Challenges for support leaders looking to adopt AI in an economic downturn
Everything changed in Q1 of 2020. Our lives, our work won’t be the same at least in the short term. And this will impact every single department of an organization. What is happening in support?
Zendesk recently published a report about the state of Support in March during the COVID-19 pandemic growth. We believe the main tendency in the report will be accentuated in the next months/years as people and companies make relevant changes in their habits:
“Customers are waiting longer on replies across affected sectors”
However, at the same time, in a downturn, corporations’ profits will shrink and it will be a challenge to maintain a balance between customer support quality and profits. On one hand, companies that get it right can be the industry winners after the crisis. On the other, companies that don’t, may not recover from this situation and as we can see in this Forbes article many are laying off customer support teams when their customers need it the most.
Businesses that are now forced to transition into digital transformation and remote work, will be pushed into automation and it will be one of the trends that will accelerate.
With costs-saving and efficiency in mind of firms but tickets and backlogs growing at the same time, fast implementation of human-friendly products will be the key in automation systems for support, so that teams can quickly react and adapt.
Fast implementation of human friendly products will be the key in automation systems for support.
Given that automation will play a critical role in business transformation, how do companies need to think about automation in systems that deal with customer communications? Are AI systems for support ready for this new paradigm?
What we’ll discuss in this article:
- Current AI/automation systems are costly and lack flexibility
- Support organizations face huge challenges to quickly adapt to automation
- AI/automation systems are very limited for support operations
- AI now can be like Excel 35 years ago
- We are about to see a change of paradigm. AI will be exploited by support leaders without the need for engineers
- As a support person, you can and should adapt to this new automation world
Lack of flexibility in current AI systems
Nowadays, most AI systems rely on humans training them. How do they work? You need to tell the system what to do by showing it that an apple is an apple and a pear is a pear, and that way it will learn. This process is called labeling (or tagging), and you need tagged data to train a supervised algorithm. It is a costly and time-consuming process.
A typical example is training an algorithm to classify photos from a cat or a dog. You need a lot of photos that have been previously labeled for the supervised algorithm to learn and that’s why it’s called supervised.
However, you don’t know why the technology is taking certain decisions because behind that labeling, there are sophisticated systems (that every day are getting more complex with deep learning). They are basically a black-box. If tomorrow, you want to identify a certain breed, you cannot just easily change the pattern, you would have to tag again thousands of photos of the specific breeds.
This is the complexity self-driving car systems are also facing. There are multiple complex situations while driving a car and while a human can react to similar situations, current AI systems can’t. As a human, we may have never seen an electric moped before but we can guess its speed and behavior based on other similar situations we have seen (like an electric bike).
Lack of flexibility applied to AI in text problems
In terms of language, the typical example we use in every day is SPAM. After years and thousands of people training what’s SPAM and what’s not SPAM, spam filters nowadays work with high accuracy. You can personalize them to you if you keep tagging enough data so that the system learns what’s SPAM specifically for you.
SPAM is a general problem, we all use email, and we could argue there is a broad concept of SPAM that is objectively true. There are two problems with this:
- In most language situations, especially business situations, that is not the case. For instance, take sentiment; what’s positive to me may not be positive to you; it depends on the criteria we have, which are not objective.
- Imagine you wanted to separate in different types of SPAM. You can’t just easily review the results and separate those types. The technology is a black-box behind the scenes. It just does what’s meant to do: separate SPAM from not SPAM, and if you want it to do something else, you have to tag again more data and re-train.
The challenge for customer support organizations
This problem is even more acute for specific business departments like customer service. Customer service teams usually organize their client’s tickets in different ways, and it’s the VP, director or ops manager that decides what’s more efficient for their specific business process and the nature of their business.
For instance, let’s compare a payment product (like PayPal) vs. a product that has payment functionality (like Netflix). It would be challenging to have a global system that applies to how they tag their tickets because different topics have different importance.
- In Paypal, it doesn’t matter if it’s a payment related problem, because probably 90% of the issues will be. Therefore, the tags will have to be more nuanced, for example, “international payment” or “lack of funds.”
- In Netflix, a low percentage of problems will be about payments. Hence they can work with something less nuanced. The business processes affected will be probably related to “renewal payment”.
And the list could be infinite! Different companies structure their payment processes differently and, consequently, want to tag things in different ways.
Their business leaders are the ones that know what’s the best way to optimize their processes.
Limitations of current systems for customer support
However, in existing AI systems for text, business users are constrained by technologists. They do not have a way to change a process on their own, they rely on engineers or complex AI systems. This is the second biggest frustration of support leaders when dealing with AI software.
In existing AI systems for text, business users are constrained by IT help.
The reason is that the only options they have are:
A predefined system with “universal knowledge”. For example, a system that applies a rule every time a ticket has a keyword.
A system that you train, but you don’t know how it works, it’s not transparent, and it’s not flexible. The limitations of this approach are:
- Lack of transparency: If you want to understand why it’s making a decision, you can’t! Your only solution is tagging more data to make it better which is time-consuming and costly.
- Lack of flexibility: If you want to modify something, you can’t! Your only solution is tagging data. Let’s talk about a real example in a tech company:
- You’ve set up a process for payment related issues to be redirected to a help center article.
- After a few months, you implement Stripe and you realize you need a specific help center article only for Stripe payment issues.
- You would need to tag more data to build this new system with two separate tags.
- This situation can keep happening as long as you need to adjust your business processes, which will keep happening in most companies!
Current AI systems for text have been designed to solve problems that are standard for everyone, but that’s not what happens in the reality of business.
Current AI systems for text have been designed to solve problems that are standard for everyone, but that’s not what happens in the reality of business. This has resulted in teams that end up with “armies of taggers” that are constantly re-training AI systems to adapt them to their reality.
Can it be more efficient?
Comparison with a different industry- Excel
Let’s compare it against other industries like financial operations. In the beginning, only financial experts could operate budgets and change something that the rest of the team would see. It was a black-box for other people in the organization.
With the growth of Excel and by making excel a simple interface that anyone can use, anyone can now access and change that formula. Every department in an organization now uses spreadsheets without the need for a financial person, and the business owner adjusts a spreadsheet to their business need.
Could we see a similar change in AI systems? A change that would allow anyone to benefit from AI without the need of an engineer? For sure.
It will be a combination of AI and UX with an easy way for business users to operate. The cost and speed implications are huge.
Same as with Excel, it will be a combination of AI and UX with an easy way for business users to operate. And the cost and speed implications are huge.
A flexible system for support operations
We designed Lang.ai since its inception as a system made for support managers and support ops managers to design their automation processes so that they are the ones in control and not IT teams. We believe that is the step needed for AI to become mainstream in this new automation era.
We don’t believe the future is for a few engineers to control AI systems so that their decisions impact the rest of the world, we believe the future is when these AI systems are flexible enough so that any person can use them, and adapt them to bring efficiency in their daily operation. It has been done in other industries like web design.
We don’t believe the future is for a few engineers to control AI systems so that their decisions impact the rest of the world.
How do we apply this concept to customer support?
Instead of business users having to tag every single topic they need, Lang.ai suggests tags in a system that has four key components:
- Adaptability. It adapts to the client’s data and criteria (we don’t believe language interpretation is universal)
- Constant evolution. It’s dynamic and evolves with more data, and user input and feedback (it’s not static)
- Transparency. Business users define the tags by “building formulas” with concepts extracted from the text (business user makes transparent decisions vs. the technology making black-box decisions)
- Flexibility. If they want to build a new help center article or re-organize the team and tag things differently, they just need to change the formula and re-organize the concepts!
As these business formulas are built, teams can quickly organize the information and react quickly to organize their support operations by setting up priority, triaging tickets or setting up automated responses.
What’s the future? What’s in it for early adopters?
As COVID-19 times make organizations convert faster into automation, not only companies will need to adapt, their employees will need to adjust to this transformation and the ones that adapt first will add more value to their organization.
Employees will need to adapt to this transformation and the ones that adapt first will add more value to their organization.
Business users will learn how to build these language formulas, same as they learned Excel in the past and they will own their processes. This will generate:
- Sharing of best practices for developing language formulas
- Sharing of best practices for processes in different organizations
- General knowledge vs. specialized knowledge. There will be formulas that will be the same for everyone and others that won’t! And the system will benefit everyone with better suggestions adapted to them.
The right revolution in automation using AI will be transforming support teams to do high-impact tasks, not substituting them. And we will have a better solution for our society, for companies and for customers that in the end will get better, faster, more personalized attention.
Thank you to Xavier Amatriain, Kristen Durham, Rachel Obstler, Fernando Domínguez, Fernando Agüero, Enrique Fueyo, Ángel Castellanos, Ignacio Vilela for your comments and help with the article.