The argument in brief

Most conversations about AI agents swing between two poles. Either agents will run the company by next year, or they are a demo that breaks in production. Both views miss the useful middle.

Agents are very good at clearly specified work with clear inputs and a clear definition of done. They get unreliable when the goal is vague, the context is missing, or the cost of a wrong action is high. The skill that separates winners is not access to the best model. It is the judgement to point agents at the right problems and to design the boundaries around them.

Three ways deployments go wrong

  • The wrong problem. A perfect agent pointed at the wrong problem still fails. Teams automate what is easy to measure instead of what actually hurts.
  • The vague brief. "Improve the customer experience" gives an agent nothing to aim at. It will optimise something, just not the thing you meant.
  • No one accountable. When an agent closes a ticket and the customer leaves anyway, someone still has to own the outcome. If nobody does, the agent's metrics look great while the business loses.

What good looks like

Small teams that know exactly where to aim agents can do the work of much larger ones. The pattern is consistent: precise briefs, structured inputs, guardrails on irreversible actions, and a named human who owns the result.

Agent closed ticket. Human lost client. The six-word version of why accountability can't be delegated to software.

How this applies in travel and hospitality

Travel makes the boundaries very visible. An agent can draft a rebooking option in seconds. Issuing the ticket, applying the refund or promising an exception is where the company carries the risk. The same essay logic applies: let agents prepare and propose, and keep the commit step with rules or people until the data underneath is trustworthy.

I go deeper on this in AI in travel and hospitality: what actually works.

The keynote

This essay became the keynote I delivered at BA & Beyond 2026 in Budapest, built for business analysts, product leaders and operations teams. The talk argues that the skills analysts already have, precise requirements, edge-case thinking and risk management, are the real bottleneck in successful agent deployments.