After becoming increasingly popular in 2015 and 2016, this year has truly become the year of the bot. Countless businesses have embraced bot technologies, built applications supported by bots, deployed bots for varying purposes.
There are bots “doing” marketing, sales, customer service, bots that connect employees to each other, bots that take up tasks for employees, even bots that cooperate with other bots. Look, for example at ordering bots like the one of pizza hut, marketing bots like GrowthBot, or the variety of chatbots that send personalized messages to customers and prospects via Facebook’s messengers, Slack, or other messaging platforms; or have a look at the infrastructure that agent.ai has built for customer service: bots collaborating with customers, employees, and themselves, governed by a bot, to deliver customer service.
Still, in many cases businesses and organizations have not seen a good return.
And this in spite of bots having been touted as a panacea, a technology that resolves most, if not all business problems, that ultimately makes many a human redundant…
What is a Bot?
A bot is an application that consists of
- Supported by a learning AI backend
- That emulates intelligent behavior and is
- Usually exposed by a conversational user interface
Apart from the conversational user interface this cannot necessarily be identified from the users’ side.
So, what went wrong?
Apart from the identification problem and the fact that some bots actually are not bots, there are problems on at least two dimensions. I would call these dimensions the technical and the “human” one.
On the technical side we see that:
- Bot systems are often not integrated with business systems, which leads to process breaks and inconsistent databases
- Functional scopes are too big; this has the result that delivery to expectations is sub par.
- The necessary foundation not yet being set. A well working bot needs a working AI and a (machine) learning infrastructure that is adequate for the desired bot scope. Many bots that I have seen also have challenges parsing longer sentences. Try some of the pizza bots …
- So far bots cannot really show empathy and have a challenge dealing with context. Try using irony or slang.
Then there are challenges on the “Human” side, which include that:
- There is no coherent plan for integrating bots into the business processes. This includes too broad scope that shall be delivered by the bots as well as missing escalation strategies. What needs to happen if a bot gets “out of its depth”?
- There are high expectations on the customer side. Customers as well as employees expect that their issues are served as frictionless as possible. This includes, particularly for conversational interfaces, that the bot understands them. The process needs to feel natural.
- Some bot implementations lack transparency. How often do we know whether there is a human or a machine behind the chat window that pops up on a web site? This relates to the expectations topic. If the system mimics a human then the user on the other side expects a human.
Vendors are selling more than they can deliver. There is still a gold rush going on, and an important ingredient for successful bots, a well working AI to support it, is not yet up to the mark.
Businesses and organizations engaging into full-blown big bang implementations, without giving enough thought about how to deliver a good (enough) experience to customers and employees.
The combination leads to implementation failures.
What it Takes
A winning strategy addresses above challenges.
Humans want to interact with systems as convenient as possible. Part of this convenience is that the interaction is often taking the form of a conversation – often, not always. But it is still more than often enough to mean that bots are here to stay. Will they fully replace a “graphical” user interface? I do not think so, but they will augment it.
As a consequence executives need to plan for the implementation of bots, and carefully so. A necessary approach is to think big, while acting small.
This includes to set a reasonably limited, yet important enough, starting scope and to define a process around it. This process needs to accommodate for an initial training and continuous improvement of the bot(s) and may involve having the bot (or bots) running as sidekicks to human operators until their ‘decisions’ are good enough, meaning they are consistently above a predefined threshold of right vs. wrong.
It also requires an escalation process from the bot to a human in case of the bot not being able to give the correct reply. From here on the scope can get incrementally increased, then covering more and more use cases. It goes without saying that the decision processes need to be data driven, and following the organization’s priorities.
There also is the need to keep the end user informed about whether they are conversing with a bot or a human. This helps a lot when it comes to expectation management and therefore the users’ perception of the interaction, aka experience.
And, for the employees it is important to know that bots are not implemented to replace them, but to help them do more of the interesting work. As Vinnie Mirchandani put it: The bots are there to do the dull, dirty, and dangerous work. They are performing tasks, and are not filling positions.
These measures help building trust.
It is equally important to choose an AI and bot platform that integrates well with the main business systems and the existing web platform. Ideally the bot platform already offers a growing number of pre-trained models that cover the chosen use cases and that serve two purposes:
- The time to production is decreased
- It is not necessary to have too much in-house data science skills
The vendor’s viability is also a concern, given the market still being early stage.
In summary, to be successful with a bot implementation, executives are wise to consider:
- Have a strategy
- Keep the employees working with the bots “in the boat”
- Have a bot process defined that has a meaningful escalation process from bot to human
- Start with a limited, tightly controlled scope, and iterate from there