Ok, so I won’t bore you with the back and forth details that led up to this. The important thing to know is that I needed to talk to a human at the IRS to ask a very simple question that would take just a few minutes to look into.
Me trying to call for 3 days dozens of calls dealing with this shit. I eventually figured out how to prompt the AI to let me push buttons to reach menus.
After all that I plotted out what each option does.
ALL OF THEM either lead to a pre-recorded message that says to visit the website, OR attempts to transfer you, before telling you there are no agents taking calls for that option.
Eventually, I wondered what would happen if I said to the AI agent “TAX FRAUD”.
Boom. Straight through to an agent.
I feel bad for lying to the hotline, but also, I was left no valid options to talk to a human. Once I said tax fraud, my issue which I had spent probably 100 calls in 3 days trying to get through, was over in 10 minutes. And 6 of those minutes she put me back on hold to research my issue.
Fuck AI
You use the LLM to navigate the phone tree and allow the user to skip levels and menus
A contrived example:
At a top level, Thing might branch into subThing1, subThing2, subThing3.
A user might say subThing3 or a similarly worded item and bypass extraneous layers of IVR.
It lets you skip levels and not have to be as precise beyond what an IVR implementation might miss.
“Customer service”, “agent”, “representative”, etc, etc. are all equivalent, but one might be missed.
LLMs do help with this circumstance
That’s still deterministic, what you are describing has been around far, far longer than llm technology.