Choose how to control the conversation (preview)

[This article is prerelease documentation and is subject to change.]

Conversation control determines who decides the next step: the author or the model.

Option A: Classic or deterministic orchestration

Use the classic or deterministic option when the conversation must follow a fixed, auditable path.

In classic orchestration, you design the flow topic by topic. This approach works well for cases where the agent must ask questions in a specific order, validate answers, and follow exact rules. Topics define how the conversation progresses, and classic mode is the nongenerative orchestration option.

You can still use AI for the generative answers as part of a flow within a topic or use generative answers as a fallback. In this scernaio, your agent attempts to answer the question using generative answers when it can't find a matching intent (defined in a topic) for the user's query.

Best for

• Authentication and identity verification

• Collecting information using Adaptive Cards

• Payment Card Industry (PCI) or payment collection

• Consent capture

• Regulatory or legal wording

• Dual Tone Multi-Frequency (DTMF) menu flows

• Any journey where exact steps are more important than flexibility

Trade‑offs

• Higher authoring effort

• Less flexibility in natural conversation

Simple rule

If the business says, "the agent must follow these exact steps," start with classic or deterministic logic using topics. You can switch to other modes later.

Option B: Generative orchestration

Use generative AI to determine the types of actions to take in the response (tools, knowledge sources, topic invocation) and how your agent responds to make the conversation more natural and fluid for the user. An agent that uses generative AI can also perform actions autonomously.

• Support questions asked in many different ways

• Support context-aware follow-up questions

• Support natural intent switching during the conversation without explicit transition rules

• Support one-step corrections to entities

• Support conversation context across turns

• Support multiple intents in a single turn

In generative orchestration, the agent can dynamically select the following options:

• Topics

• Tools

• Knowledge sources

• Events from other agents

This approach applies even though topics still exist. The difference is that the model decides when and how to use them rather than following a fixed path.

Generative orchestration reduces manual scripting but raises the bar on the following configurations and settings:

• Prompt quality

• Tool descriptions

• Guardrails

• Testing and evaluation

Best for

• FAQ and knowledge-based interactions

• Troubleshooting

• Open-ended discovery

• Multi-step self-service that needs tool calls

• Scenarios where callers speak naturally rather than following a menu

Trade‑offs

• Requires strong instructions and guardrails

• More testing investment

Simple rule

If the business says, "customers ask this question in many different ways," generative orchestration is usually the better fit.

Option C: Hybrid design

Most real enterprise agents shouldn't be fully scripted or fully generative.

Hybrid control combines

• Deterministic logic, where risk, compliance, or precision matters

• Generative logic, where flexibility improves experience

Example

• Greeting and intent discovery > generative

• Identity verification > deterministic

• Capture consent before proceed > deterministic

• Knowledge lookup > generative

• Use adaptive cards or rich messages in messaging → deterministic topic

• Authentication in messaging > always initiates authentication topic

• Payment capture or sensitive confirmation > deterministic

• Final summary or next step guidance > generative or deterministic depending on risk

Simple rule

Don't try to make the whole journey either fully scripted or fully generative.