TL;DR: An AI agent (also called a bot) can contain a conversation without solving the customer's problem, so an AI agent evaluation framework needs to measure resolution as well as containment. evaluagent's AI Agent Observability brings every bot conversation under the same quality framework as your human agents, often with an additional AI agent scorecard. It covers resolution, answer accuracy, policy compliance, escalation quality, and customer experience. The AI agent scorecard also checks every conversation for four common failure modes: hallucinated responses, non-compliant advice, broken handover triggers, and misrouted conversations. A four-step workflow (define, score, fix, report) turns those scores into decisions that leadership and compliance can check.
AI agents are going live in customer service with no scorecard, often evaluated only by the vendor that built them, so an AI agent evaluation framework needs to measure resolution as well as containment.
An AI agent is software that acts on behalf of a user or system to perform tasks. In customer service, it handles conversations on its own by generating answers and taking actions. Containment means the customer didn't reach an agent, so answer accuracy needs a separate check. This article covers what to measure, how to score conversations, how to detect four common failure modes, and how to compare AI agents with human agents under one quality framework.
Evaluating AI agents independently of the vendor
Independent evaluation gives you a view of AI agent performance that doesn't come from the vendor that built the bot. That independence is the basis of evaluagent's AI Agent Observability, which brings every AI agent conversation under the same quality framework used for human agents, often with an additional AI agent scorecard. It flags hallucinated content, off-policy advice, and compliance breaches against your organization's own knowledge base.
Grading every bot conversation against your own definition of good shows whether the bot gave correct answers from your knowledge base and whether customers reached their goal. To test a containment number, pick one intent the vendor reports as contained, compare its repeat contact rate, and read those transcripts to see whether customers got what they needed or gave up.
Vendor reporting is useful for operational metrics such as volume and containment, but it doesn't independently verify answer accuracy or customer outcomes. Containment doesn't tell you whether the problem was fixed.
evaluagent doesn't sell AI agents or contact center as a service (CCaaS) platforms, so it has no stake in grading its own homework.
Defining containment four ways
Resolution matters more than volume because a bot can keep many conversations from reaching a human agent without solving customer problems. Containment measures volume. Resolution measures outcome. A contained conversation that fails to resolve the issue creates a repeat contact and frustrates the customer.
Containment has no single definition. One practical way to break it down is into four, each answering a different question about whether the bot helped.
Compare the four definitions
| Working definition | What it measures | What it can miss |
|---|---|---|
| Bot kept the conversation | No human handover occurred | Whether the issue was solved |
| Bot handled the issue automatically | Task completed without human intervention | Whether the completion was correct |
| Intent resolved | Specific intent reached a resolution state | Whether the customer achieved their goal |
| Customer achieved their goal | The customer's actual outcome | Abandoned conversations and unresolved issues with no follow-up contact |
Each definition builds on the previous one. You need all four to understand what your containment number means.
Audit each definition
- Bot kept the conversation: Pull a random sample your vendor counts as contained and read the transcripts. A conversation with no transfer can mean the customer got the answer, gave up, or came back through another channel.
- Bot handled the issue automatically: Review conversations where the bot claims to have completed a task. Check each claimed completion against the record in your own system.
- Intent resolved: Group conversations by intent to see which resolve cleanly. evaluagent groups bot conversations to provide intent-level performance, which shows which intents the bot handles cleanly and which generate frustration. Fix intents with low resolution and repeated failed handovers first.
- Customer achieved their goal: Look at repeat contacts and predicted repeat contact scores. A customer who comes back with the same problem within 7 to 30 days likely didn't get the issue resolved. In evaluagent's xMetrics, xRepeats is a predicted measure of repeat contact, and it covers conversations where no survey came back.
Choosing AI agent evaluation metrics for resolution, accuracy, compliance, and escalation
Five metrics worth tracking are resolution, answer accuracy, policy compliance, escalation quality, and customer experience. Each needs its own measurement method.
| Technical metric | What it measures | Business outcome | One way to audit |
|---|---|---|---|
| Intent resolution rate | Share of conversations each intent resolves | Automation effectiveness | Group conversations by intent |
| Answer accuracy | Share of responses that match your knowledge base | Customer trust | Grade against your knowledge base |
| Policy compliance | Share of responses that follow your policies | Regulatory compliance | Flag off-policy advice |
| Escalation quality | Handover timing and context | Customer experience | Review handover conversations |
| Customer experience | Predicted satisfaction (xNPS) and repeat contact (xRepeats) from conversation content | Satisfaction | Score conversations with xMetrics |
Measuring resolution, accuracy, and customer experience
Problem resolution is the core metric. Check whether the customer's issue got solved and how you know. A conversation only counts as resolved if the customer achieved their goal. Repeat contact analysis shows where it didn't.
Answer accuracy is separate from containment. Measure it by grading responses against your own knowledge base and procedures. A bot can contain a conversation by giving a wrong answer that satisfies the customer in the moment. Accuracy checking catches this by comparing every bot response against your source of truth. The Context Engine evaluates whether agents give the correct answer by grounding every AI score in your own knowledge base. Models often hallucinate when they lack pertinent knowledge. Flagging answers that don't trace back to your documentation catches fabrication.
Customer experience on automated conversations needs its own measurement. Use predicted customer metrics and sentiment analysis on bot conversations. Predicted metrics such as xNPS and xRepeats cover every conversation, even when no survey came back.
Measuring policy compliance and escalation quality
Policy compliance must be checked on every conversation. Measure it by flagging off-policy advice against your own policies. In UK financial services, the Financial Conduct Authority (FCA) Consumer Duty requires firms to deliver good outcomes for retail customers. Firms must assess, test, understand, and evidence those outcomes on an ongoing basis. Firms must also regularly monitor the outcomes retail customers receive from customer support. evaluagent flags off-policy responses and compliance breaches on bot conversations, with an exportable audit trail.
Escalations are a quality signal. Audit them by checking whether the bot handed over at the right moment with the right context. Gartner reports that 87% of customers say it's essential to be able to reach a human agent when a company uses GenAI for customer service, and advises against making AI a mandatory first step. A good handover gives customers a clear path to a human agent and passes the human agent the context to continue. A failed handover happens when the bot hands over too late or without context. Review handover conversations and ask whether the human agent had everything they needed.
Detecting four common AI agent failure modes
The AI agent scorecard checks for four common failure modes, and each needs its own detection method.
- Hallucinated responses: fabricated content that sounds plausible but is wrong
- Non-compliant advice: responses that violate your policies or regulatory requirements
- Broken handover triggers: escalations at the wrong time or with wrong context
- Misrouted conversations: customers sent to the wrong place or handled by the wrong intent
A study of large language model (LLM) agent failures found that early mistakes cascade into later steps.
Identifying hallucinated and non-compliant responses
Hallucinated responses are fabricated content that sounds plausible but is wrong. Identify them by grading responses against your own knowledge base. Hallucination happens when the bot generates an answer with no basis in your documentation. evaluagent's fabrication detection flags bot responses that invent information by grading each response against your organization's knowledge base.
Non-compliant advice violates your policies or regulatory requirements. Identify it by checking every response against policy documents. evaluagent checks required disclosures and compliance language on every conversation, with an exportable audit trail.
Fixing handover and routing failures
Broken handover triggers cause frustration. Fix them by identifying which intents trigger handovers, reviewing the context, and adjusting escalation rules. Start by listing intents that trigger handovers most often, then review a sample and check timing. If the bot hands over too early, it wastes agent time, and if it hands over too late, it frustrates customers. evaluagent can help you surface escalation triggers so you can improve handover quality.
Misrouted conversations send customers to the wrong place, so identify them by reviewing transcripts for intent mismatches and routing errors. A misrouted conversation happens when the bot misunderstands the customer's intent. Review conversations where the customer had to correct the bot multiple times, because repeated corrections point to an intent the bot keeps misreading.
Using one framework for human and AI agents
You don't need a separate quality program for AI agents. The same quality standard, your definition of good, applies to every conversation, whether a human or an AI agent handled it. Both are held to the same core questions: did the agent give the right answer, follow policy, and resolve the issue? What differs is how they fail. AI agents can invent answers, hold on to conversations they should hand over, or misread what the customer wants, so their scorecards need criteria built to catch those failures.
Adding an AI agent scorecard
Many teams add an AI agent scorecard under the same framework, with the criteria human agents don't need. Human agents don't hallucinate, so fabrication checks belong on the AI agent scorecard alone. Build it around hallucination detection, off-policy checks, and handover quality, the failure modes covered earlier.
Score bots on your main scorecard as well. Its core questions give you a like-for-like comparison between bots and human agents. Whether your bot comes from an AI agent vendor such as Cognigy, Sierra, or Decagon, or from your own team, evaluagent holds every conversation to the same definition of quality. Side-by-side bot vs human reporting in SmartView shows both agent types on one framework.
Auditing bot interactions and unifying quality metrics
Every production conversation should be evaluated. AI Agent Observability grades every response against your source of truth. evaluagent can show automated scoring working on your own conversations within days, and deployment takes weeks, not months.
One definition of quality across all bots, regardless of vendor, gives you portable quality history and cross-vendor comparison. If you run bots from different vendors, each reports in its own way. Cross-vendor scoring grades conversations from different bot vendors and in-house bots against the same quality definition. Scores and trend reporting sit in evaluagent. When you switch vendor, the historical quality record comes with you.
Establishing your AI agent governance workflow
Governance is the workflow that turns evaluation into action. The model is define, score, fix, report. Each step feeds the next, and the cycle repeats as your AI agents change.
Define quality benchmarks and score every conversation
Start by defining what good looks like for your AI agents: upload your knowledge base, procedures, and policies, then write guidelines describing what a good bot response looks like. The Context Engine attaches your documents to the line items on your AI agent scorecard. The Line Item Builder converts your existing scorecards into draft automated line items, so you don't need to write prompts yourself.
Expect some setup: scorecard criteria for bots need testing against real conversations, and AI scores can run hidden while your team calibrates them against human reviewers.
Then score every conversation for accuracy and risk. Automated scoring with explainable evidence grades every conversation, and human review focuses on the riskiest interactions. Every score shows the evidence and rationale behind it. You can see the conversation passage behind each score and the rule it mapped to. Auto Work Queues point human review at the conversations most likely to carry risk. Human review is configurable to fit your program.
Fix failures and report results to CX leadership
Use the scores to fix what is broken. Identify which intents fail, which prompts need work, and which content is missing. The intent-level view shows where to start. If an intent fails because the bot lacks information, add it to your knowledge base. Rewrite unclear prompts. Adjust the escalation rules when the bot hands over too late.
Then report the evidence to leadership. Trend history, the bot vs human comparison, and the audit trail give leadership the proof they need. The audit trail records compliance checks. evaluagent logs every human override with who, when, and why. Leadership can see whether automation is improving and which intents are ready to expand.
Benchmarking AI agents against human performance
Humans and bots can use the same or different scorecards, and evaluagent still unifies them under one framework so you can meaningfully compare performance.
Benchmarking answers the question leadership asks: are the bots as good as the best human agents?
- Compare AI agents with your human teams: Score AI agents on the same main scorecard as your human agents to get a like-for-like comparison, criterion by criterion. Then add an AI agent scorecard to catch what the main one won't: invented answers, off-policy advice, and poor handovers.
- Compare intent resolution: Compare bot intent resolution rates against human performance on the same intents. The intent-by-intent comparison shows you which intents to assign to bots and which to route to humans.
- Identify intents ready for expansion: An intent is ready to expand when it resolves a high percentage of conversations, shows few repeat contacts, and has strong handover quality. Fix intents with low resolution rates before you scale them.
Auditing your AI agent vendor's performance
Four questions show what your vendor's reporting covers and what it leaves out.
- Define containment: Ask your vendor to define containment. Does it mean the bot kept the conversation, or that the customer's issue was resolved? Ask whether the vendor's metric requires proof of resolution or only proof that no handover occurred.
- Ask how they detect hallucinations: Ask whether the vendor grades responses against your knowledge base and can show the evidence. Ask whether evaluation checks every response against your documentation. Confidence scores alone fall short, because LLMs can fabricate answers with unwarranted certainty.
- Ask how they check policy compliance: Ask whether the vendor flags off-policy advice and compliance breaches on every conversation. Ask whether it checks every conversation or only a sample, and how it catches breaches in unsampled conversations.
- Ask about data portability: Ask whether you can export your quality history, scores, and audit trails in a usable format if you switch vendors. Your quality history is your evidence, and losing trend data means losing the ability to show improvement.
Building an independent AI agent evaluation framework
Containment shows how many conversations the bot kept from reaching a human agent. The five metrics above show whether those conversations helped customers. Scoring all five under the same framework as your human agents gives leadership and compliance evidence of how the AI agents perform. The define, score, fix, report workflow then shows which intents are ready to expand.
See how evaluagent's AI Agent Observability evaluates a week of your own bot conversations against your definition of good. Book a demo.
FAQs
How is evaluating AI agents different from human QA?
AI agents sit under the same quality framework as human agents. Scoring them on your main scorecard gives a like-for-like comparison on whether the agent gave the right answer, followed policy, and resolved the issue. An additional AI agent scorecard covers failures only bots make, such as hallucination and broken handovers.
Can I compare bots from different vendors?
Yes. evaluagent's cross-vendor scoring grades conversations from different bot vendors against one definition of quality, so you can compare performance directly. Your quality history stays in evaluagent when you switch vendors.
How do I show compliance how our AI agent is checked?
Evaluate every conversation for policy compliance, flag breaches, and keep an audit trail of the checks. evaluagent's AI governance features include labeled AI outputs and configurable human review from 0-100%. Every override is logged with who, when, and why.
What's the difference between containment and resolution?
Containment counts the conversations the bot kept without human handover, and resolution checks whether the customer's issue was solved. A bot can contain a conversation without resolving it.
Do I need to evaluate every bot conversation?
Yes. Sampling misses hallucinations and compliance breaches. evaluagent's AI Agent Observability evaluates every AI agent conversation under the same quality framework used for human agents.
Key terms glossary
AI agent: Automated software that handles customer conversations on its own, generating answers and taking actions. This article also calls it a bot.
AI Agent Observability: evaluagent's product for independent evaluation of AI agents and bots, brought under the same quality framework applied to human agents, often with an additional AI agent scorecard.
Containment: The percentage of conversations the bot handled without transferring to a human agent, which is the most common working definition. Containment doesn't measure whether the issue was resolved.
Resolution: Whether the customer's issue was solved in the conversation.
Quality framework: The business's agreed definition of what good looks like, covering the things that matter enough to measure. Several scorecards can sit under one framework.
Scorecard: The set of line items used to grade one type of conversation. Teams often run a main QA scorecard, a vulnerability scorecard, and an AI agent scorecard under one framework.
Answer accuracy: The share of bot responses that match the organization's knowledge base and procedures. evaluagent measures it by grading each response against your own source of truth.
Policy compliance: The share of bot responses that follow the organization's policies and regulatory requirements. evaluagent measures it by flagging off-policy advice and compliance breaches on every conversation.
Escalation quality: Whether the bot handed a conversation to a human agent at the right moment with the right context. It is measured by reviewing handover conversations.
Customer experience: How customers rate or are predicted to rate a bot conversation. evaluagent measures it with predicted metrics such as xNPS and xRepeats, plus sentiment analysis, on conversations where no survey came back.
Hallucination: Fabricated content generated by an AI agent that sounds plausible but has no basis in the organization's knowledge base or documentation.
Off-policy response or non-compliant advice: A bot response that violates the organization's own policies or regulatory requirements.
Handover: The transfer of a conversation from a bot to a human agent. Handover quality measures whether the transfer happened at the right time with the right context.
Misrouted conversation: A conversation where the bot misunderstands the customer's intent and sends the customer to the wrong place or handles it under the wrong intent.
Intent: The purpose or goal behind a customer's message. Intent-level reporting shows which intents resolve cleanly and which cause frustration.
Context Engine: evaluagent's capability that grounds AI scoring in the customer's own knowledge base, procedures, and definition of good.
Line Item Builder: evaluagent's feature that converts existing scorecards into draft automated line items, so QA teams start from drafts they review and edit.
Auto Work Queues: evaluagent's feature that points human review at the conversations most likely to carry risk.
SmartView: evaluagent's dashboard capability that shows bot and human results side by side for leadership reporting.
xMetrics: evaluagent's predicted customer metrics, such as xNPS and xRepeats. Each is a predicted score generated from conversation content when no survey came back.