Start with the decision, not the interface
A conversational screen can hide very different kinds of work. “Assign cases from this region to this queue” is an explicit rule. “Explain why this customer is dissatisfied across several messages” requires interpretation. Asking a model to perform both without distinguishing them makes testing harder.
Salesforce’s architecture guidance distinguishes traditional automation, agentic work and hybrid orchestration. Its examples are about Salesforce’s own products; they support the architectural distinction, not a claim that REDA has every Agentforce capability. [1]
- Request arrives
- AI proposes category
- Rules validate route
- Team handles exception
Divide the work into three responsibilities
| Responsibility | Suitable starting approach |
|---|---|
| Known conditions and exact outcomes | Flow or code with explicit validation and error handling. |
| Reading, summarising or proposing from variable text | AI output assessed against evidence and a defined purpose. |
| Consequential or ambiguous decisions | An accountable reviewer with a clear escalation route. |
This is a starting framework, not a rule that every Flow is safe or every AI step requires a separate approval screen. Scale, transaction behaviour, permissions and failure consequences still matter.
A worked example: routing a service request
Illustrative scenario. An incoming message contains a maintenance complaint and a billing question. AI proposes two categories and highlights the relevant phrases. A configured process checks allowed categories, record access and the required routing fields. A coordinator reviews mixed or unsupported requests instead of allowing the model to invent a new queue.
The routing action should receive a validated category and record identifier, not an unrestricted instruction to “sort it out”. Keep the existing rule for assigning an approved category to its queue. The model helps interpret the request; the controlled action handles the update.
Test the boundary between the two
- A request that clearly matches one category should take the expected route.
- A request containing conflicting instructions should become an exception.
- A failed record update should not be described as completed.
- Submitting the same request twice should not silently create duplicate work.
Also test volume and transaction limits for the chosen design. A demonstration of one successful request does not establish that a bulk process, scheduled job or portal channel behaves the same way.
Where REDA AI fits
REDA AI can be configured around business instructions, available context and permitted tools, with different experiences for staff, automation and external users. An integration with existing automation still needs design and testing; it is not created merely by writing a prompt.
Bring the existing Flow or process diagram to the first discussion. We can identify where it should stay deterministic and where an AI-assisted step is worth evaluating. Explore the platform.
Sources
- Salesforce Architects — Determining agentic vs. traditional workflow automation. Accessed 22 September 2026.