What can synthetic customers do?

Synthetic customers can do a few things:

  • Send an initial message to the agent
  • Reply to agent messages
  • Wait for a response from the agent
  • Close a conversation

What are the limits of what I can do with fake data or external systems?

To get the most out of Syntrix today, it is helpful to keep the following current architectural boundaries in mind:

  • Historical data ingestion: Syntrix generates scenarios based on the specific personas and parameters you configure within the tool. We do not currently have the ability to ingest or "learn" from your organization's existing historical conversations. (Stay tuned for changes on this front!)
  • External system connectivity: Syntrix operates as a standalone environment. We don't yet have a way of integrating with external systems (training “sandbox” systems, your CRM, etc.).

As we continue to develop Syntrix, these boundaries will expand, allowing for deeper customization and more seamless integration into your existing technical stack.

How can I ensure my synthetic customers use specific data?

If a specific test case requires a particular piece of information—such as an account ID or order number—the current method is to include that data directly within the scenario's definition.

While embedding data in the scenario is effective for transaction-specific details, avoid using it to specify core identity information (like the customer's name). The latter is randomly generated by the LLM.

Can data flow from external systems into the simulated conversation?

The synthetic customer is only aware of two things: 1) the scenario and 2) the conversation in which they are participating.

While the agent may interact with an external system (like a CRM or a billing sandbox), there is no feedback mechanism that allows an external system to send information back to the synthetic customer in real time.

Example: Adding a mobile data top-up
  • The scenario: A synthetic customer is configured to ask the agent to add a 10GB data "bolt-on" to their account because they’ve run out of data.
  • The action: The agent opens an external telco provisioning sandbox (a mock version of the billing and network system) and successfully applies the data pack to the customer's account.
  • The gap: The provisioning system generates a transaction ID (e.g., TXN-9982X) and updates the "Current Data Balance" field to 10GB.
  • The result: There is currently no way for that external telco system to "push" that new balance or transaction ID back into the synthetic customer’s knowledge.

While the agent can type "I've added that 10GB for you, your transaction ID is TXN-9982X" to the customer, the synthetic customer won't "know" or "verify" that the system update actually occurred. The simulation remains independent of the external system's state.

In a live production environment, systems often talk to each other. But in a simulation, the synthetic customer's "brain" is isolated from the results of actions taken in external tools.

Are the prompts used by Syntrix customizable?

While prompts aren't directly editable, Syntrix is designed for deep customization. We expose the most critical parts of the prompt through scenarios and personas. By customizing these, you can shape the synthetic customer's behavior to meet your specific requirements, without needing to manage the complex, interlocking pieces that make the AI agents work.

Looking ahead, you'll also be able to tailor your evaluations with customizable scorecards, so stay tuned.