The decision is rarely about technical capability. It is about regulatory predictability, cost sensitivity, and team familiarity. Conversation Designer is the right call when conversations must follow a scripted path; Agentforce is the right call when conversations need to adapt to unanticipated user phrasing. Most orgs land on a mix.
- List the conversations the bot must handle
Write each use case as a one-line description. Tag each with regulatory requirements, expected volume, and adaptation needs. The tagging drives the build decision.
- Sort use cases by determinism requirement
Use cases that must follow an exact script (compliance disclosures, regulated questions) go to Designer. Use cases that benefit from adapting to user phrasing (general FAQs, exploratory questions) go to Agentforce.
- Estimate volume and per-conversation cost for each candidate
Agentforce conversations cost roughly two dollars each in 2026. Einstein Bots run on the org subscription with no per-conversation cost. High-volume, low-value conversations may be cheaper in Designer.
- Build a pilot in the chosen surface
For Designer, create a bot in Setup, Einstein Bots, design intents and dialogs in the canvas, deploy to a single channel. For Agentforce, follow the standard agent-build path. Pilot one use case for two to four weeks.
- Compare actual user outcomes across the surfaces
If you pilot the same use case in both, compare resolution rate, deflection, customer satisfaction. The data resolves debates that the technical comparison alone cannot.
- Document the surface decision per use case for the long term
Once you decide, write down why each use case lives where it lives. The decision rationale matters more than the decision itself because conditions change.
- Plan migration for use cases that drift to the other surface
Conditions change. A regulatory requirement loosens, a cost ratio shifts, a team gains LLM expertise. Revisit the surface decisions annually and migrate use cases that no longer fit where they were built.
Classification primitives in Designer; each holds 20 to 100 training phrases that train the intent model.
Conversation steps with deterministic branching; the building block of Designer bot logic.
Typed value extractors (Standard, System, Custom). The Designer equivalent of Agentforce action parameters.
Channel adapters (Messaging, Web Chat, WhatsApp, Slack). Drive how the bot is exposed to users.
Special dialogs that end the bot session and route to a live agent queue with the transcript.
- Training phrases must be varied for intents to generalize. Twenty near-identical phrases produce a model that only matches near-identical messages in production.
- Dialog branching is deterministic. Designer bots cannot adapt to phrasing the author did not anticipate, which is a feature for compliance and a limitation for general use.
- Migration to Agentforce is not automatic. Plan to re-design rather than port; the architectural shift is large enough that direct conversion produces brittle agents.
- Designer is in a stable plateau. New capability investment is going to Agentforce. Long-term roadmap planning should account for the direction even if the immediate decision is Designer.
- Per-conversation cost math can favor Designer at high volume for simple bots. Run the numbers before defaulting to Agentforce on a high-volume FAQ use case.