Making Conversational AI feel more… conversational
Over the last year, Virtual Agents and ChatBots have exploded in adoption driven by the pressures and cost of re-staffing call centers post-COVID. Most of them are horrible, and I’m not alone in that sentiment. According to a recent study only 40% of people were willing to use a Virtual Agent to retrieve account information — and that was the high water mark across industries and use cases.
From a corporate standpoint Virtual Agents make business sense. You can drive long-term cost reduction and higher customer satisfaction with the immediacy of a bot handing first-line customer interactions. The secret sauce, however, comes in the implementation. Spending the time (and using the right tooling to reduce that time) can make all the difference in driving customer adoption, lowering near-term cost, and staying on-plan.
I spend a lot of my time leading teams building virtual agents for companies across industries using IBM’s Watson technologies. The one nearly universal ask I get is both straightforward and intensely complex: “We want a Virtual Agent that can answer almost anything well, but don’t want to have to build and maintain every possible answer.” Costs can quickly spiral out of control on both the build and maintenance of a Virtual Agent if the number of directed conversations gets too large.
At Cerebral Blue we’ve recently begun delivering an architecture I built to meet this demand — generating on-demand, topical, relevant answers with a healthy amount of repetitive variety to provide a backing plane for a small set of curated conversations.
In Watson Assistant, instead of the default Watson Discovery search integration we employ a custom search “extension”. In our architecture, Watson Discovery becomes a “context” engine — providing a curated topic context back for a given natural language query. We populate the discovery collection with properly scrubbed and indexed informational articles of 1–2K words each. At run-time, during a customer interaction, the Discovery context is then used to provide just-in-time training to a set of AI algorithms that generate a natural-language response.
All customer queries that don’t match a pre-defined path in the Virtual Agent then flow to our search extension.
This integration yields very good responses for unknown questions, and should go a long way to driving customer confidence in the virtual agent — the key to fostering a positive experience and increased willingness to utilize the assistant in future interactions.
As a demonstration, here is an assistant with only our search extension— no preprogrammed dialog or answers are present…
Reach out to us at https://cerebralblue.com to see how we can bring this innovation to your company.
