The Conversation Behavior Framework
The flow is not the work. The judgment behind it is. The Conversation Behavior Framework (CBF) is how ICX designs conversational and agentic AI: pick the right engine, design the behavior, manage the state, and tune the rhythm. Four lenses, one method.
A chatbot flow can look complete in a deck and still break in production. The copy reads well, the diagram is tidy, and yet the system answers when it should ask, remembers the wrong address, or holds the floor so long the conversation feels robotic. These are not writing problems. They are behavior problems.
The CBF makes the judgment visible. Instead of starting with the script, ICX starts with the decisions: what should the system do when it is uncertain, constrained, or unable to proceed safely? The four lenses below turn that judgment into something a team can design, review, and test.
One method, four lenses
Every ICX engagement runs the same four lenses against one question: what should the system actually do when it is uncertain, constrained, or unable to proceed safely?
The Determinism Map
The Determinism Map decides, for each intent, whether to use rules, an LLM, or a hybrid of both. It scores six real conditions instead of guessing. Not every job needs an LLM, and a beautiful flow the stack cannot deliver is just decoration.
What it scores
- Determinism needed (same input, same output)
- Compliance and audit requirements
- Latency budget before the turn feels broken
- Channel (web, WhatsApp, voice) and its limits
- Memory and state the step depends on
- Backend access the system can actually reach
Rule of thumb: rules run process, validation, routing, and policy. An LLM handles explanation, clarification, and messy language around them. Most real systems are hybrid.
The Behavior Decision Set
At every turn, an assistant chooses one move. The Behavior Decision Set names seven of them and gives each a trigger, so behavior is designed rather than accidental. A system can sound friendly and still overstep, or sound concise while hiding uncertainty. The seven moves keep it honest.
The seven moves
- Answer when confidence is high and the action is safe
- Ask when a required value is missing or ambiguous
- Verify when confidence is low and the action is risky
- Repair when a turn failed, without blaming the user
- Stop when the request is out of scope or unsafe
- Escalate when stakes or emotion run too high
- Hand-off to a human with enough context to avoid repeats
The State Ledger
Good copy cannot fix outdated state. If a user changes a delivery address and the system still confirms the old one, the sentence is fine but the state is wrong. The State Ledger gives every value a status and four operations that move it.
The four operations
- Persist a value the current task still depends on
- Update by replacing an earlier value (a correction replaces, it never appends)
- Expire a value when a change invalidates it (a new address can void a delivery slot)
- Clear a value the user cancelled or that is no longer safe to use
The rule: make state changes visible. "I will replace the old address with the new one" shows the user what changed and what still needs confirming.
Turn Rhythm
A bot sounds robotic because of pacing, not vocabulary. Adding slang to a clumsy turn does not help. Turn Rhythm tunes how the conversation moves so it feels speakable, not just readable.
The checklist
- Yield the floor early. The opening turn sets one expectation, then hands back.
- Lighten as context builds. Stop restating what is already clear.
- Make confirmation proportionate. Explicit for risky steps, light otherwise.
- Design for turn-taking, not delivery. Help the user move, do not make them read.
How does ICX apply the framework?
ICX runs the four lenses in order on every conversational or agentic project, then verifies with real evidence instead of the prototype's happy path.
Map the engine
Score each intent on the Determinism Map so the architecture is decided on evidence, not habit.
Design the behavior
Give every intent designed triggers for the seven moves of the Behavior Decision Set.
Draw the ledger
Map persist, update, expire, and clear for every multi-turn task, and make each change visible.
Tune and verify
Tune Turn Rhythm on the real channel, then test with transcripts and evals.
Framework FAQ
What is the Conversation Behavior Framework?
It is a behavior-first method ICX uses to design conversational and agentic AI, built from four lenses: the Determinism Map, the Behavior Decision Set, the State Ledger, and Turn Rhythm. The premise is that the flow or prompt is not the work. The judgment behind it is.
When should a chatbot use rules instead of an LLM?
Use rules when the journey is narrow, regulated, risky, or must be auditable. Use an LLM for open-ended questions and messy language. Most production systems are hybrid: rules for process, validation, routing, and policy, and an LLM for explanation and clarification.
What are the seven moves in the Behavior Decision Set?
Answer, ask, verify, repair, stop, escalate, and hand-off. Each has a trigger condition, so an assistant makes the right move when it is uncertain, constrained, or unable to proceed safely.
Why does an AI assistant confirm the wrong information?
Usually because its state is wrong, not its copy. A correction should replace an earlier value, not sit beside it. The State Ledger keeps the system relying on what is currently true.
The Conversation Behavior Framework is ICX's own synthesis, informed by the public work of Dr. Carmen Martinez on agentic UX and conversation design.