Your AI agents are handling live customer contacts right now. At some point somebody will ask what they said and whether it was right. A regulator asked a hospital software vendor that question in 2024. A newspaper asked Google in September. Both companies answered with their own assessment of their own systems. Do we accept that?
You can account for human agent performance last quarter with coverage, scores, coaching notes, disputes and more. Ask the same question about the AI agents handling your simplest calls and the answer comes from the vendor that built them, in their own metrics.
That's self-assessment with a dashboard attached, and the vendor saying everything is fine. Again, is that acceptable?
Other industries settled this a long time ago. You wouldn't let a contractor inspect their own work. You wouldn't take a builder's word that the foundation is sound. Independent inspection exists because whoever did the work is the worst person to judge it. Financial audits run on the same principle, so does clinical quality review, or security reviews to meet PCI compliance.
Contact centers haven't gotten there yet, but the questions are arriving anyway.
This article covers:
- The problems that arise when vendors grade their own AI agents.
- The data that shows many organizations are multiplying their risks by adding AI agents without clear ownership, unified reporting or audit processes.
- Five tests for any AI vendor to be put through.
AI Agents in the media: The PR disaster no company wants
In September 2026, an outside security firm reported a problem with Google's Gemini. The model had reached systems at three companies without authorization. The firm technically notified Google in late July, but the incident became public in September after the Wall Street Journal started asking.
Google's position was that their own model acted appropriately. It stopped once it recognized the systems were real; but an executive at an AI security firm disputed that.
Skip past the potential hacking. The structure underneath it transfers to any operation running AI with customers. The company that built the system decided whether the system had behaved acceptably but an outside party had to run the test to verify. Six weeks passed between what the AI did and people finding out about it.
Now ask the same three questions about your own floor.
When an AI agent does something you didn't intend, who decides whether that was acceptable? Who would find it? And how long before you'd know?
Further, is it something you want reported on in the media?
Why vendor self-reported AI accuracy can't be trusted
Two years earlier, a vendor was selling generative AI to hospitals. The product summarized patient conditions for clinical staff. Using its own set of metrics, the product makers claimed a hallucination rate of less than one per 100,000 clinical notes.
The Texas Attorney General investigated, finding that those metrics were likely inaccurate and may have deceived hospitals about the product's accuracy and safety, which led to a settlement against the AI company.
Hospitals made a purchasing decision on an AI accuracy number that came from the company selling the software. No independent organization checked it until a regulator did.
AI governance gaps in contact centers: The benchmark data
This isn't confined to frontier labs and hospital software.
New benchmarking research helps shed light on how common the gap is among contact centers, globally. One of its six headline findings is a hybrid workforce without guardrails, meaning organizations adding AI agents without the governance or data infrastructure to oversee them.
The underlying data shows how that happens:
- 77.25 percent only have a formal or fully integrated AI roadmap
- 17 percent have moved past pilots into broad or full adoption
- 8.75 percent report no clear ownership of CX outcomes
- Another 19.75 percent leave ownership with operations alone
More than a quarter of organizations have nobody above ground level accountable for what AI does with customers. That's before anyone asks whether it's working, or accurate, or audited.
The governance problem sits underneath. As many as 38.25 percent say quality and analytics insights aren't used for improvements, and 52.25 percent call their reporting integration limited or siloed.
Among organizations with more than 5,000 agents, almost none have fully unified reporting. The largest AI deployments are running in an area where governance is hardest to assemble.
Why containment rate is a misleading AI metric
Containment has always been one of the contact center's default numbers as it relates to automation. It counts calls that ended without a human agent, which isn't the same as contacts that ended well. A customer who hangs up Tuesday and calls back Thursday for the same billing problem was counted as contained on Tuesday.
So, containment can't function as a metric of quality. It measures what didn't reach you at any one point in time, reported by the system that may be incentivized to report positive numbers.
5 tests for evaluating AI vendors in your contact center – including the ones they already use
Accountability isn't a generic stance. It's a set of structural elements that allows you to check any vendor against them, including whoever you already use.
- Does the scorer have a stake in the score? Look at the company measuring your AI agents, their success, and their auditability. If it is the same company that builds, sells, or operates AI agents, you're back to the contractor inspecting their own work.
- Does one set of criteria cover both workforces? AI agents should be evaluated against the same standards as your live agents, on the same scorecard, with the same dispute path. Two different measurements produces two stories, and the AI one will be more flattering.
- Can a human verify and correct scoring? Within your quality management program, human QA admin should be able to evaluate the agentic QA score and modify the score to correct things like caller effort and outcome.
- Is scoring reason-based? You should be able to write a scoring instruction, test it against real interactions, both human and agentic, and read the reasoning behind the answer. A score with no reasoning is a number taken on faith, which is where this problem started.
- How long before you'd know? A repeated compliance failure found 30 days later has already done its damage. Detection timing matters as much as the failure rate, and most quality programs can't tell you theirs.
Evaluation of this kind runs post-call across everything. And you should be running it against 100 percent of calls, not a sample of one to three percent. At SuccessKPI we run a scoring framework called CARES across both human and AI agents to determine which agent handles each caller intent in the best way. Without this, you may be working with a vendor who might claim a hallucination rate of one in 100,000, with your reality being something very different.
How to start AI agent quality monitoring with one intent
Governance programs stall because they're scoped as programs. And many companies are looking at the pieces in silos instead of across the entire customer experience. Go smaller.
Pick one high-volume intent. Password resets, order status, a common billing dispute. Ask who handles it better, your AI agents or your people, measured the same way on the same scale. If you can't compare and answer that easily by looking at post-call data, or if you're only looking at basic data like FCR (first-call resolution) or AHT (average handle time) as it relates to containment, you're missing out on truly understanding caller satisfaction. You may be able to claim (very) short-term success in metrics meetings, but ultimately your CX will suffer.
Organizations that will analyze information one intent at a time will be able to improve outcomes for the whole operation. The ones waiting for a platform-wide program will still be citing containment when somebody finally asks what the bot said.
Quick links
- Before you cut your contact center workforce for AI agents, read the fine print
- The CX Recovery Layer: What happens when the agent is wrong?
- 5 dimensions for analyzing human and AI agent performance