By Ric Kosiba, Sharpen Technologies
I think it is inevitable that the older you get, the more curmudgeony you get. It certainly is true of me.
I spent last week at a fantastic conference for contact center salespeople, where representatives from 16 contact center vendors all got to pitch their companies and products in front of potential sales partners. It was really cool to hear how each company tried to differentiate itself. I find sales pitches fun!
What made this so interesting was that a good percentage of the pitches had obviously been written by their marketing departments, were read verbatim, used a bunch of jargon, and were silky smooth. And it mentioned AI a lot, but in a vague way.
Later, when we had breakout sessions, we were able to chat with the individual sales professionals. So I asked them: “What do you think contact center AI is?” The answer was universal: it’s mostly chatbots/IVR.
The dirty secret is that, for the most part, chatbots are pretty similar, vendor to vendor. We all outsource voice transcription to one of a few companies, and the main differentiators are the tools used to allow customers to traverse the IVR/chatbots decision tree or to pull phrases from the transcription results. And there, the differentiators are fairly small.
What other AI is left? There are purpose-built, large effort AI systems developed to solve very specific problems, like routing calls to the most “compatible” agent, or auto QA, or contact sentiment, or aids to after-call work. All valuable if done right, but they are bigger and more expensive projects for sure. Very few of these AI projects come right out of the box ready to go.
With AI being such a huge marketing point for many contact center vendors, there must be a next big AI thing, right?
Technology as a Substitute for Decision-Making
Technology can be a great substitute for human decision-making when the business problem is well-defined, and the solution is combinatoric in nature (there are many, many solutions to consider). Agent scheduling is a perfect example of this – our problem is to find the best set of schedules given a series of real-world business constraints (agent preferences, 40-hour work week, break and lunch rules, day-off patterns, etc…). The number of valid schedule combinations is huge. The computer’s job is to sift through the huge number of possible shifts to find the set of schedules that most reduces our costs while maintaining all these constraints as well as possible. The computer is good at sifting through these combinations, using a bit of algorithm magic.
But technology can be a terrible substitute for human decision-making, when the conditions are poor for it to work: if the problem is not well defined, if the solution-space is infinite, or if data cannot drive a correct answer. In our world, automatic forecasting sometimes falls into this case. When history is no longer a guide because of changes in our business environment, say that the economic recession or the pandemic changes our customer’s behaviors, then no algorithm can get a forecast “right” because history is “off.” (Note… see older articles in
On Target about using capacity planning to manage risk!)
So, we may have a bit of a problem with AI in contact centers today. Sure, all the chatbots are fine, but the bigger AI efforts will still be large and costly, and potentially, projects that depend on streams of historical data may be problematic.
Machine Learning is Hot, But Expert Systems are Downright Sexy
One of the earliest types of AI is one that we don’t hear much of anymore is the expert system. Years ago, this was the most practical AI model in development. It involves finding an expert, or a team of experts, eliciting how they solved some important problem, and then coding up their decision-making process into a computer program. I expect AI like WebMD is an expert system. In my youth, I worked on an expert system to determine if a satellite was broken — I found an expert and coded his rules around how a satellite fails. Worked great.
The thing is, different sorts of problems are best solved using different sorts of “AI” or other math technologies. Each problem likely uses different techniques. While machine learning approaches are great for honing in on solutions given a stream of refreshed data, expert systems are great for solving complex problems that have acknowledged specialists and less data.
This is fun — often common sense or leaning on your “engineering judgment” (a real term in engineering circles) is a great way to solve the sorts of problems that don’t have a ton of data or the ability to do an A/B test.
Better yet, in your contact center, there are bound to be experts who know both your products and your customers, and who have experience listening to and critiquing calls.
Introducing Micro-Bots with Expert Systems
What if we changed the current AI paradigm, where we hire a bunch of expensive data scientists and spend months testing and tweaking a machine learning algorithm? And instead just set up a new process flow, based on our best practices and expert ideas, and test to see that it works.
Welcome to the world of micro-bots, where we solve a smaller but significant problem with automation, common sense, and your expertise. Here, we follow a model with three simple steps:
- A trigger to fire off a bot
- A model to determine a course of action
- The action (what the bot does)
Let’s discuss each of these in turn, and then we can all use our imagination and our engineering judgment to conjure up our own bots (read through to the end to see how you can win a prize for your bot ideas!).
In contact center micro-bots, we can set up bot triggers a number of ways. The obvious is to trigger a bot with a change in state, meaning when an event within the contact stream happens, we can turn on a bot. For example, we can turn on a bot whenever a contact enters a queue, or is put on hold, or is transferred, or the contact is ended. Or any other change of condition.
Bots can be triggered at a specific time, at 8AM or, say, on every half-hour.
Bots can be triggered when a condition is met, for example when a customer is on hold for too long, or when an agent has been in a pause state for greater than some threshold. Maybe a bot can be triggered when a customer has been transferred twice.
Models are typically Boolean expressions using data within contact center or CRM databases. The simplest may be something like:
If (the customer is a member of ‘gold status’) OR (if the customer is a member of ‘silver status’ AND the number of calls in 24 hours is greater than 1) then… (insert action you want your bot to take).
But here is where it can get fun. Data can be combined, small bits of code can be written to develop new metrics, or models can be built to populate tables in the database. So external models can serve as a decision element, say a “likely to buy” model, or a “likely to attrit” model, by simply adding its results to the database.
In contact center systems, the ability for a micro-bot to act can take a whole bunch of forms.
- Contacts can be re-routed or prioritized (or de-prioritized)
- Different prompts or messages can be played to the customer
- The customer can be added to a list or the contact can be flagged
- Agents can be shown information and prompts
- Supervisors and management can be notified (say, emailed or slacked)
- The customer can be added to an agentless outbound campaign
- Pretty much anything!
This ability to act makes micro-bots very powerful!
Figure 1: Micro-bot process flow
Likely Upset Customer Bot: Let’s say that we have done some analytics work with your contact center data that showed CSAT is highly driven by Active Contact Resolution (ACR), as you’ve seen in recent editions of On Target. How do we create a bot to help improve the experience when customers have to call us back? Here’s a simple bot:
Trigger: Call enters queue
Model: If a customer has called us before within the last 24 hours OR if they have given us a low CSAT score this month OR if they have spent more than 45 minutes of total handle time in the past week OR if they are on the retention risk list
Action: Play a message telling the customer that we appreciate their business, we’ve noticed they have called recently, and that they will go to the front of the specialist queue. AND reroute the contact to a Tier 2 agent at high priority. AND alert the agent that the customer is at risk of churn. AND flag the customer as a retention risk and place them on a surprise and delight list. AND add the customer’s previous calls to a QA-to-analyze list.
Would the customer be better handled given the bot? Our intuition says yes, but it certainly is a testable proposition. Since setting up this bot is so quick and easy, we can set it up, gather statistics on customers who have met the conditions of the model, and look at differences in CSAT.
Intraday Management Bot: Here is a fun one. How can you alert supervisors to the important stuff that is happening within their team? Certainly, monitoring their dashboards and reports is helpful, but what if you could build management ideas into a series of bots?
Trigger: Hourly, starting at noon
Model: (When the busyness of an agent (personal occupancy, see On Target) is greater than 10% of the group’s average AND the agent has worked more than 5 hours today) OR (the agent has greater than 95% adherence for more than 26 hours while working AND the agent has not received kudos in the past 2 weeks)
Action: Send a note to management and the supervisor recommending a compliment. For example:
- “Lucas’s busyness has been over xx% for more than x hours. It will help him avoid burnout if you can afford to give him a quick, extra 10-minute break.” (Or let him go home early, etc.)
- “Lucas has been #1 in [Performance Management Stat] has for x days. Treat him to a $5 Starbucks gift card?”
- “Lucas’s personal CSAT scores is x% higher/lower than the average among other agents in [queue name], make a big deal of it.”
See how these can be both fun and helpful?
Ideas, Tests, and Gift Cards
What has made developing micro-bots so interesting is that the pain associated with building them is so low, and the capabilities of this sorts of automation is so high that there are few barriers to testing them. We can all really roll out a new one in less than an hour. And we get to be our company’s expert and our bots reflect our great ideas.
So here is my challenge to you all. Please think of some cool bots you would like to build for your contact center. Email me any cool, fun, or funny ideas for a bot, and I’ll send a gift card to the first 10 great ideas that hit my inbox.
Ric Kosiba is a charter member of SWPP and is the Chief Data Scientist at SharpenCX. He can be reached at firstname.lastname@example.org or (410) 562-1217. SharpenCX provides unified contact center software that empowers agents to deliver engaging customer experiences with an all-in-one, customizable platform.