The pressure to deploy AI agents in customer service is intense. Some could blame Klarna for starting their frenzy when they announced they were replacing ⅔ of their customer service reps with AI agents in 2024 (then quietly rolled it back in 2025). But no matter where it started, it means every earnings call, every board meeting, every industry conference reinforces the same message: automate with AI or fall behind. The promise is compelling: handle more volume with fewer people, reduce costs, improve response times. But here’s what everybody forgets: expensive mistakes can also multiple when you automate the wrong things.
Where this list came from
Loris has analyzed more than a billion customer service interactions. After analyzing over 500 million customer service conversations, we’ve identified clear patterns in where AI agents succeed and where they catastrophically fail. The difference isn’t about technology capability. It’s about understanding which customer interactions are genuinely automatable and which ones require human judgment, empathy, and expertise.
This isn’t an anti-AI argument. Loris is an AI company and we believe AI agents absolutely have a place in modern customer service. But that place is specific, and the easier it is to switch on AI agents across your entire customer service operation, the higher the cost of getting it wrong.
Here are the five types of customer service issues you should never hand to an AI agent, no matter how sophisticated your technology.
1. Regulatory Complaints and Legal Threats
An AI agent can detect frustrated customers. That’s the easy part. What it can’t do is distinguish between someone venting about a policy and someone describing a pattern of issues that constitutes a regulatory violation. Depending on how it’s designed, it may try to deflect legitimate customer concerns instead of escalation to a human agent or a supervisor. That nuance matters.
Regulated companies may have strict obligations to log complaints for regulatory reporting. A customer who says “I’m going to report this” is obvious. But what about the customer who describes repeated unauthorized charges without using the word “complaint”? Or the one whose issue clearly violates UDAAP standards but never mentions the BBB or CFPB?
AI agents miss these nuances because they’re trained to predict the right patterns. If they haven’t seen this before they won’t understand the legal implications or regulatory context. When a regulatory complaint gets routed through standard AI workflows, the organization often doesn’t realize what happened until weeks later when the formal complaint arrives.
The real cost isn’t just the missed escalation. It’s the audit that follows, the compliance review, the potential fines, and the executive time spent explaining to regulators why your systems failed to flag an obvious issue. One mishandled regulatory complaint can cost more than a year of agent labor.
Some organizations try to solve this with better AI training, adding more compliance keywords to their models. But this creates a different problem: over-flagging. When your AI starts marking every frustrated customer as a potential regulatory issue, your compliance team drowns in false positives and the system becomes useless.
The smarter approach is two-fold: 1) use AI to analyze both human and AI agent communications for potential violations in a way that’s consistent and 2) route potentially risky interactions directly to your human agents human from the start. But make sure to train your AI to recognize when a conversation might involve regulatory risk and immediately route it to a human agent with compliance training.
2. High-Value Customer Churn Signals
Your AI agent sees a cancellation request. What it doesn’t see is that this customer has been with you for five years, represents thousands in annual revenue, and is having their first bad experience after dozens of positive ones.
A major food delivery platform discovered this the hard way. Their premium customers, the ones with the longest order history and highest lifetime value, had a 100% reorder rate within 30 days after positive support interactions. But when those same premium customers had negative experiences, that reorder rate dropped to 91%.
Think about what that means. Your most valuable customers are the most sensitive to poor service experiences. And AI agents, by design, provide the same level of response to everyone. They don’t factor in customer lifetime value, purchase history, or relationship tenure when deciding how to handle a cancellation request.
Human agents do this instinctively. They see the account history, recognize a valuable customer, and adjust their approach accordingly. They might offer a higher-tier resolution, escalate to a retention specialist, or simply invest more time understanding what went wrong. An AI agent follows the cancellation script.
The math makes this even more obvious. If you’re optimizing for cost per contact, automating cancellation requests looks like a win. If you’re optimizing for customer lifetime value, it’s a disaster. The $3 you save on the AI interaction costs you thousands in lost future revenue.
Some companies try to solve this by giving AI agents access to customer value scores and special retention playbooks for high-value accounts. But this assumes the AI can effectively execute a retention conversation at a time when customer emotions may be running high. Which brings us back to the core problem: retention isn’t a transaction, it’s a negotiation. It requires reading between the lines, understanding unstated concerns, and building trust in real time. Those aren’t AI strengths.
Forecasting Your AI-to-Human Agent Ratio
Get a clear framework that tells you where you should apply AI agents – and where human agents work best – to maximize your AI investment without sacrificing customer experience quality.
3. Complex Troubleshooting Requiring Judgment Calls
Decision trees work beautifully when customer problems follow predictable patterns. But self service is now handling more and more of the straightforward inquiries. If a customer has tried all those, the conversation is probably more complex.
An AI agent can walk someone through password reset steps or guide them through basic account configuration. These are linear processes with clear success criteria. But the moment troubleshooting requires judgment about what might be causing an issue, AI agents struggle.
Consider a customer reporting that a feature “isn’t working right.” That could be a technical error message. But it could also be a gap between expectation and reality. A skilled human agent asks clarifying questions, tests hypotheses, and adapts their approach based on the customer’s technical sophistication and patience level. They might realize what the customer thought the issue was has an entirely different root cause, like a service not working because they didn’t renew their subscription.
AI agents can’t always make these kinds of inferential leaps. They match the customer’s description to known issues in their training data and suggest solutions. When the problem doesn’t match a known pattern, they either guess wrong or escalate. Both outcomes are expensive.
The wrong guess is worse than no guess. The customer invests time trying a solution that was never going to work, their frustration compounds, and when they finally reach a human agent, that person inherits an angry customer halfway through a failed troubleshooting process. Your Average Handle Time goes up, your CSAT goes down, and you’ve wasted everyone’s time.
Organizations often discover this pattern too late. They deploy AI agents for technical support, see decent initial metrics, then notice that escalation rates are climbing and customer satisfaction is falling. What’s happening is that AI is successfully handling the easy 60% of issues and systematically failing at everything else. The human agents are left with a concentrated pool of difficult problems and frustrated customers.
4. Emotionally Charged Situations (Even Simple Ones)
Sentiment analysis can tell you someone is upset. It cannot tell you what they need.
A customer contacts support angry about a delayed delivery. On paper, this is simple: apologize, explain what happened, offer a resolution. An AI agent can execute this script perfectly. But here’s what the script doesn’t capture: why this particular delay matters so much to this particular customer.
Maybe the delivery was a birthday gift. Maybe they rearranged their schedule to be home for it. Maybe this is the third delay in a row and they’re questioning whether to continue doing business with you. The factual resolution (refund, discount, expedited replacement) might be identical across all three scenarios, but the conversation required is completely different.
Human agents read emotional context and adjust accordingly. They know when a customer needs to vent before they’ll accept a solution. They recognize when someone is upset about something bigger than the stated issue. They can tell the difference between “make this right” and “I need you to understand why this hurt me.”
AI agents optimize for resolution speed. Emotionally charged customers need to be heard first and helped second. When you reverse that order, even perfect solutions feel inadequate.
The data on this is striking. For new customers having their first support interaction, a negative experience with an AI agent resulted in a 10% reorder rate. A positive experience with a human agent resulted in a 40% reorder rate. Same issue, same resolution, four times the retention based solely on the quality of the emotional connection.
Some organizations try to solve this by training AI to recognize emotional escalation and route to humans. This works better than letting AI handle the entire interaction, but you’ve still subjected a customer to a frustrating AI experience before they get real help. First impressions matter.
5. Product or Service Recommendations That Drive Revenue
AI agents can suggest and guide. But they can’t sell. And even if they can, do customers want AI agents to sell to them? Probably not, or at least not yet.
There’s a meaningful difference between answering “which plan includes feature X” and helping a customer understand which plan is actually right for them. The first is information retrieval. The second is consultative selling.
Good sales conversations involve understanding context that the customer may not have explicitly stated. What are they trying to accomplish? What have they tried before? What’s their technical sophistication? What’s their budget sensitivity? How does this purchase decision connect to their broader needs?
Human agents gather this information through natural conversation and building rapport. They pick up on cues, ask follow-up questions, and tailor their recommendations to the specific customer context. AI agents pull from product documentation and feature matrices. They can tell you what’s included in each tier. They can’t tell you which tier matches your unexpressed needs.
The revenue impact is direct. When customers get consultative guidance, they’re more likely to upgrade, add complementary products, and stay with the higher tier long-term. When they get feature comparisons, they optimize for immediate cost and miss the bigger picture.
Organizations automate these conversations because they look like information exchanges. “Which plan has X” seems like a straightforward question. But customers asking about features are often really asking “will this solve my problem” or “is this worth the upgrade.” Those are sales conversations disguised as support questions.
The cost of automation here isn’t just the lost upsell in the moment. It’s the customer who buys the wrong tier, has a poor experience, and churns. It’s the expansion revenue that never materializes because nobody helped the customer understand what was possible. You’ve traded efficiency for growth.
The Pattern Across All Five
Notice what these situations have in common. They’re all cases where a good customer service agent knows what to do instinctively, no matter which industry they’re in. These are also situations where the cost of getting it wrong far exceeds the cost of the interaction. AI agents are genuinely powerful for high-volume, low-risk, transactional work. The moment you introduce emotional complexity, financial stakes, or regulatory risk, the ROI math flips.
This creates a clear framework for automation decisions. Don’t ask “can AI handle this technically?” Ask “what happens if AI gets it wrong, and can we afford that outcome?”
For password resets, order tracking, and basic account updates, the downside of AI mistakes is minimal. Worst case, the customer tries again or escalates to a human. For regulatory complaints, churn signals, and revenue conversations, the downside is severe and often irreversible.
What to Do Now
The AI agent market is dynamic. As foundational AI model providers improve their AI, the AI agent providers benefit. But the solution isn’t to wait around until things are perfect. It’s to understand where you can deploy AI strategically for use cases it can handle and that you want it to handle.
But before you automate anything, you need to know what use cases are the right fit. Otherwise, you’re experimenting with your customer base. The way to avoid trial and error is to analyze all your customer service interactions, both human and AI powered, to understand patterns. Which use cases have the fewest exchanges? And in which ones does sentiment typically improve or stay flat? What percentage of each issue type gets resolved versus escalated?
With this visibility, you can make informed decisions about automation instead of guessing. You can route simple transactional work to AI and reserve human attention for situations that require judgment, empathy, or expertise.
This is how the best customer experience teams are approaching AI. Not as a replacement for human agents, but as a way to make human agents more effective by removing the work that doesn’t require their capabilities. The goal isn’t maximizing automation percentage. It’s maximizing the value delivered per interaction while managing cost.
AI agents will continue improving. But the fundamental constraints around judgment, empathy, and risk assessment aren’t a bug. They’re human traits that are incredibly hard to emulate. The organizations that understand this and deploy AI accordingly will deliver better customer experiences at lower cost than those chasing automation for its own sake.