Where AI must defer to humans
All business coordination is conversation. By classifying each utterance in a workflow as a speech act, we can draw exact, principled boundaries between what an AI agent can do autonomously and what requires human authority.
Grounded in Fernando Flores & Terry Winograd’s Language/Action Perspective (1986) and John Searle’s Speech Act Theory (1969).
Five speech acts, one hard boundary
Every utterance in a business process falls into one of five categories. The first two — requesting and stating — involve no commitments. An AI can execute them at scale. The moment communication shifts to promising or declaring, a human must authenticate the act. Click any row to explore.
| Speech Act | Definition | AI Autonomy | Human Role |
|---|---|---|---|
| Directives | Requesting action from others. “Please upload a police report.” “Fill out this form.” | Autonomous | None. AI can trigger follow-ups, gather missing data, and route requests without limit. |
| Assertives | Stating facts or beliefs. “The roof is 15 years old.” “The loss run shows 3 claims.” | Autonomous | Human reviews only when confidence scores are low or data is highly complex. |
| Expressives | Expressing psychological states. “We are sorry for your loss.” | Supervised | AI drafts routine acknowledgments; humans manage catastrophic or liability-adjacent communications. |
| Commissives | Committing to future action. “We will quote this at $5,000.” “We will repair your car.” | Strict HITL | The hard boundary. AI drafts the quote or settlement; the human authorizes the promise. |
| Declarations | Changing legal or institutional reality. “Your claim is denied.” “The policy is bound.” | Strict HITL | The hard boundary. Only a licensed professional or regulator-approved engine can change state. |
Mapping speech acts to phases
The Language/Action Perspective decomposes all coordinated work into four-phase loops between a Customer (who needs something) and a Performer (who delivers). Each phase is dominated by specific speech acts — and the AI/human boundary cuts cleanly across them.
Click a phase to see its speech acts and AI delegation pattern
The Architectural Takeaway
Stop viewing AI as a “decision engine” and start viewing it as a preparation engine. Grant the AI full read/write access to internal data and full permission to issue Directives to external users. But gate every state-changing API — Bind, Deny, Pay, Cancel — behind a human authentication token. The Speech Act category becomes a first-class attribute in your service architecture.
This isn’t conservative caution — it’s a precise match between what AI can do (process language) and what it cannot do (bear consequences). AI has linguistic competence but zero legal or financial agency. It has no skin in the game.
The insurance lifecycle as action loops
The insurance industry is a network of commitments — promises to pay, promises to perform, promises to indemnify. Every policy, quote, and claim is a speech act with legal force. Here’s how the full lifecycle decomposes into action loops, and where the AI boundary falls in each.
No commitments made in this loop. The AI is a hyper-efficient intake clerk — every speech act is a Directive or Assertive. It can run this loop at unlimited scale.
Renewals replay the Quote & Bind loop: the AI does all the preparation, but the human authorizes the new commitment. The pattern recurs because the type of speech act — not the workflow name — determines the boundary.
The System Architecture
Design the system so the AI Agent has full read/write access to internal data and full permission to issue Directives (requests) to external users. However, the core APIs that control state changes — Bind, Deny, Pay, Cancel — are physically gated behind a human authentication token.
The Speech Act category is the access-control attribute. Directives and Assertives get scope: agent. Commissives and Declarations require scope: human. The boundary isn’t a policy document — it’s enforced in the API layer.
“Management is not merely decision-making, but the ongoing generation and maintenance of a network of commitments.”
— Fernando Flores, Management and Communication in the Office of the Future, 1982
The candidate review process as action loops
A hiring workflow is a network of nested commitments. The hiring manager opens a loop with a request; the recruiter promises to fill it; interviewers promise to evaluate; and each promise spawns its own sub-loop. Step through the progression to see how each speech act builds on the last.
Loops Within Loops
The candidate review process isn’t a linear pipeline — it’s a network of action loops. Each time the recruiter needs something from another performer (an interviewer, a hiring committee), a new Customer–Performer loop opens inside the main one. The recruiter who is the Performer in the main loop becomes the Customer in the scheduling sub-loop. This nesting is how all complex coordination works: commitments spawn commitments.
The AI boundary follows the same rule here as in insurance: Directives and Assertives (gathering requirements, scheduling, collecting feedback) are AI-safe. But the final Declaration — “she’s our candidate” — must come from the hiring manager. The speech act type, not the industry, determines the boundary.