AI agents in aviation: drafting yes, signing no

An AI agent can assemble a file, pre-fill a form and flag a missing record. It cannot sign the aircraft's airworthiness documents: a signature is the act by which an authorized person attests and commits their responsibility, and that gesture cannot be delegated to software.

The boundary is worth stating clearly, because the tools are changing nature. The assistants of the last two years searched and summarized. Recent agents act: they fill, send, trigger. And in the same month, an aircraft crossed the United States without the onboard pilot touching the controls.

What September showed: the agent that acts

On September 18, 2026, Joby Aviation announced that a converted Cessna Caravan, its J208, had covered 3,199 miles across the United States with zero control inputs from the onboard safety pilot: taxi, takeoffs, navigation and landings were flown by the system, with a remote pilot acting as pilot in command supervising it from California and a base in South Carolina, up to 2,323 miles away, handling flight plan updates and radio. The system, acquired with Xwing, claims more than 400 flights and 800 automated hours.

On the office side, document agents are moving too: an agent receiving an incomplete file can now identify the missing record, draft the follow-up email and prepare the tracking form without human intervention.

These two facts are not of the same kind. Joby's flight is about certifying an airborne system, an authority's work on a specific product. The office agent is about an operator's decision: which tasks to hand over, and which to keep.

What the European framework is preparing, without having settled it

EASA publishes its doctrine in stages. The Concept Paper Proposed Issue 03, opened for consultation on June 3, 2026, describes levels of autonomy, from level 0 of low automation up to advanced automation where the human may be remote or absent from the operation, through human assistance and human-AI cooperation. The consultation closed on August 12, 2026; the text remains a proposal preparing rulemaking work, not an authorization.

In parallel, EUROCAE working group WG-114 is writing ED-324, a development and approval standard for aeronautical products embedding AI: the WG-114 group page lists a target publication date of 31 December 2026; the standard remains a draft and is binding nowhere today.

In other words: Europe is building the framework that will one day allow highly autonomous systems to be assessed. Today, that framework signs nothing in anyone's place, and no text transfers airworthiness responsibility to software.

A signature is an act, not a checkbox

In an aircraft's file, each signature attaches a fact to a person. The certificate of release to service is signed by authorized staff attesting that the work was done to the applicable data (point 145.A.50(a)). The technician who records an action in the technical log commits their reading of the situation. The CAMO issuing a recommendation for the airworthiness review certificate answers for the file's consistency (point M.A.901(d)).

These acts assume two things an agent does not have: a verifiable authorization, and an enforceable responsibility if the attestation is false or careless. An agent can therefore produce everything that comes before the signature: the assembled file, the pre-filled form, the flagged gap. The final gesture belongs to a person.

A fictional case: the file prepared by the agent

Take an invented example. Karim is the airworthiness manager at an operator of four turboprops. His team is testing an agent that prepares monthly packages: it gathers technical log entries, lists deferred defects and their deadlines, and pre-fills the summary intended for the CAMO.

In the first month, the agent flags an anomaly a manual check might have missed: a deferred defect whose repair category is about to expire. Karim checks the original record, confirms, and signs the summary after correction. The gain is real: the file arrives complete, the error is caught early.

The limit is just as real. Had the agent misread the deadline date, Karim would have signed a wrong summary. The tool prepares; the person checks, corrects and commits. The day the signer can no longer retrieve the original record or redo the reasoning, the agent no longer assists: it replaces, and that is a different problem.

Write the boundary down before the first file

A few decisions to write before entrusting preparation to an agent:

• What it produces: the documents it assembles or pre-fills, and those it never touches.
• What bounds it: the data it may access, what is off limits, and what triggers a stop.
• What the person keeps: review of the original records, the right to refuse the draft, signing only after checking.
• What stays traceable: what the agent produced, changed or proposed, kept as a piece of the file rather than an attestation.
• What resumes without it: the manual method that works if the tool is unavailable.

These bounds do not come from a regulation on agents, which does not exist yet. They extend an older requirement: the person who signs must be able to show what their attestation rests on.

What neither the agent nor the autonomous aircraft changes

Joby's flight shows a machine can hold a complex mission alone under supervision. Agents show software can hold a file alone up to its final form. Neither transfers responsibility: someone must be able to answer for what was produced, and that person signs.

To prepare the evaluation of such a tool with a vendor, our article AI in aviation: what evidence before using it? offers a trial sheet. Here, the question is simpler and older: in your file, who attests, and to what?