An insurer's empathy model marks a claims rep down for how she handles a bot

The scoring model is the part of this already producing wrong numbers. The quality AI that grades the California claims rep on "empathy" marks her down for not being accommodating enough to the bot ^[10]. She can usually get the rating overturned, and she gets it overturned by asking ^[11]. That appeal happens one call at a time, and it runs through a supervisor.

Two calls at about five minutes is roughly ten minutes of one rep's shift, spent on a single law firm's bot ^[20]. "If it's still talking back, we have to stay on until it disconnects," she said ^[7]. At today's volumes that cost comes out of a shift, ten minutes at a time: Business Insider reports machine callers are still a small share of the total, and does not report the share ^[3]. Scoring and authentication are rules, they apply to every call, and a rep cannot change either from her desk.

A second worker, in customer retention at a major credit card company, said it is relatively rare for her to end up on the phone with an AI ^[12]. Business Insider lists the questions the industry has not settled: how to authenticate an AI caller, whether to engage it at all, how to tell an authorized agent from a scammer, and whether a rep is even allowed to ask if the voice is AI ^[4]. At her employer that last question is settled, and reps may not ask ^[15]. She can tell anyway, she said, because the voices "don't have a lot of voice in their voice, and you can hear there's almost a hesitation when you answer a question that a normal human customer wouldn't have" ^[16].

One control holds. Only the account holder can close a credit card. "There are certain functions that only the customer can do, so they have to be the one on the line," the retention agent said ^[17].

The volume is still small ^[3]. One machine caller at the credit card company spent its time working out how to reach a person: "I had one just recently where it was like, 'If I had to get to a representative like you, how would I do that?'" she said ^[13]. "What I think they're trying to do is map out the system," she said ^[14]. Neither of the bots these two women described appeared to be doing anything nefarious, according to Business Insider ^[21].

TTEC put the question to its clients, and the general answer was to take the call but proceed with caution ^[19]. That leaves the judgement with the person holding the line. Engage every caller and you spend handle time on principals you cannot verify. Screen for machines and you may be turning away a customer's authorized delegate. The two decisions actually available now are whether the empathy model may grade a call the company believes was machine-placed, and whether a rep may ask the caller what it is; both take effect on every call that arrives after they are made.