When Bots Argue: How AI Agents Are Escalating Chaos in Customer Service and Hiring
We are entering a new reality, one where AI agents are no longer just assisting humans, but negotiating directly with each other on our behalf. Customer service bots are disputing claims, escalating complaints, denying refunds, and enforcing policies autonomously, often without meaningful human oversight. AI avatars and digital clones are interviewing candidates, looping, glitching, and freezing in the middle of conversations.
What looked like a dystopian sci-fi movie just two or three decades ago has now become our reality. The problem is that AI agents still lack human judgment and the ability to read between the lines. They can process enormous amounts of data, follow rules, and optimize for efficiency, but they often struggle with nuance, context, emotion, sarcasm, urgency, and ethical gray areas that humans navigate instinctively every day.
As more AI agents begin interacting directly with other AI systems, these gaps become even more dangerous. Two bots can reinforce each other’s rigid logic, intensify conflicts, or create deadlocks without anyone realizing the situation is spiraling.
What looked like a dystopian sci-fi movie just two or three decades ago has now become our reality. The problem is that AI agents still lack human judgment and the ability to read between the lines. They can process enormous amounts of data, follow rules, and optimize for efficiency, but they often struggle with nuance, context, emotion, sarcasm, urgency, and ethical gray areas that humans navigate instinctively every day.
As more AI agents begin interacting directly with other AI systems, these gaps become even more dangerous. Two bots can reinforce each other’s rigid logic, intensify conflicts, or create deadlocks without anyone realizing the situation is spiraling.
Customers using bots to negotiate telecom discounts
Large telecom providers are already facing these issues and adjusting to this new agent-to-agent economy. For example, Trevor Mata, Lead Digital Growth Strategy at AT&T, shared this story:
“A customer deployed an AI agent to negotiate a billing credit. The business’s AI agent met them in the chat, analyzed the account, and offered a baseline offer. The customer’s AI decided the baseline wasn’t enough. It didn’t just argue; it escalated. To bypass the chat bot’s logic, the customer’s AI switched to voice generation. It called in, spoke to a real human agent, and used natural-sounding speech to push for an optimized demand.”
This raises the issues of efficiency versus exploitation. Businesses are concerned about customers exploiting the system and deploying bots to “farm credits at scale.”
As Trevor points out, “the goal of AI in CX shouldn’t be to out-negotiate our customers.” In this new environment, the winners will not be the companies with the most aggressive bots, but the ones that build systems fair enough that customers do not feel the need to deploy one in the first place.
“A customer deployed an AI agent to negotiate a billing credit. The business’s AI agent met them in the chat, analyzed the account, and offered a baseline offer. The customer’s AI decided the baseline wasn’t enough. It didn’t just argue; it escalated. To bypass the chat bot’s logic, the customer’s AI switched to voice generation. It called in, spoke to a real human agent, and used natural-sounding speech to push for an optimized demand.”
This raises the issues of efficiency versus exploitation. Businesses are concerned about customers exploiting the system and deploying bots to “farm credits at scale.”
As Trevor points out, “the goal of AI in CX shouldn’t be to out-negotiate our customers.” In this new environment, the winners will not be the companies with the most aggressive bots, but the ones that build systems fair enough that customers do not feel the need to deploy one in the first place.
AI-to-AI handoffs can compound misinformation
Another customer story comes from Victor Smushkevich, Founder and CEO of Call Setter AI:
“We ran a test deployment where our AI voice agent answered an inbound call from a homeowner whose prior contact had been handled by the HVAC company’s outbound AI dialer (a different vendor). Both agents thought they had context about the appointment. Neither did. The homeowner asked a simple question: “Is the technician still coming?” Within three exchanges, the two AI systems had given contradictory ETAs because each was hallucinating from a slightly different memory of the prior call. The customer ended up frustrated, called the human dispatcher, and we lost the booking.”
Victor believes the biggest lesson for CX leaders deploying AI on both ends is not the risk of AI-versus-AI conflict, but what he calls “AI plus AI compounding ambiguity.” As he explains, “Two confident systems with imperfect memory will produce a customer experience worse than either system alone, because each one anchors the other on bad context.”
He recommends the following operational principle: “Whenever an inbound AI agent detects that the prior touchpoint was also AI-handled, confidence should be lowered and the interaction routed to a human checkpoint before any new commitment is made.” This approach introduces timely human oversight, helping resolve issues before inconsistencies escalate.
“We ran a test deployment where our AI voice agent answered an inbound call from a homeowner whose prior contact had been handled by the HVAC company’s outbound AI dialer (a different vendor). Both agents thought they had context about the appointment. Neither did. The homeowner asked a simple question: “Is the technician still coming?” Within three exchanges, the two AI systems had given contradictory ETAs because each was hallucinating from a slightly different memory of the prior call. The customer ended up frustrated, called the human dispatcher, and we lost the booking.”
Victor believes the biggest lesson for CX leaders deploying AI on both ends is not the risk of AI-versus-AI conflict, but what he calls “AI plus AI compounding ambiguity.” As he explains, “Two confident systems with imperfect memory will produce a customer experience worse than either system alone, because each one anchors the other on bad context.”
He recommends the following operational principle: “Whenever an inbound AI agent detects that the prior touchpoint was also AI-handled, confidence should be lowered and the interaction routed to a human checkpoint before any new commitment is made.” This approach introduces timely human oversight, helping resolve issues before inconsistencies escalate.
Machines misunderstanding the intent
Buyer intent is often treated like the pot of gold at the end of the rainbow—highly coveted, but rarely as clear or attainable as it seems. Anyone who uses LinkedIn Sales Navigator has seen firsthand how imperfect its buyer intent feature can be. AI still struggles to identify true buyer intent, and that is exactly what our next story explores, shared by Marc Bishop, Director at WYTLABS:
“A software vendor tested autonomous renewal outreach with midmarket accounts nearing annual contract decisions. One buyer used a sourcing agent to compare feature gaps against lower-priced competitors in real time. The vendor’s renewal agent responded by defending roadmap commitments and discount limits from prior quarters. Soon, both systems were debating future intent, which no machine can validate with certainty.
I found the danger was not hallucination but negotiated fiction becoming part of the record. The sourcing agent treated aspirational product language as a binding promise, while the vendor agent treated policy templates as settled law. Human sellers usually soften ambiguity with judgment, tone, and reciprocal trust built over time. Agentic negotiations need narrow scopes, approved fallback offers, and immediate handoff once contradictions appear.”
“A software vendor tested autonomous renewal outreach with midmarket accounts nearing annual contract decisions. One buyer used a sourcing agent to compare feature gaps against lower-priced competitors in real time. The vendor’s renewal agent responded by defending roadmap commitments and discount limits from prior quarters. Soon, both systems were debating future intent, which no machine can validate with certainty.
I found the danger was not hallucination but negotiated fiction becoming part of the record. The sourcing agent treated aspirational product language as a binding promise, while the vendor agent treated policy templates as settled law. Human sellers usually soften ambiguity with judgment, tone, and reciprocal trust built over time. Agentic negotiations need narrow scopes, approved fallback offers, and immediate handoff once contradictions appear.”
AI misfires in hiring
Vaibhav Kakkar, Founder and CEO of Digital Web Solutions, shared this story about an AI fiasco during a hiring interview:
“We saw a clear failure in an AI driven interview workflow. A candidate joined a screening call with a video avatar as the hiring manager. The system pulled answers from a wrong profile and asked about experience never claimed. When the candidate tried to correct it, the system saw it as evasive and reduced trust. The impact was not only technical,” said Vaibhav. “It broke the human connection. We would expect a person to notice this, but the system did not. By the time we reviewed the session, the candidate had already shared it in public. We learned that, when identity confidence drops, the system should pause and pass control to a human.”
“We saw a clear failure in an AI driven interview workflow. A candidate joined a screening call with a video avatar as the hiring manager. The system pulled answers from a wrong profile and asked about experience never claimed. When the candidate tried to correct it, the system saw it as evasive and reduced trust. The impact was not only technical,” said Vaibhav. “It broke the human connection. We would expect a person to notice this, but the system did not. By the time we reviewed the session, the candidate had already shared it in public. We learned that, when identity confidence drops, the system should pause and pass control to a human.”
Key takeaways for CX leaders
For CX leaders, the lesson is simple: Efficiency is not enough. In an agent-to-agent world, speed without judgment scales confusion, friction, and bad outcomes fast. The winners will not be the companies with the toughest bots, but the ones with the clearest guardrails, smartest escalation paths, and the discipline to know when a human needs to step in.
These stories also reveal a deeper shift in identity and accountability. When an AI agent speaks for a customer, who is the business really serving: the person, or the system acting in their name? That tension will shape the next era of CX, and the companies that lead will be the ones that build around trust, context, and human override from the start.
CX leaders should also prepare for a new kind of customer behavior: AI bots that negotiate relentlessly, spot policy loopholes, and extract discounts or credits at a scale no human could match.
Ultimately, the goal of AI in customer experience should not be to out-negotiate or outmaneuver customers. It should be to remove friction and deliver fair resolutions so quickly that there is no need for an intermediary bot in the first place.
These stories also reveal a deeper shift in identity and accountability. When an AI agent speaks for a customer, who is the business really serving: the person, or the system acting in their name? That tension will shape the next era of CX, and the companies that lead will be the ones that build around trust, context, and human override from the start.
CX leaders should also prepare for a new kind of customer behavior: AI bots that negotiate relentlessly, spot policy loopholes, and extract discounts or credits at a scale no human could match.
Ultimately, the goal of AI in customer experience should not be to out-negotiate or outmaneuver customers. It should be to remove friction and deliver fair resolutions so quickly that there is no need for an intermediary bot in the first place.
Related resources
Want more like this resource? Find helpful links below.
