A Gartner VP briefs CX pros on selecting conversational AI providers

Before you start talking to vendors about conversational AI for CX purposes, Adrian Lee would like to have a word.

Speaking at a virtual event hosted by Engageware this week, the vice-president and analyst with Stamford, Conn.-based market research firm Gartner Inc. warned of AI agent sprawl, security issues and other pitfalls that make deploying the technology difficult for CX leaders.

This is not to suggest Lee and his team don’t see generative and agentic AI as viable solutions for CX teams. In fact, he said that by next year Gartner expects up to 70 per cent of customer support interactions to happen through conversational AI tools that will be pushed into production.

Among the near-term challenges at the moment, Lee said, was the breadth of choice CX leaders need to assess as they evaluate the potential for their brand.

“There are over 3000 conversational AI solution providers out there, all of them promising great customer experiences delivered through their AI solutions and platforms, but this does not make your decisions any easier,” Lee said.

Those vendors share at least one thing in common: a tendency to downplay the work involved in getting conversational AI up and running. Lee showed a very busy slide with ten different boxes carrying labels like “dialogue processing” and “organization” integration with various arrows connecting them.

“I guarantee you this: every time you speak to a vendor, you’re going to see a very simplified marchitecture in terms of what the conversational AI platform is,” he said. “These are complex solutions and require complex architectures as well, so beyond just thinking about the design and also which language models are being employed in the use of your conversational AI experience, there are many other components.”

Some providers may even be guilty of “agent washing,” Lee said, where they suggest there is one AI agent to handle everything. CX leaders should pay close attention to details such as what business processes are being invoked in a conversational AI interaction, what knowledge bases are being drawn upon and what kind of data pipelines need to be set up to make AI agents successful.

Much like a valued human employee who gains institutional knowledge over time, Lee said conversational AI should never be set up to treat customer incidents in isolation.

“That conversational AI agent should be able to reason continuously, learn, and be proactive,” he said. “How do you be proactive without being creepy or invasive? It comes down to timing, and it comes down to design.”

See Also

50% of consumers cite lack of empathy as biggest AI pain point

Three key areas Lee said CX leaders should ask about in selecting conversational AI tools include:

1. Input integrity: How does a particular solution provider help you to manage any kind of prompt injection? This represents one of the most common cyber-threats, Lee said.
2. Content moderation: How do you filter out hate speech. How do you filter mentions of competitors within that conversational experience? Lee said vendors should have a full set of controls that govern everything from personally identifiable information (PII), personal health information (PHI) and other forms of data protection. These should be turned on by default rather than requiring CX teams to go to extra efforts to dig them out and activate them, Lee said.
3. Output integrity: All vendors are very skilled at talking about how their systems help mitigate hallucinations from large language models (LLMs), Lee said. However they should also offer retrieval augmentation generation (RAG) techniques and paths in order to make sure that the content or inputs are always brought from the most relevant and most correct sources. “If the solution provider can show you how to help you manage bias detection within the system, then you have the beginnings of a great CX to be delivered through that solution provider,” Lee said.

The potential for AI agent sprawl, lack of standardization and security risks remain the biggest impediments to conversational AI in CX, Lee said. It also pays to recognize how many cooks you’ll have in the kitchen.

“When you look at identity and access management controls, conversational platforms are not just used by one person from the CX organization,” he said. “They’re used by a variety of different folks who need different levels of access as well. How easy is that to set up? That’s part of the internal security guardrails.”

Lee’s presentation and all other sessions from Engageware’s event will be available on-demand for registered attendees.