Picture two aircraft on the same apron at dawn. The first is having its wing scanned by an autonomous drone that maps every rivet in ninety minutes — a survey that once cost a technician most of a shift on a cherry-picker. The second is being released to service by a licensed engineer who has just absorbed the gist of forty thousand pages of maintenance manuals in the time it took the kettle to boil, because an AI assistant read them first. Neither aircraft is being flown, fixed, or signed off by a machine. Both are being handled by professionals whose hands are now free for the work that genuinely needs judgement.
That is the honest state of artificial intelligence in aviation in 2026 — not the pilotless airliner of the headlines, but something quieter, more useful, and far more immediately relevant to anyone running a maintenance organisation, a CAMO, or a safety department.
The two-speed reality
The single most useful idea a decision-maker can carry into a budget meeting is this: “AI in aviation” is two different stories moving at two different speeds.
The fast track is non-safety-critical AI — predictive maintenance analytics, fuel-burn optimisation, drone-captured inspection imagery, generative-AI document search, safety-data mining. These tools do not fly the aircraft or release it on their own; a human always makes the final call. They are deployed today, at scale, and paying for themselves.
The slow track is safety-critical, certified AI — machine learning embedded in avionics that performs a safety function. Here, despite years of breathless coverage, the truthful status is that nothing has been certified yet, anywhere. The current front-runner, Daedalean’s visual traffic-detection system, is described by its own chief executive, Bas Gouverneur, as likely to become “the world’s first safety-critical product for civil aviation based on machine learning”; according to reporting in late 2025 it had cleared two of four FAA approval stages, with certification still pending and pursued in parallel with EASA.[1] Until that line is crossed, treat certified airborne AI as a multi-year research and certification track, not a near-term purchase.
Confuse those two speeds and you will either waste money chasing a robot co-pilot that regulators will not approve this decade, or miss the unglamorous tools already cutting your inspection times and fuel bills.
Where the regulators have drawn the line
Europe has thought about this longer than anyone. EASA’s AI Roadmap 2.0 (2023) built a human-centric “trustworthiness” framework, and its concept papers turned that philosophy into workable guidance: Issue 02 (March 2024) covered Level 1 and Level 2 machine learning, and the proposed Issue 03 (2026) extends it to Level 3 — advanced automation — while introducing concepts such as human-AI teaming and operational explainability.
The genius of the EASA scheme is that it classifies AI not by how clever it is, but by how much authority it holds and how the human stays involved:
- Level 1 — assistance to the human (1A augments, 1B supports decisions).
- Level 2 — human and machine work as a team (2A cooperation, 2B collaboration).
- Level 3 — advanced automation (3A safeguarded, with a human able to override; 3B non-supervised).
EASA’s own timeline is candid about pace: first approvals of Level 1 AI around 2025, Level 2 and the first safeguarded Level 3 systems around 2035, and genuinely autonomous AI not before 2050. The agency calls the 2035 date “not a formal target but rather a prediction.” In other words, the regulator is telling you, in writing, that the autonomous aircraft is a generation away.
The FAA reached the same junction by a different road. Its Roadmap for Artificial Intelligence Safety Assurance (August 2024) leans on “overarching properties” — intent, correctness and innocuity — and on a sharp distinction between “learned” AI (trained, then frozen and testable) and “learning” AI (which keeps adapting and is therefore far harder to trust). The FAA warns explicitly against personifying AI and prioritises the frozen, testable kind. It is no coincidence that the leading certification candidate is machine-learned, with fixed, deterministic behaviour.
ICAO plays the third role: it does not certify, it harmonises. Its 2025 Assembly carried multiple AI working papers and proposals to weave AI through ICAO’s programmes, and a global seminar on AI oversight was floated for 2026. The takeaway for a manager: EASA, the FAA and ICAO are converging, not competing. They agree on the essentials — a rigorous safety case, an accountable human, an incremental rollout, and shared industry standards (notably the EUROCAE/SAE AS6983/ED-324 process standard now being finalised by a 600-strong committee that includes Airbus, Boeing and the regulators themselves).
The EU AI Act: already in your regulatory bloodstream
If your operation touches Europe, the EU AI Act is not a separate burden bolted on top of aviation rules. Through Article 108, the Act’s high-risk requirements flow directly into EASA’s existing safety regime: human oversight, technical robustness, data governance and clear accountability become aviation requirements rather than a parallel paperwork exercise. EASA has already opened this front through rulemaking task RMT.0742 and its first formal AI proposal, NPA 2025-07. The practical message is reassuring and demanding at once: AI in safety-related products will be governed by familiar aviation processes — but to AI-Act substance.
What is actually working right now
Strip away the speculation and a concrete, fundable picture emerges.
Predictive maintenance — the mature workhorse
This is the most proven large-scale use case. Lufthansa Technik’s AVIATAR platform supports, by the company’s July 2025 account, more than 5,000 aircraft across 40-plus customers; Airbus says its Skywise platform connects well over a hundred airlines; Collins’ Ascentia, GE’s analytics and Air France-KLM’s Prognos all play in the same arena. The promise is simple to explain to a board: spot the failing component before it fails, convert unscheduled groundings into planned events, and stop replacing parts that still have life in them. A healthy note of caution belongs here too. As far back as 2023, Ahmed Safa of Emirates Engineering told Aviation Week that the airline had “not really seen any tool in the market that delivers demonstrable and measurable outcomes for predictive maintenance,” adding that it remained “still early days.”[2] The technology is real and scaling, but the returns are uneven and operator-specific. Demand evidence, not slideware.
Drone and automated inspection
Donecle says it operates a fleet of inspection drones across fifteen countries, with approvals from manufacturers and regulators including Airbus, Boeing, EASA and the FAA, and reports inspecting aircraft up to ten times faster than by hand. In October 2024, Delta TechOps announced it had become the first US carrier to gain FAA concurrence for fleet-wide drone inspections, reporting that a narrowbody general visual inspection fell from roughly sixteen hours to under ninety minutes.[3] Tellingly, the vendors frame this as teamwork: the AI is better at spotting hairline cracks, the technician is better at the unexpected. The human still validates and signs.
Operations, fuel and air traffic
SITA’s OptiFlight builds a machine-learning “digital twin” of each individual tail to recommend the most efficient climb, with SITA reporting climb-out fuel savings averaging around 5%. In air traffic management, AI is firmly in the decision-support and shadow-trial phase: NATS’ Project Bluebird tests AI on a digital twin of UK airspace, and SESAR-funded work aims to predict congestion an hour ahead instead of twenty minutes. Controllers and pilots remain in command; the AI advises.
Generative AI in the back office
The fastest-moving, lowest-risk frontier is paperwork. GE Aerospace says it built a generative-AI assistant that turns hours of records searching into minutes; start-ups are layering voice-driven assistants over MRO systems to surface the right manual page and draft documentation. EASA intends to treat such tools as “operational tools” approved within an organisation’s existing Part-145, Part-M, DOA or POA approval — not as certified airborne AI. For most maintenance and CAMO leaders, this is where the near-term productivity actually lives.
How maintenance and CAMO work will change
EASA’s own worked examples are a window into the near future, and they land squarely in your world. The first is usage-driven corrosion control: instead of inspecting every airframe on a fixed calendar, a model trained on each aircraft’s route history, weather exposure, de-icing record and fleet-wide findings predicts where corrosion is likely — so the CAMO schedules a focused inspection at the moment it is most economic and least risky. The second is AI damage detection in NDT images, where computer vision flags cracks in X-ray, ultrasonic or thermographic scans for an inspector to confirm. In both, EASA is explicit: the inspector validates, the human decides.
So the CAMO role shifts from administering schedules to orchestrating data — from “when does the calendar say?” to “what is this individual aircraft actually telling us?” The Part-145 engineer is freed from documentation archaeology and routine imagery to concentrate on diagnosis, repair and the judgement calls that release an aircraft to service. The work does not vanish. It moves up the value chain.
Should professionals be afraid for their jobs?
The short, evidence-based answer is no — and three forces explain why.
First, the industry is short of people, not flush with them. Boeing’s 2025 Pilot and Technician Outlook projects a need for 660,000 new pilots and 710,000 new maintenance technicians over the next two decades. Oliver Wyman’s 2025 survey found two-thirds of MROs struggling to recruit technicians, with more than 40% of the licensed workforce nearing retirement. AI is arriving as a way to close that gap, not to deepen a surplus that does not exist.
Second, regulation keeps a human accountable by design. Every framework above is built on it. The EASA maintenance use cases require inspector validation. The FAA forbids treating AI as a person. EASA’s Issue 03 adds a formal “responsibility-scheme” assessment for higher-authority systems. Releasing an aircraft, signing an inspection, commanding a flight — these remain human acts, in law.
Third, the economics are about augmentation. Across every case study, AI removes the drudgery — the manual photographing, the manual searching, the manual labelling — so skilled people spend more time where their expertise pays. The honest caveat is that “augmentation” is not “no change.” The professionals who thrive will be those who add new competencies: reading and interrogating AI outputs, knowing when to override, basic data literacy, and the discipline to treat an AI recommendation as a hypothesis to be checked rather than an answer to be trusted. EASA’s Issue 03 already writes “competence considerations” for organisations into the framework. The risk is not that AI replaces the engineer; it is that the engineer who refuses to learn the new tools is out-competed by the one who does.
What decision-makers should do now
Sort your AI portfolio by speed. Move now on the fast track — predictive maintenance, fuel optimisation, drone inspection, generative-AI document search. Treat certified airborne AI as long-horizon R&D.
Build the compliance scaffolding early. Map every AI system to an EASA level, revise your concept of operations, and prepare your risk and (where relevant) ethics assessments. If you touch the EU, track NPA 2025-07.
Invest in people, not headcount cuts. Fund data literacy and AI-oversight skills across maintenance, CAMO, operations and safety. Frame AI to your teams honestly: it is a tool that makes them more valuable.
Demand measurable outcomes from vendors. Adopt the Emirates posture — require documented, fleet-specific results, and run a validation trial before any fleet-wide rollout.
- Keep the human accountable — by design and by law. Build human validation into every workflow. It is both a regulatory necessity and your reputational insurance.
The aircraft of 2050 may well fly themselves. The maintenance organisation of 2027 will not run itself — but it will increasingly be run by professionals who let AI do the reading, the watching and the predicting, so they can do the deciding. The question for leaders is not whether to let the machines in. It is whether your people will be ready to supervise them.
References:
Bas Gouverneur, quoted in AOPA, “Future Flight: Daedalean PilotEye,” October 2025. Certification status (two of four FAA Stages of Involvement, third commencing summer 2025; concurrent FAA/EASA review) as reported at that date and subject to change.
Ahmed Safa, Emirates Engineering, quoted in “Are Predictive Maintenance Tools Falling Short?,” Aviation Week, 8 February 2023. Statement reflects the airline’s position as reported at that date.
- Per Donecle and Delta Air Lines / Delta TechOps public statements (Delta announcement, October 2024). Figures are company-reported and not independently verified.
Editorial note. Company and product names are referenced for factual, informational purposes only; no endorsement, partnership or affiliation is implied. Performance figures and status claims reflect statements publicly reported by the companies concerned or by the cited publications, are accurate to the best of our knowledge as of the date of publication, and have not been independently verified. This article is general information, not legal, technical or investment advice.
