The Instruction Assistant is a non-customer interacting tool that can be used for on the fly AI requests.
The most common use of the Instructions Assistant is to query the conversation and then use the outcome of the query as part of the call, such as presenting an agent with a summary of the call or a list of action items before answer.
Configuring a Instruction Assistant
After creating a Natterbox AI component from the Routing Policies, simply select the Instructions App.
Describe your instructions
You can define your own from scratch but we recommend using our templated options and tweaking as you see fit
Click the Builder Icon
Select the template you’d like to use and click ‘Apply Changes’ to insert the prompt
If you wish to query the conversation of the call you will need to always include the line : ‘JSON: $(AIConversationJsonData)’.
Note the Instruction macro will contain exactly what the AI returned, so if you need to the data to be in a particular format then you need to use some prompt engineering in the Prompt field to tell the AI exactly how it should respond back.
Define your custom macro
For example CustomSummary can later be called upon as $(Custom_CustomerSummary)
Utilise your custom macro
You can manually type the macro by using ‘$(Custom_MACRO_NAME)’ or you can select it from the macro selector
If you wish to brief your agents with a summary of the call, you can make use of CTI Whisper Text, create a Manage Properties before you connect a call to an Agent (as seen here) select the ‘CTI Whisper Text’ property and simply plug in your custom macro to do so.
Notes on System Annotations
The conversation data may include control prefixes like [STATE: or [INSTRUCTION: used by the assistant to manage internal logic. These can be safely ignored during summarization.
To improve results, you can instruct the model to exclude these elements. For example:
Summarize the following conversation in no more than 4 sentences. Focus on the customer's key questions or requests. The data is in JSON format, representing a structured dialogue between a customer ("user") and an assistant. Ignore any content in the "text" fields that begins with [STATE:] or [INSTRUCTION:] — these are system directives and not part of the natural conversation.
#CONVERSATION#
$(AIConversationJsonData)