A practical primer for owners, directors, and flight departments –
what the terms mean, where the useful applications are, and why you should keep people in command.
Why this matters
If you’re in the aviation industry, you likely did not set out to become a software specialist.
In aviation, knowing what an aircraft can do is part of the job. Knowing its limits-and its level of performance-is just as important. Moreover, that’s a useful starting point for understanding artificial intelligence.
Most operators already work with sophisticated software that requires oversight and sound judgment. AI software introduces a way of working with computers we haven’t experienced before. We’re still learning what it can (and can’t and shouldn’t) do—and still finding the words to describe it.
Part of the challenge is the language. The technology is evolving, and so is the vocabulary we use to explain it. Technical terms spill into everyday conversation, often without clear definitions. For anyone trying to understand what AI could mean for their operation, that can make an unfamiliar subject feel unnecessarily difficult to approach.
The first distinction to make is this: AI is not a single technology.
AI encompasses a range of models, tools, and systems with different capabilities and limitations. Before exploring how those pieces fit together, let’s make sense of the language used to describe them.
The basic vocabulary in aviation terms
| Term | Plain-language meaning |
|---|---|
| Model | The trained system at the core of an AI product is the model. The model is “trained” on and then uses patterns learned during “training” to generate likely words, numbers, classifications, images, or actions from the input it receives. The model is not the app on the screen. It’s best to think of the model as the engine of an aircraft; the surrounding product provides the controls, systems, data, and operating limits. |
| LLM | LLM stands for Large language model. It is a type of AI model trained on extensive text and code. It can generate, summarize, translate, and analyze language by predicting likely sequences of words or tokens. LLMs are one kind of model. There are many others (computer vision model, speech models, forecasting models, etc…). For example, ChatGPT is primarily an LLM. Tesla’s full self-driving AI is primarily computer vision. |
| Foundation model | A large, general-purpose model trained on a broad mix of information. Other products and tasks can be built on top of it. A useful analogy is a type-certified airframe: the same aircraft type can be equipped for very different missions, depending on the configuration and operating environment. |
| Frontier model | A model at the current leading edge of capability. |
| Prompt | The question, instruction, and constraints given to the AI Model. It is the equivalent of a clearance, trip sheet, or briefing. Clear inputs usually produce more useful outputs; vague inputs leave more room for error. |
| Hallucination | A confident-sounding answer that is wrong, invented, outdated, or unsupported. AI models provide probabilistic answers. Constraints in questions can help to avoid hallucinations. It is always best to check answers against rules and regulations. |
| Training and inference | Training is how the model is built. Inference is using that trained model to answer a live request. In aviation terms, training is the simulator and recurrent preparation; inference is today’s flight. |
| Tokens | Tokens are the small pieces of language an AI system reads and writes. |
| RAG | Retrieval-augmented generation. The system looks up relevant, current documents before it drafts an answer. That is like pulling the current chart, AFM section, MEL item, or trip folder instead of relying on memory. |
| Fine-tuning | Additional training that makes a general model better at a particular organization’s work. It is like a tailored completion or mission kit. |
| Guardrails | Guardrails are the rules and controls that limit what the AI system may do. While they often take the shape of guidance, guardrails include permissions, approval steps, blocked actions, and scope that users or organizations set on the AI. |
| Harness | The surrounding systems that make an AI model useful and controllable. It can include tools, data connections, memory, permissions, a separate workspace, logging, and stop conditions. If the model is the engine; the harness is the aircraft’s airframe, avionics, controls, limitations, and operating procedures. |
| MCP | he Model Context Protocol is a standard way for an AI system to connect with other software and information sources. For example, it can let an AI assistant search a company’s documents, check a calendar, read approved email, look up information in a database, or create a draft in another application. Think of it as a common adapter: different AI products can connect to the same kinds of tools without building a completely new connection each time. The connection still needs to be approved and limited by user or organizations’ permissions. |
The Levels of AI
When someone talk about AI it can mean a variety of things. AI can receive and respond to questions, like in the common chat style. AI can also create imagery or video based on almost any request. Higher-level and higher functioning AI can perform real work.
Each step up this ladder gives the technology more ability to act, which means the department needs stronger permissions, monitoring, and human oversight.
In general, the levels below, Chat, Copilot, Agent and Operation are the levels by which you can use and interact with AI. Let’s take a quick look at each.
1. Chat
You ask a question and the system answers. Nothing in the real-world changes unless a person takes the answer and uses it. Chat can help draft a passenger letter, explain a maintenance term in plain English, or create a first draft of a procedure. The chat takes no further action. It is up to the user to copy and paste that answer into a document for example.
2. A copilot with tools
The system can look something up, read a file you provide, search an approved source, or fill in a draft. A person still reviews the result and takes the consequential action. This is like an autopilot: it can follow the path you load, but you brief it, monitor it, and disconnect whenever you choose.
3. Agent AI
Agentic AI is a technical term for a practical idea: giving AI a goal and enough independence to work through the steps. It plans, uses tools, checks the results, and adjusts its approach without asking for approval at every click.
An agent can work alone, or it can share the job with other agents. Like an aircraft arriving at a busy airport: approach control, tower, and ground control each handle a different part of the arrival. Each has defined responsibilities, responds to changing conditions, and coordinates with the others. That division of responsibility is a useful way to picture several AI agents working together.
For an operator, an agent might assemble a draft trip folder using passenger and aircraft information, identify missing details, and flag them for review. Another might monitor an inbox and prepare suggested responses.
The important question is how much authority it has: what can it do on its own, and where must it stop for a person’s approval?
4. A controlled operation
At this level, AI connects to real company systems and carries out parts of a workflow. That requires clear permissions, records of what the system did, a safe place to test changes, and human approval before important actions occur. The department should also be able to review the system’s work afterward. If AI connects to scheduling, maintenance tracking, owner communications, or vendor spending, the organization should deem these controls essential.
Where AI may help a flight department
The safest early uses are usually the ones that reduce administrative work without making operational decisions. Examples include drafting routine correspondence, summarizing internal meetings, organizing information from a trip folder, comparing documents for differences, creating a first pass at training material, and turning notes into a more consistent checklist or procedure draft.
The value is not that the system replaces the people. The value is that it can handle a time-consuming first pass, leaving experienced people to review, correct, and decide. A good analogy is a capable first officer who prepares the work so the captain can focus on judgment and accountability.
As the stakes rise, the controls must rise with them. An AI system may help prepare a schedule, but a human should approve any publication of the schedule. It may identify a maintenance record that deserves attention, but that does not mean it should close the discrepancy. A system may draft a message or response, but a person should decide what is sent.
Four questions to ask before you say yes
What level is this? Is it chat, a copilot with tools, or an agent that can keep taking steps? The answer explains the required amount of supervision.
What information does it use? A model’s general training is not the same as access to your current manuals, records, policies, or schedules.
What is it allowed to touch? List the systems and people in scope: calendar, email, maintenance platform, accounting files, vendor portals, and owner communications. Permissions should be specific, limited, and review-able.
Who is responsible when it is wrong? The answer should be a named role or person, not “the system” or “the vendor.” Responsibility that cannot be identified will eventually land on the operator.
A sensible starting approach
A flight department does not need an enterprise-wide AI strategy on day one. Start with one contained problem that is repetitive, easy to check, and low consequence if the first draft is imperfect. Define what information the system may see, what it may produce, and what it may not do. Keep a person responsible for the final action.
Run the trial long enough to learn where the system helps and where it creates work. Save examples of good and bad outputs. Ask whether the result is accurate, traceable, and worth the review time. If the answer is yes, expand carefully. If the answer is no, stop or redesign the workflow. A small, well-controlled test will tell you more than a polished demonstration.
The goal is not to make every flight department look like a technology company. The goal is to use new tools without giving up the habits that make aviation safe: defined authority, current information, clear procedures, disciplined verification, and a person who remains accountable.
What does not change
AI can produce a useful first draft of many things in business aviation: letters, trip summaries, training notes, procedure drafts, and internal briefings. Used this way, it can save time in the same way a good first officer saves time: someone else prepares the work, and an experienced person decides whether it is fit to use.
AI could also potentially invent a tail number, airport, regulation, contact, or slot time and present it with complete confidence. While frustrating, that is not a reason to reject every tool. It is a reason to verify important information against the source and to build that verification into the workflow.
You already know how to live with a powerful, limited system. You do not let the autopilot own the airplane, and you should not let an AI model own the department. Use the chat where it helps. Put a proper harness around anything that can act. Keep a qualified person with clear authority in command.
Join the ACSF AI Working Group
If you’d like to learn more about the AI ACSF Working Group or apply to become a part of the group, please visit: The ACSF AI Working Group
Share your thoughts: AI in Aviation Survey
