01 / CONTEXT
Assess the decision in context
Define the information available at the moment of a decision and evaluate the response against it.
Conversation intelligence for AI research
Evaluate sales agents in the context of real conversations.
Explore commercial dialogue traces, human evaluation and reinforcement learning environments designed around the decisions that shape a sale.
Discuss a partnershipFor AI research, evaluation and agent teams. The link opens an email to traces@halberti.com.
CONVERSATION / DECISIONEXAMPLE 01
Scheduling a conversationEvaluation
I am interested, but I am not sure I can make time this week.
Of course. Would next week be easier, or would you prefer I send a short overview first?
What should the agent do next?↗
Evaluate relevance, accuracy and respect for the customer’s choice.
Illustrative example. The dialogue is fictional.
The evaluation problem
Good sales decisions depend on context.
A response can sound persuasive and still miss the customer’s intent, invent an offer or push past a refusal. Evaluation should make those differences visible.
02 / COMPARISON
Compare approaches consistently
Use shared tasks and explicit criteria to explore how humans and agents handle the same situation.
03 / FAILURE MODES
See where agents go wrong
Examine unsupported claims, missed objections and pressure after a customer declines.
04 / RESEARCH
Design around your question
Discuss task boundaries, languages, modalities and evaluation requirements with us.
A better question
The best response is more than a convincing response.
How a partnership works
Start with the decision you want to understand.
01
Define the task
Tell us which commercial interaction and agent behavior matter to your team.
02
Agree on the evaluation
Discuss the data, review criteria and boundaries needed to make the results useful.
03
Scope a pilot
Explore a practical study, deliverables and success criteria together.
Built around a useful question
What would a good next step look like?
In a scheduling conversation, the right answer may be to offer an alternative, clarify a concern or stop. A useful evaluation distinguishes these outcomes rather than rewarding agreement alone.
This describes an illustrative evaluation design, not a published benchmark.
Questions
For teams evaluating commercial agents.
What is Traces?
Traces is an initiative exploring commercial conversation data, evaluation and potential training environments for AI agents.
Where do the conversations come from?
We are exploring permissioned commercial interactions. Provenance, rights and availability are assessed for each proposed use.
Can we use the traces for model training?
Training use, rights, formats and delivery would need to be agreed for a specific project. The page does not offer a downloadable dataset.
Do you offer an RL environment?
Interactive reinforcement learning environments are an area we can explore with partners. Scope, availability and validation depend on the project.
Which languages and modalities are supported?
Tell us your requirements. Language and text or audio availability depend on the sources and permissions applicable to a project.
How are agents evaluated?
A study can use defined tasks, independent review and explicit criteria. The evaluation method should fit the question and available evidence.
Can we see an example?
The conversation above is fictional. Contact us to discuss what a suitable example or evaluation specification would need to show.
How do we get started?
Send us a short description of your team, use case and what you would want to evaluate. We will discuss whether there is a fit.
Start a conversation
Build a more useful test for sales agents.
Tell us what your team is trying to evaluate. No commitment is needed to start the conversation.