CitadelAI
The only aviation AI tool you need.
CitadelAI answers the regulations from the rule text and hands you the passages behind the answer — then reads your own manuals beside them, works your safety data, and drafts the document at the end of it. The assistant is called Chuck.
How do alcohol limits for flight crew differ between the FAA, EASA and the UK?
Chuck
- FAA — no person may act as a crewmember within 8 hours of consuming alcohol, or with an alcohol concentration of 0.04 or greater. That is the rule itself, not a means of complying with it.114 CFR § 91.17
- EASA — the binding requirement is that the operator holds a policy on psychoactive substances; it sets no figure.2CAT.GEN.MPA.170 The numbers sit one rung down, in the AMC: operator instructions should be no less restrictive than 8 hours before reporting time and a blood alcohol level below 0.2 per thousand.3AMC1 CAT.GEN.MPA.100(c)(1)
- UK — a person must not act as flight crew if they know or suspect their condition renders them unfit, including by reason of the effects of any psychoactive substance.4ANO 2016, art. 166 Ramp inspections include alcohol testing on random selection or reasonable suspicion of being in excess of the prescribed limit.5ARO.RAMP.105 The corpus refers to that limit but does not define it — the figure is set in primary legislation I do not hold, so I will not quote one.
Note the weight: the FAA figure is binding regulation, the EASA figures are AMC — a means of compliance, not the rule. The UK limit I have declined rather than guessed.
Ask it something else
What it does
One place for the thinking work of a compliance department.
Not a chatbot bolted to a rulebook. CitadelAI does the reading, the comparing, the analysis and the drafting — and shows its working on every part of it that has a rule behind it.
Answers questions
Ask what a rule requires and get the answer built from the rule text, with the passages it used shown beside it. Where the corpus does not support an answer, it says so instead of inventing one.
Tackles problems
The awkward ones that span two rulebooks, three departments and a manual nobody has revised since the last audit. Chuck works them through rather than returning a definition.
Analyses data
Read across a body of occurrence reports, audit findings and hazard logs, and get the themes a case-by-case read cannot see — with the denominator stated and the limits admitted.
Simplifies complex work
Turn a requirement set into a worked grid, a 200-page investigation into a summary, and a rough narrative into the report the authority expects — in the format you have to file it in.
Holds the project together
Group the documents, chats and grids for one audit, revision or investigation, and share them with the colleagues working it with you. The context does not reset every session.
In use
How you can use Chuck.
Six real questions, the sources each one reaches for, and the shape of what comes back — a cited answer, a worked grid, a drafted section, a set of themes.
Fatigue management under Part-NCC
EASA · answered from the rule text
Part-ORO states its own scope, and it reaches non-commercial operations with complex motor-powered aircraft — so the management-system requirements are yours.
ORO.GEN.005Subpart FTL states a narrower scope: aircrew assigned to commercial air transport with aeroplanes. It does not bind your operation.
ORO.FTL.100That management system has to identify hazards, evaluate them and manage the associated risks. Fatigue is one of those hazards.
ORO.GEN.200(a)(3)
Sources
- ORO.GEN.00538%
Regulation (EU) 2024/1111
- ORO.GEN.200(a)(3)34%
Regulation (EU) No 965/2012
- ORO.FTL.10028%
Regulation (EU) 2024/1111
Prompt
We're a Part-NCC operator taking on a second base. Do we need a fatigue risk management system, or is that only for commercial air transport? Show me the provisions that decide it, not a summary.
Reads
- Regulation corpus — EASA
Ask the rules
Put the question in the words you would use with a colleague. The answer is built from the rule text and hands you the passages behind it — reference, source, and the amendment that passage belongs to, weighted by how much each one informed the answer. Where the corpus does not support an answer, Chuck says so rather than producing a rule that does not exist.
Illustrative. The provisions are real; the scenarios are written for this page.
The difference
Retrieval built for regulation, not for documents in general.
Aviation regulation is structured, versioned and ranked by legal weight. General document retrieval treats it as prose. Six places where that difference is built into the engine rather than promised in the copy.
Chunked at the regulation's own boundary
Conventional retrieval splits a document every few hundred tokens, which cuts provisions in half and staples unrelated ones together. The corpus is split at the regulation's own unit instead — one section, one chunk — and each chunk's identifier comes from its structural position rather than being assigned. A provision keeps its identity across re-ingestion, which is what makes amendment tracking exact rather than approximate.
It drafts the rule it is looking for, then searches for that
A pilot asks how long they can be on duty. The rule speaks of maximum daily flight duty periods and a numbered table. Nothing lexical connects the two and the embeddings do not close the gap either — so before searching, the engine writes the provision it expects to find, in the corpus's own register, and searches with that instead.
Legal weight ranks the results
A second-stage model reads your question and the provision together, so an entry that merely shares vocabulary with the question ranks below the one that actually sets the limit. The aviation precedence ladder is then applied as a multiplier — binding regulation above binding standard, above AMC, above guidance. A nudge, never an override: hard-sorting by weight would bury the answer you needed.
Amendments are detected, not asserted
Identifiers are structural and hashes are taken of the verbatim text, so re-ingesting a framework classifies every provision by comparison: same identifier and same hash is unchanged, same identifier and a different hash is amended. That comparison is what puts an amendment state on the citation you are reading, rather than a date somebody typed into a spreadsheet.
Framework entitlement is a hard filter
An organisation entitled to FAA and EASA is never served UK provisions — excluded at query time by the same mechanism that separates one customer's data from another's, not filtered out after the fact. A rule from a framework you do not operate under is not slightly wrong; it is confidently wrong.
Your data never leaves the boundary it was promised to
The corpus, your documents, the index built over them and the speech-to-text all run on our own UK infrastructure. Only the question and the passages retrieved to answer it reach the model provider, inside the EU, for the length of the request — not used for training, and not retained afterwards.
Next
See it against your own manuals.
CitadelAI is licensed per seat, from three seats, and set up by our team with the frameworks your organisation operates under.