Most Hotel 'Agent' Problems Are Actually Workflow Problems, One Framework Argues

“Most hotel work is a workflow dressed up as an agent problem.” That’s the central claim of Darko Bosancic’s framework for hoteliers deciding how much AI sophistication a given task actually needs. The piece splits implementation into four tiers — Chat (simple Q&A, no cross-system decisions), Workflow (a predefined sequence that runs identically every time), Agent (a system that makes runtime decisions about which tools to use), and Orchestration (multiple agents and workflows coordinated by a governance layer) — and argues hotels are systematically over-buying tier three when tier two would win on cost, latency, and reliability. Pre-arrival messaging, review responses, and standard reporting are named as tasks where a workflow beats any agent, because their paths can be flowcharted in advance. The proposed decision matrix runs on two axes: whether a task’s path is predictable, and how many systems it touches. Under 20% complexity calls for a simple workflow; 20-50% complexity touching one or two systems calls for a single agent with guardrails; only tasks with over 50% complexity across multiple systems justify a full orchestration layer.

The framework lands as a useful corrective to the adoption-gap numbers Canary Technologies published last week: 98% of hotels have begun using AI in some form, but only 32% have embedded it across operations — a gap Canary’s own guide attributes partly to hotels reaching for agentic use cases (reservation management, guest messaging, front desk operations) without first sorting which of those actually need agent-level autonomy versus a simpler workflow. It’s also a sharper, more tactical version of the argument Are Morch made three weeks earlier: that most hotels do not need a 2030 AI strategy, they need confidence about 2026 — AI “amplifies” operational excellence rather than replacing it. Bosancic’s decision matrix is what that confidence looks like translated into a budgeting tool: match the AI tier to the actual task complexity, not to the vendor pitch.