Comparing a game of chess with a customer service interaction may seem unexpected at first. Yet, when you look closely at the structure and progression of both, the analogy becomes surprisingly insightful.
While their objectives differ radically—checkmate vs. customer satisfaction and problem resolution—both follow a similar escalation in complexity as the interaction unfolds.
Conceptual Parallels
| Chess Game | Call Center Interaction | Meaning / Analogy |
| Openness strategy | Call opening / reception. | Set the tone and take control from the start. |
| Tactical combination | Handling of objections. | Quick thinking to turn things around. |
| Late-game accuracy | Closing of the call. | Ensure resolution and satisfaction before finishing. |
| Sacrifice | Offer compensation or a gesture of goodwill. | Short-term loss for long-term gain (loyalty or retention). |
| Checkmate | Customer satisfaction and resolution | Achieve the desired result in an efficient manner. |
| Critical error | Communication error / violation of rules | A costly mistake that affects the results. |
| Pat (blocking) | Deadlock / escalation | Neither side achieved its goal. |
| Time pressure | High call volume periods | Decisions under pressure; Compromise between efficiency and precision. |
Now that the parallel between the 2 activities is clarified, the behavior of the AI on these interactions becomes interesting to observe:
An illustrative example of the limitations of LLMs (large language models) comes from documented experiments with chess.
In March 2024, Chess.com held a showdown between ChatGPT and Google’s Gemini, where both systems could perfectly explain the rules of chess when asked directly, but then violated those same rules repeatedly during the game. Both bots constantly attempted to make illegal moves, and when they were informed of the error, they continued to come up with invalid moves.
Nikola Greb, an NLP data scientist and former ELO 2000+ junior chess champion, played several games against ChatGPT-4 in January 2024 and documented that the model played “like a grandmaster” in the opening first moves, but deteriorated significantly as the game progressed. ChatGPT-4 began to hallucinate, coming up with impossible movements even after being warned. Greb concluded that the overall rating of the system was below 1500, and observed something crucial: “No implicit rule learning has taken place – ChatGPT-4 still hallucinates at chess, and continues to hallucinate after the warning about hallucination. This is something that cannot happen to a human.
This disconnect between what an LLM can “say” and what it can “do” reveals a fundamental limitation: they don’t have real mental models of the world. In the context of customer service, this means that a bot can perfectly recite company policy but apply it incorrectly in specific situations, or it can explain how a product works without being able to diagnose a problem with it.
The Chatbot Chess Tournament 2025
In January 2025, a chatbot chess tournament aired on the GothamChess channel pitted professional chess engine Stockfish against seven generative AI chatbots, including ChatGPT, Google’s Gemini, and X’s Grok. The results were exactly what you would expect when language models try to play chess: decent opening moves followed by increasingly chaotic attempts to circumvent the laws of the game. The Snapchat chatbot decided that the pawns could move sideways like a tower, and when the error was reported, it repeatedly refused to continue saying “I’m sorry. I can’t engage in such a conversation. Let’s keep our conversation respectful.”
The problem of memory and context
LLMs have strict memory limits. While newer models offer wider windows of context, they still treat each conversation as relatively isolated. This means they can “forget” crucial information provided at the beginning of a long conversation, forcing customers to repeat themselves.
In one of the following articles, we will see how to avoid putting the customer in failure while making the best use of the undeniable capabilities of AI…