How do you know whether an AI agent is good?
It sounds like a simple question until you try to answer it. A fluent response may contain the wrong fact. A voice can sound natural but interrupt at the wrong moment. A task can succeed once, then fail when the user changes one phrase. An agent may retrieve the right policy and still take the wrong action.
That difficulty is turning Conversational AI evaluation into a valuable career skill.
A Bot Jobs snapshot on 6 September 2026 contained 180 current vacancies. Evaluation appeared across levels and disciplines: an applied AI internship at Razer, a chatbot product internship at Shopee, a dedicated AI voice quality role at Spotify and a principal Voice AI product role at ServiceNow.
This is not a statistical study of the whole labour market. It is a useful specialist signal. Employers are no longer treating evaluation as a final check performed by one quality team. They need product managers, conversation designers, engineers, analysts and customer-facing specialists who can define quality, test it systematically and improve a live service.
For job seekers, that creates an unusually practical opportunity. You can demonstrate evaluation judgement without training a foundation model or having access to a large production system.
Evaluation turns “good” into something a team can improve
Traditional software testing asks whether a known input produces an expected result. Conversational AI is harder because language varies, outputs can change and the system is larger than the model. A voice agent may include speech recognition, turn detection, a language model, retrieval, tools, text-to-speech, routing and human handoff.
Good evaluation creates a shared definition of acceptable behaviour, evidence about failure and a repeatable way to check a change. A designer may define a repair rubric, an engineer may automate regression tests and a product manager may set release thresholds.
Evaluate the task and the experience
One score is rarely enough. A useful evaluation programme separates at least two dimensions.
Did the agent complete the right task?
Task evaluation checks what happened, not merely what the agent said. Depending on the use case, that might include:
- identifying the user’s real goal;
- gathering the required information;
- retrieving an approved source;
- choosing the correct tool and parameters;
- applying policy and permissions correctly;
- completing, declining or escalating the task as intended;
- recording the right outcome in the system of record.
An answer can be polite and coherent while failing every important item on that list. The expected action therefore belongs in the test case alongside the expected language.
Did the interaction work as a conversation?
Experience evaluation examines whether the journey was clear, efficient and appropriate. For chat, that can include relevance, tone, context retention, clarification and recovery. Voice adds timing, pronunciation, accent handling, interruptions, audio quality and the natural use of silence.
Deepgram’s 4 September guide to paralanguage in voice AI makes the point neatly. A transcript preserves words but loses much of their delivery. The guide separates five families: prosody, voice quality, disfluency, timing and silence, plus non-speech vocalisation.
Those details can change the meaning of a turn. “Yes” may be a confident confirmation, a hesitant concession or a brief signal that the user is still listening. A voice evaluation that inspects only the transcript will miss the difference.
Six parts of a credible Conversational AI evaluation
1. Start with the user outcome
Define what success means before choosing metrics. “The agent answered” is not an outcome. “The customer changed an eligible appointment to an available time and received correct confirmation” is.
Also define the boundary. Which requests may the agent complete? Which require confirmation, specialist approval or a human? Evaluation becomes much clearer when the intended authority of the system is explicit.
2. Build a representative test set
Happy-path examples are useful, but they cannot establish quality. Include ordinary requests, paraphrases, missing information, contradictions, policy exceptions, unavailable tools and cases that should be refused or transferred.
For voice, vary speaking rate, background conditions, numbers, names, interruptions and pauses. Deepgram’s 3 September explanation of filler words in speech-to-text shows why small transcription choices matter: cleaning every “uh”, “um” or backchannel may remove information from a live dialogue.
Reflect actual users and conditions. Without production data, state your assumptions and how you would replace synthetic examples with consented, redacted cases.
3. Use rubrics people can apply consistently
Define what each score means and provide examples near the boundary between pass and fail. Compare dates, statuses and tool results with known answers. Use human or model-assisted judgement for clarity, empathy and naturalness, then sample the results and measure reviewer agreement. Disagreement may expose an unclear rubric or a decision that genuinely needs context.
4. Test reliability, not one lucky run
Generative systems vary, so run important cases more than once and report the spread. For latency, include percentiles rather than only an average. For quality, report results by scenario as well as overall. An aggregate score can conceal a serious failure in one high-consequence journey.
5. Diagnose the failing layer
“The agent got it wrong” is not yet an actionable finding. Use an error taxonomy that separates:
- speech recognition or turn-detection errors;
- intent, reasoning or context errors;
- retrieval and source-quality errors;
- tool selection, field mapping or integration errors;
- policy, permission or safety errors;
- response and text-to-speech errors;
- routing, escalation or orchestration errors.
This helps the right person fix the problem. It also stops teams from rewriting a prompt when the real issue is stale knowledge or a failed API.
6. Keep evaluation connected to production
A pre-release suite becomes more valuable when it grows from live evidence. Turn confirmed failures, complaints, unexpected tool calls and poor escalations into regression tests before changing the system.
AWS’s Agentic CX Designer release on 2 September reflects this production focus. Its technical walkthrough describes saved tests, debugging, reviewable versions, rollback and analytics alongside conversation design. The interface is vendor-specific, but the career lesson is transferable: testing, release control and monitoring belong in the build process.
Current jobs show an evaluation career ladder
The refreshed listings show how this skill appears at different career stages.
Applied AI Intern, Voice: Razer
Razer’s Applied AI Intern, Voice role in Pennsylvania includes offline evaluation of models, prompts and pipelines, with measures covering quality, latency and reliability. It shows that a reproducible benchmark and thoughtful failure analysis can be credible early-career technical evidence.
AI Chatbot Product Manager Intern: Shopee
Shopee’s AI Chatbot Product Manager Intern role in Singapore covers evaluation standards, cross-functional review and analysis of poor conversations. It shows that junior product candidates need to connect a bad response with user impact and a measurable decision.
Senior Product Quality Analyst, AI Voice: Spotify
Spotify’s London-based Senior Product Quality Analyst, AI Voice owns rubrics and campaigns across naturalness, tone, pacing, pronunciation, accent and audio quality. It converts language, listening and research skills into repeatable quality operations.
Principal Product Manager, Voice AI: ServiceNow
ServiceNow’s Principal Product Manager, Voice AI in California owns dialogue quality, model selection and evaluation strategy. At senior level, the evidence informs investment, launch decisions and customer promises.
The pattern extends beyond dedicated quality posts. NICE’s remote US Technical Customer Success Manager tests and optimises customer deployments, showing how field issues can become product evidence after release.
Build an evaluation portfolio employers can inspect
Choose one narrow assistant, such as appointment rescheduling or order-status support. Use a prototype, public demo or recorded conversations.
Create an AI agent evaluation dossier with seven parts.
1. Outcome and boundary brief
Name the user, task, available systems and intended outcome. State what the agent may do, what needs confirmation and what must reach a person.
2. Scorecard
Define five to seven measures across correctness, conversation quality, reliability, safety and performance. Explain each one and set a release threshold without pretending one number replaces review.
3. Test set
Write at least 25 cases across routine tasks, paraphrases, ambiguity, missing information, policy exceptions, tool failures, unsafe requests and escalation. For voice, add accents, names, numbers, interruptions and silence.
4. Results
Run important cases three times. Record the expected action, actual action, evidence, score and failure category. Preserve audio or transcripts only with permission.
5. Error analysis
Group failures by layer. Prioritise the two issues with the greatest frequency and user consequence, explaining the likely cause without overstating certainty.
6. Improvement experiment
Change one thing: a rubric, prompt, knowledge source, tool description, threshold or conversation repair. Rerun the affected tests and the wider regression set. Report improvements and regressions.
7. Decision note
Recommend release, a limited pilot or a hold. State the evidence, remaining risk and next test. This shows judgement beyond a spreadsheet of scores.
For a more technical starting point, ServiceNow’s open-source EVA evaluation framework assesses complete multi-turn spoken conversations across task accuracy and conversational experience. You do not need to adopt its exact tooling. Study the separation between outcome and experience, then apply the same principle to your own case study.
Make evaluation visible on your CV/Resume
Avoid writing only “tested chatbot” or “performed QA”. Describe the system, method and decision your work enabled. For example:
- Built a regression set covering routine, ambiguous, policy and integration scenarios for a customer-service agent.
- Defined scoring rubrics for task success, repair quality and appropriate escalation, with reviewer guidance and boundary examples.
- Analysed voice-agent failures across speech recognition, turn-taking, reasoning, tool use and synthesis.
- Compared model or prompt versions using repeated tests and documented improvements, regressions and release risk.
- Converted recurring customer issues into reproducible cases for product and engineering teams.
Use genuine numbers when you have them: cases evaluated, reviewers aligned, error types reduced or task completion improved. Label portfolio work clearly if it was not a production deployment.
Quality is becoming shared career currency
Conversational AI teams do not only need people who can build. They need people who can say what good looks like, find out whether the system meets that standard and explain what should happen next.
The latest vacancies show evaluation moving across the career ladder and job families. Engineers, designers, linguists, researchers, product professionals and customer specialists can approach it differently, but all can build inspectable evidence.
Start with one narrow journey. Define the intended outcome, test the awkward cases and make your reasoning visible.
Then explore current Conversational AI roles on Bot Jobs, including openings in Conversation Design, Engineering, Product Management and ML/Data Science. You can also create a job-seeker profile so specialist employers can find you.
