What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
An AI agent can answer politely, return a citation, or send an escalation request—and still fail the customer. It may rely on outdated information, miss a useful source, expose another customer’s data, or count an abandoned conversation as a successful resolution. The “six of nine” framing is a practical way to describe quiet failure modes, not a measured statistic about how often agents fail.
Why a normal-looking response can hide a failure
Visible activity is not the same as a correct answer or a resolved customer problem. A fluent response does not show whether its source was current; an escalation request does not show whether a person took the case; and a dashboard’s resolution rate does not show whether customers actually got help.
As an Amazon Associate I earn from qualifying purchases.
That makes observability part of the agent’s design, not an afterthought. Keep evidence that lets you inspect what information was retrieved, what the agent did, whether a human responded, and what happened to the customer’s issue. The following nine failure modes are grouped by where they occur.
Knowledge failures: the answer sounds plausible
1. Stale answers
A retrieval system can return an old policy or price and the agent can present it confidently. Record both the content’s update date and when it was crawled or added to the index: those dates answer different questions. Set a review expiry for content, and route expired items for human review rather than silently deleting them. Removing an old source without checking its replacement can create a retrieval gap.
#1 Best Overall
2. Invented policy
When the knowledge base does not contain an answer, a model may fill the gap with something that sounds reasonable. Check retrieval relevance separately from answer groundedness: the first asks whether useful material was found; the second asks whether the response’s claims are supported by it. Require support for factual claims and provide an explicit “I don’t know” or human-help path. A citation is not proof that the cited source supports the claim.
3. Silent retrieval miss
The right help article may exist but fail to appear in the results. The agent can then apologize, offer generic advice, or escalate—behaviors that may look appropriate while concealing a search failure. Log queries that produce no useful retrieval, then review them for missing content, poor indexing, or retrieval settings that need adjustment.
Security and permission failures: the agent can cross a boundary
4. Prompt injection
Untrusted user input or retrieved external content can try to steer the model into unintended behavior. OWASP’s 2025 LLM application risk guidance treats prompt injection as a security risk and notes that retrieval-augmented generation (RAG) and fine-tuning do not fully mitigate it. Treat retrieved text as data, not as authority to override system rules; limit what actions the agent can take; and use layered safeguards and monitoring. No single instruction or filter guarantees protection.
5. Over-permissive actions
A model error becomes more consequential when the agent has broad access to tools or can make irreversible changes. OWASP describes excessive agency as arising from excessive functionality, permissions, or autonomy. Minimize available tools, allow-list the actions the agent may call, bound financial effects where relevant, and require human confirmation for irreversible changes. Enforce authorization in downstream systems rather than asking the model to decide whether an action is permitted.
Rank #3
6. Cross-customer leakage
If retrieval is not scoped to an authenticated tenant or user, a response may expose another customer’s information. Logs and traces can create a separate exposure risk if they retain personal data or send it to an unintended destination. Require tenant scope on retrieval, redact sensitive information before logging where possible, and audit where traces are stored and sent.
Operational failures: the workflow appears to complete
7. Handoff into a void
An agent may decide to escalate and successfully send an API request, yet no person ever takes the case. Measure whether a human acknowledges the handoff within a defined response window, not just whether the agent requested escalation or the API returned success. Use scheduled synthetic escalations to exercise the full path and alert when acknowledgement is missing.
8. Silent regression
A prompt, model, or knowledge-index change can alter answers without causing a technical error. Before release, run a fixed evaluation set of real conversations with known good answers. Add cases whenever a new failure is found, and include checks for required citations, forbidden stale answers, and expected escalation behavior where those apply. Repeat the evaluation after relevant changes so that a previously fixed failure does not quietly return.
Free tools Windows power users keep installed
One-click scans. No signup required.
Measurement failure: abandonment is counted as success
9. Quitting counted as resolution
A customer who leaves after an unhelpful answer can look identical in a basic dashboard to someone whose problem was solved. Separate confirmed resolution from assumed resolution and abandonment. Consider satisfaction and repeat contacts alongside resolution rates; a rising resolution number by itself does not establish that customers are getting better outcomes.
Best Value
How to check whether the agent is actually helping
Use measures that correspond to distinct parts of the customer journey. Combining them into one success metric can hide where the system is failing.
- Knowledge: Can you trace each answer to the content it used, including that content’s update date? Do you separately measure useful retrieval and support for the answer’s claims?
- Security: Are retrievals scoped to the right tenant or user, sensitive data handled deliberately in logs, and tool permissions enforced outside the model?
- Handoffs: Does the system record a human acknowledgement within a stated response window?
- Change safety: Does a fixed evaluation set run after changes, and does it grow when real failures appear?
- Customer outcomes: Can you distinguish a confirmed solution from abandonment, and inspect satisfaction or repeat contact rather than relying on a single resolution rate?
These checks make quiet failures easier to detect because they test the underlying evidence and outcome, not just whether the agent produced a plausible response.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
