draft a compliance matrix, not a summary
AUTHORITY    Rule set and amendment in force
REQUIREMENTS Which paragraphs, or the full list
MANUAL       Exposition and revision to map against
COLUMNS      Requirement / manual ref / evidence /
             gap / owner / status / reviewer note
EVIDENCE     What evidence exists, and where
STATUS       Proposed only; reviewer sets final status
RULES        Leave a cell blank rather than guess

Naming the columns is what stops a matrix coming back as three paragraphs. The method behind the columns is in how to build an aviation compliance matrix.

Intake: corrective action plan

TASK         Structure a CAP for review
FINDING      Finding text, verbatim, with reference
ISSUED BY    Internal audit / authority / customer
CONTAINMENT  What was done immediately
CORRECTION   What was fixed in the sampled case
ROOT CAUSE   What is known, and how it was found
ACTION       Proposed corrective action
OWNER        Name or role, due date
EVIDENCE     What will show completion
EFFECTIVE    How and when effectiveness is checked
RULES        Challenge "staff reminded"-type wording

If root cause is still open, say so. The tool should then produce root cause questions, not a finished cause. See how to write an aviation corrective action plan.

Intake: certificate or recency note

TASK         Prepare a readiness note for my records
CERTIFICATE  Type and issuer
SCOPE        Categories, ratings, limitations
DATES        Issue, expiry, last recurrent
RECENCY      Relevant recent experience, if any
RECORD       Personal note or company record?
OUTPUT       Reminder list / CV line / renewal checklist
RULES        Do not infer validity. Mark unknowns.

The tool can organise dates and list what to check. Whether a certificate, rating or authorisation is valid is for the issuer and the rules that govern it.

Intake: reply to an authority or customer

TASK         Draft a reply for my review
FROM         Authority / customer, and their reference
ISSUE        Their text, verbatim
DEADLINE     Date the reply is due
FACTS        What has been done, with dates
EVIDENCE     What will be attached
POSITION     Accept / partly accept / clarify
SIGNATORY    Who signs and sends
TONE         Formal, factual, no commitments beyond
             what is listed under FACTS
RULES        Do not add actions or dates I did not give

This is the intake where guessing does most damage. A tool that "helpfully" adds a commitment creates an obligation the organisation never agreed to.

Ask only enough to control the risk

Nobody wants an assistant that turns every request into an interrogation. For a low-risk explanation, two questions may be enough: which authority, and who is the audience. For a procedure draft, the set is larger. For a finding, the requirement and the sample are non-negotiable.

The test for each question is whether a wrong guess would change the output in a way the reviewer might miss. If it would, ask. If not, state the assumption and move on.

A good assistant can offer the choice itself:

• "I can give a general explanation now, with assumptions labelled."
• "Or I can ask a short intake and prepare a review-ready draft."

That gives speed when speed is safe and structure when structure is needed.

What a good assistant should ask back

Even with the intake, a well-built tool should push back when something is missing or odd: a source revision that looks old, evidence that does not support the proposed wording, a request that drifts from drafting into deciding. If you ask for a finding to be closed, a manual amendment approved or an aircraft released, the right response is to stop at preparation and say who decides. Where that line sits for each kind of task is covered in the human review gate.

If you skip the intake, the output should say so: a general note with assumptions listed, not something dressed as a review-ready draft.

In Avioverse

Some of this intake is carried for you in Avioverse. Each conversation keeps its regulation boundary (EASA, FAA or both), Metis can draw on the records and context you have saved in the workbench (depending on your plan), and modules such as Audits and Procedures supply the structure a finding or procedure needs. Metis can ask when a gap matters. The draft still goes to a competent person before it is used.