Can you legislate AI?
The CallMiner Research Lab weighs in on the proposal recently released by the European Union on how to regulate artificial intelligence.
After my internet went down for the umpteenth time last week, prompting yet another call to the call center, again I heard, “I’m so sorry to hear you are having internet trouble.” I was at the end of my rope, pretty much since I had signed up for this promised super-fast internet it went down 4-5 times a day for weeks. I began replying, “I’m sorry too.”
Working in software I know that the only way to solve this issue is to keep fixing each possible bug until it works. In addition, having spent a good part of my career in call centers, I realize some agents, especially those in out-sourced environments, are required to use empathy in calls. Even with all this background, I just wanted the internet to work.
Hearing my hundredth sorry was not getting me closer. While sitting on hold, forced into listening to looping ads about how reliable this service is, I began to ponder, how sorry are they?
Sorry is proffered innumerable times a day in call centers all over the earth. Somewhere right now an agent on the front lines is plying a ‘sorry’ to apologize for an issue. Every year, millions of ‘sorries’ are offered as a salve for some imperfection in a product or service. But how sorry are they really? In a business environment does “sorry” even matter? Is it worth the effort to even say?
Sorry is prolific in general conversation. So ubiquitous that governments have passed Apology Legislation to curb the impact of apologies such as sorry as being legally binding. Five Canadian provinces and 39 states in America have “Apology Legislation”. This is an effort to limit liability if someone says they are sorry for something.
It seems legally we have a strong vote that saying you are sorry does not actually mean you in fact are. This is especially problematic in health care, where a provider may be truly sorry for the condition of a patient, but not want to be sued for liability for an expression of that empathy. So legally, even if you really are sorry, you can’t be held liable for such sentiment.
‘Sorries’ are cultural too. Linda Geddes, in her article “Why do the British say Sorry so much,” offers “the average Brit says ‘sorry’ around eight times per day” and that “one recent poll found that there would be 15 British ‘sorries’ for every 10 American ones.” Sorry is said a lot in English, so much that I wonder if the word really carries real empathy in general conversation?
The quick hit English tutorial website “Espresso English” offers a lesson in twenty ways to say you are sorry, ranging from a simple use such as “I did not hear you” to the heart felt “death of a friend or relative”. There are many ways to be sorry. A good way to think of the range of apology in a call center is personal empathy for a life situation, and service or product failure empathy of an issue. This range presents a problem to an analyst like me – sorry as data is totally open to interpretation.
If I tell you “sorry” is said across our research clients 2.5 million times a month, it is easy to interpret that as a vapid use of the word. But if the transcript says, “I’m so very sorry to hear your spouse has passed away” the interpretation of the 2.5 million may change. “Sorry” data, like all conversational data, is complex.
At the CallMiner Research Lab we work in a world of communication. Specifically, using AI to understand and improve the conversations between a company and its customers. Spoken and written, we have access to billions of conversations and over a trillion words a year. We know words, who said ‘em, wrote ‘em, how they were said, or used, and how they were responded to. With this backdrop here is what our data tells us about sorry.
Agents say they are sorry more than customers, the ratio for Agent to Customer sorry use is about 1.75:1
The most common uses of sorry in call centers are, service or product failure empathy using a simple sorry in front of the issue such as, ‘I’m sorry’, ‘Oh, sorry’ ‘I’m sorry to hear that’, ‘I’m so sorry’ and ‘Okay, sorry’. This apologizing for a service or product failure is about 10X more frequent than the next use.
Next most common is trouble hearing or understanding with phrases like, ‘I’m sorry what was that’, ‘Sorry say that again’.
Machines bring up a distant number three, with automated messages doing their best to throw some artificial empathy at us, ‘I’m sorry the person you`re trying to reach is unavailable at this …’ I bet we can all finish that one in unison.
An average client in CallMiner Research Lab uses sorry about 35,000 times a month, but that is not fair as client’s range in size. So, let’s normalize a bit.
Frequency of use is a good way to even the playing field. It gets interesting when comparing how frequently sorry is said in customer interactions regardless organizational size. The most frequent use of sorry in our database is a law firm, I didn’t expect that, but that is why research is so cool. Here is the top 10 list ranked by frequency:
What does this tell us? Well, it’s sort of scattershot at the top, but BPOs have made empathy into a business practice. If you work in the industry, you quickly learn that outsourced labor is managed very closely and is often rigorously scripted. Empathy is a product feature and BPOs use it a lot.
Looking at the other end of the spectrum, three of the five lowest frequency users of sorry are Vacation and Travel companies. This in no way means they are crass or not sorry, but it seems they are not using the word with nearly the frequency of a BPO. I have heard their calls, looked at their interactions, these are some very good and empathetic companies. When you have an issue with travel and end up calling the provider it’s not usually good. Yet all three of these are lowest frequency use of sorry.
I was interested about this, so I spoke to one of them regarding the use of the word. It was confirmed they absolutely use empathy in response to dissatisfaction but combine it with strong ownership. If agents say they are sorry, they need to own the solution personally to ensure the customers issue is resolved.
Using a machine learning algorithm, we can look at how sorry is used in a set of contacts by context. Without going into too much detail, we have models that will group or cluster words and phrases by the context in which there were used. Dog, Cat, Parakeet, Parrot, “Pot belly pig”, are all “pets”. It’s more complicated than that, but I’m sure the idea is clear.
Let’s look at the use of sorry in the most frequent client, the law firm. It is used a lot, but in what context? There are two main uses, first personal empathy for difficult situations faced by clients, and there is an entire cluster of them.
"hear about your loss"|”very sorry for your loss"|"terribly sorry to hear"|"certainly glad"|"certainly certainly"|"hear about his passing"|"truly am"
This makes a lot of sense, it’s a law firm dealing in litigation around life’s situations; they should use personal empathy, and lots of it. The second context sorry is used in is apologizing for an interruption on the call such as, “i’m sorry to interrupt".
The eCommerce Company is also interesting, they have three primary contextual uses of sorry, first is delivery time frames,
"we're sorry to keep you waiting"|"keep you updated"|"keep me in the loop"|"keep your eye"|"keep you posted"|"keep you waiting"|"keep you in the loop"
Second a contextual cluster of difficulty hearing,
"sorry couldn’t hear"|"barely hear"|"hear me i cannot hear"|"cannot hear"|"hardly hear" And a large cluster of recorded apologies, like a voice message or voice recorder apologizing for messages and wait times. "we’re sorry your call cannot be completed"|"holding your call is important to us please remain".
How about all those BPOs? Almost all the clusters we sampled are apologizing for hold times, "patiently waiting sorry for the long hold" or service/product delays, "some delays"|"slight delay"|"waiting sorry for the slight delay”. Very logical.
Knowledge management is something that BPOs must deal with. Many programs are complex, and the accuracy requirements are rigorous. As such, agents spend time in knowledge management systems looking for correct answers or asking their supervisors or peers. That typically equates to time on hold. Every BPO I know will require an apology if hold is longer than a specified time, and at the end of any hold time no matter the length.
Service/Product delay is logical as well. As the arrangement is a third party, they are privy to only the information that the first party shares, and escalations are frowned upon. Apologies can tend to deescalate a call. All part of the complexity of being a great BPO.
Finally let’s look at the lower use Vacation/Travel Companies. Regardless of the client when we use this algorithm, the machine is trying to find the context around the use of every word and common short phrases. But the machine has no idea of the actual context the way we humans do. Machines don’t know what pets are from our example above, just that they are used in the same way. This unsupervised approach works well, it clusters similar concepts very well, yet at times a word is used so much or so infrequently that the machine can’t figure out the context.
Stop words like “is” or “and’, “I” or “we” are perfect examples. The machine puts them in what we lovingly call junk clusters. A random smattering of overused and underused words and phrases that the machine can’t distinguish but believes are contextually the same.
In our Vacation/Travel Client it turns out sorry gets put in junk clusters a lot. Here are some examples.
Sorry being confused with a names and other random terms in a junk cluster:
kelvin|mavis|penelope|chrysalis|jabari|zoe|janus|sica|dolby|freddie |issa|gladiator|noggin|shah|rhoda|roddick|florrie|lupita|kb|cissy|ravi |jamil|devlin|peregrine|hydra|franny|hegel|kimmel|ramone|moby|nat|"ed ed"|"- kay"| "david david"|"sorry if i mispronounced"|"joe joe"|"matt matt"|"julie julie"|"hey sis"|"- #"|"tony tony"|"baby doll"|"ray ray"|”katie katie"
Sorry also appears in a total junk cluster of words rarely used by the client:
dorky|salter|snappy|frail|washy|flicks|chatty|wakeup|itchy|robots |doolittle|keener|dyslexia|sabbath| quito|momma|mozilla|grouchy|achy|fasten bearer|thicker|rune|lofty|bashful|lite|gaseous|"sorry i misunderstood"|"small letters"|"hey bud"|"tai chi"|"hail mary"|"young woman"
This tells us that sorry is infrequent, but what are they using instead? “Apology” it seems, note there is a sorry in here as well.
apologized|"sorry for the inconvenience"|"apologize for the confusion"|"understand your frustration"|"apologize for any inconvenience"|"sincerely apologize"|"apologize for that delay"|"apologize about the inconvenience"|"truly apologize"
So, they do apologize but they use a more formal way to do so. Just as my contact confirmed, it is just coached to be used in a way that they hope is heard, meant, and felt.
Let’s loop back to the question, should companies use blanket empathy, such as sorry? What are the options? Don’t use it at all, use it however you want, and use it when you mean it.
The best option would be the last obviously, but how can an organization with more global agents on the phone than the population of Ft. Lauderdale ever coach that? When I worked in collections call centers the absolute best debt collectors were the one who cared about the debtor at the human level. They really cared and felt responsible to help a consumer find resolution, not just collect debt. Maybe from all we have seen, actually caring is the key.
I recommend coaching agents to use empathy that matters, ideally that they believe, but at least is topical to the level of dissatisfaction at hand. That does not mean every other word is sorry, but a combination of empathy and strong ownership goes a long way. Strong ownership builds trust and decreases the frequency empty empathy needs to be used.
Using our software as these companies do allows them to not only detect the dissatisfaction on the customers side, but the empathy and ownership on the agent’s side. If the agent doesn’t do it correctly, a helping hand can quickly be extended. As an AI person, I do love a machine in the loop helping the humans.
Let me offer a suggestion – if you say you are sorry, own it. If you work in a large internet provider whose service I pay for, it might sound something like “I’m so sorry your internet is not working again, I see it has been weeks, I want you to know we will take care of this while we are on the phone today, so you don’t have to deal with this again.”
That is some empathy I could believe in.