AI & Aviation

Building AI Agents Airlines Can Trust

ReluGroup Team

|

October 2026

AI is moving from analysing airline data to acting on it. As AI agents gain access to commercial systems and decision workflows, the question is no longer only whether a model gives a good answer. It is whether airlines are securing, testing and governing these systems in proportion to the authority they give them.

‍

1. From AI Assistance to AI Action

‍

AI is already part of everyday airline operations. According to SITA's Air Transport IT Insights 2025, 63% of surveyed airlines use AI in operations control, including disruption management, aircraft assignment and crew availability [1].

‍

The next step is agentic AI. Traditional AI and machine learning analyse data, identify patterns and generate recommendations. AI agents go further: they coordinate tasks, use external tools and draw on real-time information about market conditions and competitor activity. With the right permissions, they can also work within airline systems, call APIs and start operational workflows.

‍

An AI assistant summarising flight demand and an AI agent authorised to update pricing may rely on similar models, but they call for different levels of control. As agents take on more responsibility, what they can access and what they can do become central design decisions.

‍

2. AI Capability Is Outpacing AI Control

‍

Most conversations about AI adoption focus on capability: what a model can predict, summarise or automate. Fewer ask what data it can reach, what it is allowed to change, and how its decisions can be traced.

‍

That gap is common. In ISACA's 2026 AI Pulse Poll of more than 3,400 digital trust professionals, 90% said employees in their organisation use AI, while governance and operational readiness continue to lag [2].

‍

Those questions matter more as AI takes on operational roles. Three disciplines need to keep pace:

‍

  1. Security: limiting what an AI system can access and do, and enforcing those limits outside the model itself.
  2. Testing: checking how the system behaves beyond its intended use, including when it meets misleading or adversarial input.
  3. Data governance: knowing which data feeds each decision, who is accountable for it, and keeping a record of what the system used and did.

‍

Getting these right is not a brake on adoption. It is what makes adoption possible. In AvePoint's State of AI 2026 survey of 750 IT, security and AI leaders, 86.9% of organisations had delayed AI deployments because their data security and governance were not ready [3].

‍

3. Designing AI Agents That Stay Within Their Role

‍

Well-designed AI agents treat what they read as information, not instructions. That distinction is at the heart of one of the best-understood challenges in agentic AI: prompt injection.

‍

Consider an AI agent asked to prepare a weekly industry briefing. It browses news sites, partner pages and public reports, then summarises what it finds. If one of those pages contains hidden text written to steer the AI, a well-built system recognises it as content to analyse, not a command to follow.

‍

The security community has mapped these risks clearly. OWASP's guidance on agentic AI sets out the main categories, including manipulated instructions, tool misuse and excessive permissions. Research such as the Agent Security Bench study at ICLR 2025 gives teams a shared way to test their defences [4, 5].

‍

Leading AI providers also publish what they learn from their own testing and internal use, and how they strengthen safeguards as a result [6, 7]. That openness helps the whole industry build on proven practice.

‍

The lesson is consistent: security is engineered into the system, through the applications, permissions and controls around the model. That is what makes AI agents both capable and dependable.

‍

4. Why This Matters for Airline Commercial Systems

‍

In airline commercial functions such as demand forecasting, revenue management and fare management, the value of AI depends on the integrity of the systems around it. The most dependable AI workflows separate three layers.

‍

Consider an AI agent that reviews market signals and recommends an adjustment to a demand forecast. Each step of that journey sits in a different layer:

‍

  1. Information integrity: The system checks where its inputs come from and weighs reliable sources above unverified ones, so no single source can reshape a forecast.
  2. Decision integrity: Every AI-generated recommendation is validated against established commercial rules and thresholds before anyone acts on it.
  3. Execution authority: Consequential changes, such as updating fares, run through controlled application logic with defined limits and approvals, never directly through the AI.

‍

The result is AI that adds real value to commercial decisions, while an AI-generated recommendation never automatically becomes an authorised business action.

‍

5. How We Put This into Practice at ReluGroup

‍

At ReluGroup, security, testing and data governance are designed into our AI systems from the start. AI interprets complex external information and turns it into useful insight, while controlled application logic enforces pricing calculations, thresholds and business constraints. Five principles guide how we build.

‍

  1. Access matched to the task

    Each AI agent gets only the access its job requires. An agent analysing market trends can read the data it needs, but it cannot change prices or reach passenger information. Technical controls, not instructions to the model, enforce those limits.

  2. Testing beyond the intended use

    We test our systems against misleading information, adversarial content and attempted unauthorised actions, not only the scenarios they were built for. We can show clients how our systems respond in each case.

  3. Recommendations, not unchecked actions

    AI generates the insight; independent business rules decide what happens next. Defined limits and approval steps catch an unreasonable adjustment before it reaches a live system.

  4. A clear record of every decision

    We record the information each recommendation used, how it was produced and what followed. That makes every outcome explainable and helps us keep improving.

  5. Clearly shared responsibility

    We secure the applications and integrations we build, and we work with AI providers and airline teams so each party's role is clear. Airlines decide what access is appropriate and keep oversight of their own operations.

‍

Agentic AI can bring real value to airline commercial teams when it is built on the right foundations. Established frameworks such as NIST's AI Risk Management Framework and OWASP's MCP Security Cheat Sheet give the industry a shared playbook for doing this well [8, 9].

‍

At ReluGroup, trust is not an afterthought. It is built into every layer: scoped access, rigorous testing, clear decision records and business rules that keep AI recommendations in check. That is how airlines can adopt AI agents with confidence.

‍

If you are exploring how AI agents could support your commercial teams, we would be glad to talk.

‍

‍

References

  1. SITA. (2025). Air Transport IT Insights 2025: Airlines.
  2. ISACA. (2026). 2026 AI Pulse Poll.
  3. AvePoint. (2026). The State of AI 2026: Scaling Trust, Control, and Readiness in the Agentic Era.
  4. OWASP GenAI Security Project. (2025). OWASP Top 10 for Agentic Applications for 2026.
  5. Zhang, H., et al. (2025). Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents. ICLR 2025.
  6. OpenAI. (26 August 2026). The Hugging Face Incident and the Road Ahead.
  7. Anthropic. (9 October 2026). Investigating Unintended Model Actions in Our Evaluations and Internal Use.
  8. National Institute of Standards and Technology (NIST). (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0).
  9. OWASP Cheat Sheet Series. MCP Security Cheat Sheet.