Is ‘almost’ ever good enough?
Now, I know this sounds like the sort of pseudo-philosophical caption you might see over a photo of a brooding Tom Hardy or Cillian Murphy getting shared on Facebook, but when it comes to customer service and experience, it’s a legitimate area of debate.
And this debate has intensified with the introduction of generative AI.
While the tech has swept through the customer support landscape – promising faster responses, reduced costs, and 24/7 service – many enterprise CX leaders are discovering that it isn’t a panacea.
But does it need to be?
In discussing this dilemma, Gintautas Miliauskas, Mavenoid CEO and Co-Founder, stated:
If you have an 80-90% accuracy rate on generative AI, companies are kind of ok with that as long as they have escalation channels into human support.
However, when the same standard is applied to human agents, the margin for error suddenly feels risky.
This distinction underscores why CX leaders must carefully consider the trade-offs between accuracy and scalability before deploying AI-driven support.
“Enterprises are experimenting with generative AI,” says Gintautas, “but the reality is, in production, the accuracy isn’t always there.”
Why Generative AI Alone Often Falls Short
Part of the allure of generative AI is its flexibility. It can handle open-ended queries and engage in conversational flows that feel natural and intuitive.
Yet, as Gintautas points out, it’s becoming increasingly clear that the accuracy isn’t always there, especially for edge cases or multi-step problem-solving.
This probabilistic nature can create brand risk. Hallucinations or non-governed outputs are particularly dangerous in safety-critical industries, while a lack of source control makes compliance difficult.
“There are just some circumstances that purely require a business or a brand to 100% know what it’s going to say every time,” Gintautas explains.
In other words, for many enterprises, the consequences of a wrong response are too high to rely solely on generative AI.
Other practical limitations include user frustration and workflow dead ends.
Customers aren’t prompt engineers. Queries like ‘doesn’t work’ can break generative-only flows, while clarifier tolerance is low.
Generative AI may explain, but without actions or guided workflows, it leaves customers stuck.
Why Deterministic (Curated) AI Also Has Limits
On the other side of the spectrum is deterministic or curated AI.
This approach emphasizes controlled, pre-defined responses and workflows. It’s ideal for high-stakes, repeatable interactions, such as safety-critical instructions, step-by-step troubleshooting, or regulatory compliance.
However, deterministic AI struggles with scalability.
“Every new issue requires content creation and ongoing maintenance,” Gintautas notes.
Rigid experiences and limited coverage mean curated systems often can’t anticipate long-tail or multi-topic queries.
Moreover, they demand continuous investment from support teams to keep content up to date.