Consider These Things When Formulating Your AI Strategy

(This week I'm sharing a post from our lead AI and machine learning engineer Alice Li. This post originally appeared on Gladly's blog.)

Artificial intelligence (AI) seems to be the topic on everyone’s lips right now. Motivated by a fear of falling behind the competition, and the drive to get their share of the estimated US$14 trillion in revenue from incorporating AI, companies are diving headfirst into implementing AI and machine learning (ML) in their business processes.

But while AI and ML can be a real game-changer for enterprises, it’s important to have a sound strategy in place, to truly maximize the return on your investment. So before you embark on buying an AI solution, here are a few things we think you should consider.

1. Start with a plan

When a technology comes in with the promise of reducing costs and improving efficiency, there’s a temptation to rush straight into implementation. But without a clear strategy in place, there’s a danger of ending up with a string of small, disjointed successes that don’t do justice to your investment.

When creating that plan, you might want to consider not only your current needs, but your future needs as well. What are the 5 or 10-year goals for your business, and how can you incorporate the building blocks to that today?

2. Prioritize ‘good’ data collection

How effective your AI solution is depends almost entirely on the data you feed into it. In fact, the hardest part of creating AI isn’t in writing the algorithms (after all, the best algorithms are all published and open source) —it’s in creating the datasets that train those algorithms.

Not only does that data need to be accurate, but it needs to be representative of the problems it’s being used to solve. If you’re feeding your AI data that’s incorrect or incomplete, the results it produces will be equally incorrect or incomplete.

Start Early and Go Long. Start collecting data early, and continue doing it over an extended period of time. That will help to account for seasonal changes or temporal anomalies to ensure your data is truly representative.

And what that ‘extended period’ means will differ according to the complexity of the data you’re trying to gather. For example, gathering data around whether a customer had a positive or negative experience is reasonably straightforward. But if you want to predict the expected volume of inbound email to your contact center, you might need to observe that over a year or more to be able to collect enough seasonal data. 

Leverage Domain Experts. Getting accurate data can be a tedious affair if you don’t approach it the right way. It’s often not feasible, for example, to have one person read and annotate 100,000 emails to get a sufficient data set for a decent ML model. So instead, try to take advantage of domain experts to help get that data.

We designed our ML-powered Answer Suggestions—which recommends to an agent which answer they should use from the company’s knowledge base—to leverage the knowledge and expertise of our customers’ human contact center agents. Gladly observes what answers agents are using when responding to specific customer issues. And as agents continue to either use the suggested answer, or pick another one, Answer Suggestions keeps getting smarter about what to suggest in the future, reducing the time an agent spends getting back to a customer.

Break down data siloes. It’s important to ensure that your data is aggregated across all the channels you offer. Especially with today’s customers who often move from one channel to the next, it’s valuable to gather that cross-channel data to get an accurate picture of the full path the customer took, the number of touchpoints needed to resolve an issue, the full time to resolution, and the customer’s overall satisfaction. To do that, it’s recommended that companies invest in acustomer service software that unites all the channels they want to provide, rather than separate systems which result in information siloes, or require high-maintenance data pipelines to unlock the data’s full potential. In Gladly’s case, for example, a unified data infrastructure lets our AI make holistic recommendations to agents, such as when it makes sense to escalate an email to a phone call, or when to rope in a team member from a different department.

3. Put yourself in the shoes of the user and measure their experience

Have clear metrics and a follow-up plan to ensure that your AI solution, once implemented, actually improves the user’s experience.

Take the example of the bank that introduced online banking to make it easier for customers to bank with them. While that, objectively, made a lot of sense, the move ended up increasing their customers’ total transaction volume. Those customers began calling and visiting the bank more, increasing costs for the bank and decreasing their overall profits. And it’s the same with any cost-saving AI you implement—be sure that the chatbot you install isn’t actually driving traffic to a more expensive channel, or frustrating your customers.

To that end, don’t just focus on response times, but keep tabs on other important metrics such as churn and CSAT, so you get a real understanding of whether your AI is actually achieving its purpose of making your customers’ experience better.

4. Use ML in the right situations

ML is a great solution when it comes to:

• Streamlining and automating repetitive tasks
• Making complex decisions that involve analyzing a large amount of data
• Detecting anomalies and patterns in data

But while AI might one day fulfill its prophecy of replacing human beings in the workforce, the reality is that today’s technology isn’t sophisticated enough to do that just yet. Nor is it something that consumers want, with a 2018 survey showing that over 50% of consumers were disappointed with their last chatbot experience, and 98% of customers admitted that they try to skip a company’s IVR to go straight to a human agent.

So where a task involves conversing with your customers with a level of empathy, or recognizing and understanding nuance, that’s probably not the best place to deploy your ML resources.

5. Create clear guardrails to keep AI in its lane

The string of recent and high-publicity missteps—from racist chatbots, to thieving AI— highlights the need for companies to put clear and consistent guardrails around their AI technology. And that’s all the more pertinent if you’re using it in a capacity that impacts your customers directly.

The nature of ML is that you’re not explicitly programming everything it does. But it does place the onus on you to install guardrails around the technology, to make sure it doesn’t surprise you in a costly way.

Put reasonable constraints on what ML can do automatically, or thresholds for when a human should be alerted to confirm that a specific behavior is acceptable. For example, if a customer is reaching out about a serious accident or a bereavement in the family. These customers should be flagged to bypass your automation and go straight to a human agent to handle them with the delicacy they require.

At Gladly, for example, we take a four-step approach to AI in Answers Suggestions:

• We start by suggesting actions to agents, streamlining customer service while keeping a human in the loop.
• We would only automate once we’ve learned enough from agent feedback that we’re confident in the ML algorithm’s suggestions.
• Even then, we would give admins controls, such as specifying topics that you want to ensure a human handles, such as bereavement or legal affairs.
• Finally, we make it easy to monitor what happens in the system, so you can audit performance after-the-fact, with powerful reporting and drill-downs.

The recent breakthroughs in ML, especially deep learning, have led to remarkable advances in solving problems such as speech recognition and natural language processing, with a potential to completely revolutionize a company’s processes and their customers’ experience from what it is today. It’s certainly an exciting time to be in customer service today.