Apple recently announced its pause on AI notification summaries for news and entertainment.
This followed a wave of backlash for the generation of inaccurate news alerts.
The next IOS update will disable the notification completely until a future update after refining the service.
Apple’s backtrack comes as hallucinations continue to sidetrack generative – and now agentic – AI projects.
Consider Amazon’s recent struggles to rebrand Alexa as an AI Agent. It cited hallucinations as a continued obstacle.
Yet, despite widespread talk of hallucinations, many contact centers pressed ahead and implemented auto-summarizations.
After all, contact center providers promised that this would be a simple first use case for service teams to help build their confidence in GenAI.
Now, those auto-summarizations don’t seem quite so easy.
Auto-Summarization Adoption Is High in Contact Centers, But It Has Difficulties
Keen to jump on the heels of the AI rollout, many contact centers have jumped on the bandwagon of case auto-summarization.
The last 18 months have seen a huge uptick in service providers implementing auto-summarizations.
Indeed, according to a recent CX Today report, 38 percent of contact centers have already done so.
Loading these auto-summarizations into the CRM post-contact has proven helpful in tracking customers’ case history, lowering handling times, and saving costs.
Yet, some contact centers have found that – while models have performed well in POCs and pilots – scaling them to enterprise-grade production has its difficulties.
That’s according to Swapnil Jain, Co-Founder & CEO at Observe.AI.
In a LinkedIn post, Jain shared the following list of what it takes for enterprise contact centers to make their auto-summarizations effective:
- Fine-tuning models for accuracy
- Resolving transcription errors
- Enforcing length constraints
- Achieving high-accuracy entity extraction
- Managing latency requirements
- Maintaining voice consistency (first-person vs. third-person)
- Handling complex scenarios like call transfers seamlessly
“These aren’t minor details,” noted Jain. “They’re essential for building true enterprise-grade AI solutions.”
Should Contact Centers Bail on Auto-Summarizations?
Before bashing auto-summarizations completely, it’s critical to remember the time before they were a possibility.
Contact center agents had to manually write up a summary, tag the interaction with a disposition code, and send it off to the CRM.
In a rush, agents would often miss key details and select the wrong disposition code.
As such, they filled the CRM with inaccurate data, which meant that contact centers struggled to track the history of many customer cases. Moreover, they couldn’t isolate why customers were calling accurately.
While auto-summarizations may sometimes hallucinate, they are a step ahead of what came before.
Yet, contact centers should strive to make sure their deployments are as accurate as possible.
That starts with a healthy dose of skepticism. As one commentator on Jain’s post wrote:
A good AI demo is 10 hours work but true production is 10000 hours of dirty work
Also, contact centers should target auto-summarization solutions that let them customize the back-end large language model (LLM) and the prompt that feeds it.
From there, run extensive tests on those models and prompts in a sandbox environment, with tools to benchmark the results.
One example of a solution that enables all this is the Five9 GenAI Studio. Discover more about that solution here.