Goals and evaluation
Define what success means for each agent, rate how it talks with competencies, and let NuTalk evaluate its calls and conversations so you know what to improve.
Overview
Two things measure an agent:
- Goals
- Yes-or-no outcomes, such as an appointment booked. Each evaluated call or conversation met the goal or not, and each goal's success rate is tracked over time.
- Competencies
- Behaviors rated on 3 or 4 performance levels, such as how the agent handles price questions. They judge the conversation itself, not its outcome.
With evaluation on, an AI evaluator scores the agent's completed calls and inactive conversations against both.
Defining agent goals
Goals are the outcomes you want your agent to reach. Write each one so it can be judged as met or not from the conversation.
Examples of goals
- Conversion
- Appointment booked, demo scheduled, sale completed, lead qualified.
- Information
- Contact details collected, survey completed, feedback gathered, issue documented.
- Resolution
- Problem resolved, question answered, complaint handled, request fulfilled.
- Engagement
- Follow-up agreed, interest confirmed, referral obtained, opt-in collected.
Set up goals
Open Goals
Open AI agents → Goals and select the agent.
Add a goal
Click Add Goal, or Generate with AI to draft goals from the agent.
Name it
Enter a clear, specific Goal Title, such as “Appointment booked”.
Describe success
Write the Success Criteria: the evaluator uses them to decide yes or no.
Pick the channel
Choose the Channel the goal is evaluated on.
Add targets (optional)
Set a count or a rate per day, week, month, quarter, semester or year, and mark one as primary. Without targets, the goal shows its success rate over the last 30 days.
Goal title: Appointment booked Channel: Main line Success criteria: The caller agrees to a specific date and time, and the agent confirms the appointment or books it with its booking tool. Target: Rate 30% per week (primary)
Competency evaluation
Competencies measure how well your agent handles the conversation. Unlike goals, which are outcomes, they rate the conversation itself.
Competency criteria
Each competency has a name, one sentence of what good looks like, and 3 or 4 performance levels from lowest to highest (by default: Needs improvement, Partially meets, Meets expectations, Exceeds expectations). A rating at the passing level (the middle one, or the third of four) or above passes. Every competency counts the same. For example:
| Competency | What good looks like |
|---|---|
| Greeting quality | Introduces itself professionally and uses the caller's name correctly. |
| Information accuracy | Gives correct facts and makes nothing up. |
| Problem resolution | Addresses the caller's main concern. |
| Tone and empathy | Friendly and patient, with a tone that fits the caller. |
| Call closure | Ends with clear next steps and a proper sign-off. |
Write competencies for your use case, such as a compliance check for your industry, your brand's voice or knowledge of a specific product. Measure one behavior you can hear in a conversation, and group competencies by type (skill, knowledge, attitude) and category.
Turn on automatic evaluation
Switch it on
On AI agents → Goals or Competencies, turn on Auto-evaluate calls & conversations for the agent. It applies to all of the agent's channels.
Choose how many
In a channel's Evaluations settings, How many to evaluate sets whether every call or conversation is evaluated, or only a share of them.
Set when chats count as finished
For chat channels, the Evaluation inactivity window sets how long a conversation goes without messages before it is evaluated.
Write the competencies
On Competencies, add what good looks like and 3 or 4 performance levels for each.
Understanding scores
Each evaluated call or conversation gets:
- for each goal, Met or Not met, with the evaluator's justification;
- for each competency, the level it reached;
- an overall score from 0 to 100.
Read them in a call's Evaluation panel, or in the Goal evaluation log and the Competency evaluation log.
Track results
AI agents → Goals lists each goal's success rate over the last 14 days and whether it is strong, on track or needs attention. Show analytics charts the success rate over time, by day, week, month or quarter, filtered by channel and agent version, and compares the latest period with the one before.
AI agents → Competencies shows, for a time window, the share of ratings that reached the passing level, the trend against the previous window, the ratings by level and recent misses.
Try it yourself
Before a version goes live, try your agent by hand in AI agents → Automated testing: open the agent and choose Try it yourself.
- Web chat
- Write as a customer would, with any version of your agent, drafts included. It uses the same instructions, tools and knowledge as your channels, but nothing is sent and nothing is booked.
- Phone call
- The agent's line calls one of your verified test numbers, and the live version answers, as a caller would hear it.
- Web voice and WhatsApp
- Talk to the live version in your browser through your website widget, or write to its WhatsApp line from a test phone.
Next to it, Scenarios runs the agent's scenarios: simulated customers (personas) talk to the versions you pick, and a judge grades each conversation with the scenario's rubric, its own checks plus the goals and competencies it imports. See the Automated testing API to run them from your code.
Best practices
- Start with one goal
- Focus on your most important outcome first. Once results are consistent, add more goals.
- Make goals specific
- “Book appointment” is better than “Successful call”. Specific goals are easier to measure and improve.
- Review low scores
- Read the conversations that scored lowest or missed a goal in the evaluation logs. Each competency's details list its recent misses.
- Iterate on competencies
- Refine your competencies around what matters most. Start simple and add more as you learn what drives success.
- Set realistic targets
- Start with targets you can reach and raise them as the agent improves. A 20–30% goal rate is often a good starting point.
Continuous improvement loop
Goals and evaluation make a continuous improvement cycle:
Release
Release a version of your agent with its goals defined.
Track goals
Follow each goal's success rate over time.
Evaluate
Review the evaluations and look for patterns in low scores.
Improve
Change the agent's instructions, try the new version and release it.