The industry is abuzz with talk of AI (artificial intelligence). Advocates claim that contact center AI can do wonders to improve CSAT and lower operating costs. But how credible are these claims? It pays to carefully consider the specifics as to when, why, and how you plan to utilize AI in your contact center. This article provides context for how current AI solutions work, allowing you to decide which problems to solve, and offers eight best practices to help select and successfully manage an AI pilot project.
Why Consider AI in your Contact Center?
Many contact centers face these common problems:
- Rising customer preference for more effective self-help/self-service options. Building these interaction flows is difficult.
- Agents spend valuable time performing repetitive tasks like answering the same questions or completing the same after-call work tasks.
- High labor costs due to finding, training, and retaining qualified agents. Furthermore, forecasting schedules to meet fluctuating contact center demand can lack precision and create additional inefficiencies.
What would the ideal solution look like? An automated solution capable of performing routine and repetitive interactions and delivering effective outcomes-- thereby eliminating the need for live agent assistance. The result? Lower, more predictable demand for live agent assistance and ultimately lower labor costs. Automation is a proven technology for reducing labor costs. AI is the next generation of automation technology. Therefore, AI is a prime candidate for modern contact centers.
For example, the right AI bot can deliver effective self-service options; augment a WFH workforce or a workforce which faces fluctuating service demands, and deflect and perform other highly repetitive tasks.
This is one example of one set of contact center problems that AI may solve. However, there are likely many more. When we talk about using artificial intelligence in the contact center, we are talking about an AI service that can be used across many different call center applications. So, while we may talk specifically about an AI chatbot, we are really talking about a chatbot powered by a shared AI service. And if we talk about AI-powered analytics, we are talking about an analytics application powered by the same underlying AI service.
Now it’s time to consider whether an AI application is the right solution at the right time.
Is it the Right Time to Put AI to Work in Your Contact Center?
Our research has uncovered some interesting data. First, in 2020, for the first time, 62% of surveyed respondents reported better First Call Resolution (FCR) with un-assisted self-service options than with live-agent assistance. Second, Millennials strongly prefer chat over voice. So, when you consider a growing preference for self-service and chat you have an interesting nexus for introducing AI self-service chatbots.
When considering AI, it’s also equally important to consider whether an organization and its contact center are ready for AI. A good way to do this is to look at overall operational maturity. Here are a couple of things to consider:
- First -- for AI to deliver transformational capabilities, it must be used in a very open environment. When we say open, we mean that all relevant systems and functions are accessible through a web API. Unfortunately, older contact center infrastructure systems do not adequately meet this very crucial requirement. A contact center built on a legacy tech stack frankly has more urgent modernization needs (think omnichannel routing, support for digital channels, flexible self-service, advanced QM, better overall integration, etc.) We will talk more about this in a minute.
- Second – the organization must already have a very strong CX mandate. This means there are operating procedures and a culture already in place that is focused on delivering the best customer experiences. If no fundamental focus exists for delivering mature CX, then AI investments and benefits will be diminished.
Therefore, if either of these two pre-conditions are not in place, it would be best to make improvements before introducing artificial intelligence in the contact center.
What is the Anatomy of an AI bot?
Primitive forms of self-service bots have been used in contact centers for some time. The earliest versions relied on automatic speech recognition (ASR) coupled with rules-based actions. These were very static and limited to very narrowly defined problems.
The most common AI bots found today represent a second iteration. These bots rely on AI-driven natural language processing (NLP) to better understand speech and its meaning. However, these bots continue to be coupled to rigid, rules-based actions.
The next generation of AI assistants will continue to utilize NLP and execute actions determined through AI learning by matching the best action for the expressed intent.
From this we can see that AI bots consist of three major components:
- A medium which allows interaction with a bot (across devices and access channels). For example, think voice, chat, or social media channels.
- A service which discerns the intent behind an interaction. As discussed, this has evolved from automatic speech recognition (ASR) to more sophisticated natural language processing (NLP).
- A service back end which takes the action and parameters to deliver an expected response – e.g., fetching or calculating an account balance.
As a bot represents a new kind of application – one with a language-based interface, as opposed to a visual and tactile interface - bots increasingly will take on a persona which makes two-way communication easier for humans. We will discuss this more later.
How Do You Build an AI bot?
AI applications require a rich and diverse technology stack. Modern IT applications, including contact center software, now live in the cloud. This means you have many build options available.
Here are the key elements:
AI service, often a cloud-based service, contains most of the AI solution elements except for the delivery backend. This typically includes the natural language processing (NLP), the machine learning (ML), and some form of integration UI and API.
You can select a general-purpose AI platform like those offered by Google, Amazon, IBM, Oracle, or similar. Think of any AI platform as “middleware”. These AI platforms are general-purpose, extensible, and easy to integrate with other cloud systems and applications using cloud APIs. However, these platforms may require more work because you must design the “conversation”. The conversation includes the intents, anticipated verbiage, responses, and conveyance of the responses the AI application must service.
As an alternative, you may select a specialized AI platform like those offered by SmartAction, Omilia, Linc, and similar. These AI platforms are often specialized by industry. This means some of the conversation has already been pre-built based on problems, questions, phrases, and answers common to an industry. Because of this, specialized platforms require less work and reduce some of the effort and time needed to build an AI bot.
Finally, you have the delivery backend. This is not an AI service, but is that component of an AI application which connects to your business systems to obtain business rules and data. The delivery backend is typically part of your IT infrastructure. For example, if an AI bot requests an account balance, the delivery backend is that part of the AI application that fetches or calculates the balance.
CCaaS as an AI Service Platform
Even with specialized AI services, there are still some critical services that are needed. This typically includes the communications services (e.g., network, voice, chat, IM, social, etc.). It’s also extremely helpful to have some sort of configuration UI to help integrate the AI bot into your work processes.
A CCaaS platform has all the underlying communications services needed to interface an AI bot with a variety of interaction methods. It also contains much of the data and business logic needed to drive conversational AI responses. Furthermore, a CCaaS may even offer a UI interface, making it much easier to integrate an AI bot with contact center work processes.
So, while many AI applications today are initially being built using generalized or specialized AI service platforms, much of that AI functionality will eventually migrate into other SaaS platforms, like a CCaaS.
8 Best Practices to Ensure Your Contact Center AI Pilot is Successful
Let’s now pivot and talk about the best practices for how to introduce AI into your contact center.
Best Practice #1: Build Organizational Alignment
Our first best practice calls for building strong organizational alignment. The objective is to make sure the AI project aligns with organizational priorities and measures of success.
Start by identifying the owner of the initiative. This is the person or persons who own the decisions and are accountable for success. The owner will likely be a business or IT leader.
Next, tie project outcomes to the organization's current strategy. Often, this comes down to growth, revenue, and profit. For example: If the organization has goals to operate at a 30% profit margin and increase market share by 10%, then how does the proposed AI project directly support these goals?
It’s best to achieve this alignment by collaborating with other stakeholders. For example, does this project impact IT, Marketing, or Operations? If so, collaboratively assess project outcomes considering the goals and priorities of these teams. Debate the impact – the costs, benefits, and tradeoffs; look for compromises as needed, and determine how success will be measured.
The goal is to get everyone “on board” and to eliminate detractors. Look for near-term success so you can plan for later expansion.
Best Practice #2: Have Very Clear Use Cases for Your AI Solution
Our next best practice calls for identifying a very clear set of use cases. Our objective is to make sure we solve the right problem while also taking care to eliminate costly scope creep.
If the purpose of your project is to lower costs or help increase market share, be specific as to HOW. Identify a well-contained set of use cases and be explicit as to what problems the project will solve. This will set clear expectations and help avoid scope creep which in turn will help ensure success.
In addition to having a very clear set of use cases, it’s also important to understand and defend why AI is needed for this project.
You can help refine your use cases by thinking about it as a type of root-cause analysis. Start with your immediate facts and assumptions and ask critical questions until you land on specific problems that must be solved. Here is an example:
Problem: labor costs in the contact center are too high.
- Q: Why are labor costs so high?
A: Because call volumes are very high.
- Q: Why are call volumes so high?
A: Because we have a lot of calls asking for help on routine problems.
- Q: Why are agents assisting customers with routine problems?
A: Because there are no self-service options.
- Q: Why would self-service options help us lower costs?
A: Because this would allow us to deflect calls.
- Q: Why would call deflection save us money?
A: Because fewer agents would be needed to handle non-routine calls.
In this example, deflecting routine calls using self-service will help improve our profit margin objective. AI can help because it will be able to better understand intent, when compared to ASR or DTMF IVR call trees, and offer the best self-service options. Finally, we will define and measure success by call deflection, FCR, call abandon rate, and CSAT.
Best Practice #3: Determine How to Build Your AI Solution
Our third best practice calls for determining how we will build our AI solution. Our objective is to make sure we have all the needed prerequisites.
When you think of building an AI-powered application you must assume that some assembly is required. Today’s solutions are not completely an “out of the box” experience. How you decide to build your bot will depend largely on your use case and the capabilities of your people and systems. Conducting this assessment will help you determine the readiness of your organization and provide guidance on how to plan your project accordingly.
With this step, there are three key considerations:
- What skills do we need, and do we have them (or where can we get them)?
- What technologies will we choose to implement our solution?
- Do we have the accessible backend services needed to provide the responses?
Let's take a closer look.
There are roughly six key roles that are needed to successfully deploy artificial intelligence within the contact center. Depending on your project, some skills may be more important than others. These skills include: AI and ML engineers, UX engineer, KB Manager, Contact Center Leader, Data Scientist, and Customer Experience Leader. Not all projects will require each of these skills. However, there are a few key roles to focus on:
- A customer expert. It is extremely important to understand the typical interactions the bot will be designed to service. This understanding may come from the contact center leader or customer experience leader.
- User Experience designer. Giving a bot the right personality is essential because it needs to conform to your brand and audience. AI adds a new degree of complexity with its voice interface.
- Other engineering roles are also crucial and needed based on project requirements for AI training and how your AI application interacts with your services backend.
Also be aware that while some skills may be implicitly provided by working with a vendor or by using a specialized AI framework, you still need to be able to operationalize your bot, which requires these skills on an ongoing basis once it’s put into operation.
Finally, consider what technology will be used to build the AI application. In this phase, you will need to determine whether to use generalized or specialized AI services, or a SaaS delivery platform like a CCaaS with native AI and ML. As part of this consideration, you must also determine whether critical elements of your IT infrastructure (e.g., delivery backend) are cloud-based. If much of your IT infrastructure is still premises-based, you may find it prohibitively difficult to integrate AI with your back-office systems. Legacy, on-premises systems often lack the low-effort integration needed to make these types of advancements possible. Rather than building an AI application, your efforts may be better spent first migrating to a delivery platform like a CCaaS.
Best Practice #4: Build Your AI Solution for Customer Experience
Our fourth best practice calls for placing a focus on the customer experience by solving a problem which is valued by the customer in a way that enhances your brand.
When you build for the customer experience you are building an interactive customer experience that:
- Reflects your brand values.
- Solves a specific problem for a specific user persona.
- Delivers services that are valued.
If these elements are not reflected in your AI application, you may do more harm than good.
To build a great AI bot, you must intimately understand your customers’ intentions and interactions. Therefore, it’s important to deeply understand and document your customer journeys for accurate understanding and insights. This will help you design the all-important conversation.
The conversation design determines what your AI bot will say, how it will say it, and what actions it will take. Without understanding who the persona is, the likely topics of conversation, and what the user wants, your AI application may never deliver the experience or value your customers expect.
There are many documented AI hiccups that illustrate this point. For example, remember Amazon Alexa? This virtual AI assistant was designed to help someone easily order Amazon products. The trouble was, it was too easy and would allow children to order products – very expensive products -- without a parent's permission. Adding insult to injury, when an incident (such as a child ordering a dollhouse) was reported by broadcast media, reporters would often quote the child as saying, “Amazon ordered me a dollhouse”. This again triggered listening Amazon devices to order additional dollhouses. In this case, perhaps the designers did not contemplate what would happen when the target persona was not the actual user.
Which brings us to a final and very important point. Your virtual assistant may not be able to help everyone, all the time. So, build your AI application to fail gracefully. This means that when it becomes clear, at the earliest moment, that the bot cannot provide the needed assistance, let the customer interact with a live agent and make the handoff from bot to live agent gracefully.
Best Practice #5: Operationalize Your AI Solution
Our fifth best practice calls for taking adequate precautions prior to placing your AI application into service to rapidly mitigate unforeseen problems. As part of your QA process, and before you place your AI bot into service, do the following:
- Review all upstream and downstream processes for any impact. For example:
- When the bot does a handoff, is the handoff at the right point in the call flow?
- Is there context with the handoff?
- Does the agent know how to receive a handoff from an AI bot?
- If a bot “completes” an interaction, is there any after-call work required and how is this accomplished?
- Is there a way to trigger and capture customer feedback?
- Pilot the initial operation.
Slowly introduce the application and observe and measure results.
- Use pre-defined KPIs as the basis for evaluation. This is where the earlier identification of stakeholders, use cases, and KPIs will pay off. The initial deployment may be bumpy. Expectations may be slow in being met. But having clear objectives, stakeholders, and success measures will help determine what corrective actions may be required.
- Once in production, monitor and optimize. Ask these questions.
- Is the bot performing as expected?
- What is customer reaction?
Again, another famous story and lesson to be learned comes from Microsoft and the pilot of their early chatbot called “Tay”. Tay was built to learn language and conversation by interacting with people. But this learning was initially unsupervised and unfortunately, Tay learned all the wrong things from the wrong people. Tay was short-lived and was decommissioned shortly after.
- Assess impact with external teams and stakeholders.
- Use predefined KPIs as performance benchmarks.
- Have plans for responding to sudden needs.
- Make immediate adjustments.
- Create new actions.
- Know when to pull the plug and what to do next.
Best Practice #6: AI Care and Feeding
Our sixth best practice calls for taking ongoing steps to ensure your AI application evolves and continues to add value in any circumstance.
Some AI applications, like AI bots, are part of the frontline in a very dynamic information exchange between customers and organizations. Therefore, you must always be prepared to do the following on a frequent and ongoing basis:
- Constantly monitor bot results for any indication customers are not satisfied with the AI experience.
- Teach your AI application to handle more diverse interactions.
- Make immediate adjustments, create new actions, and make all relevant information available to your AI application. If your business experiences a sudden change does your AI application have the skills and facts needed?
Finally, the need to monitor and adjust a bot only serves to underscore the need to retain the critical AI skills used to initially develop the bot. Think of your bot as a pet -- it needs constant care and feeding. If you relied on a vendor to provide the AI/ML engineering when building your AI application, do you still have access to those skills if needed?
Best Practice #7: Account for the Risk of Implementing Your AI Solution
Our seventh-best practice calls for taking every reasonable precaution to deal with risk.
There are many potential risks associated with any new business process, including using an AI bot. You can reduce many of the risks using the best practices we’ve discussed. However, not all risks can be foreseen. Therefore, have generalized contingency plans for the following:
- CSAT rapidly declines.
- Employee reaction is adverse.
- Business impact is undesirable.
- Governance or regulatory requirements change.
For example, while an AI application may reduce contact center labor costs, it may also increase labor costs in other departments. Say, for instance, IT labor costs rise due to maintaining new specialized skills. What is the risk to ongoing operations if this occurs?
You may not always foresee the exact cause but knowing what to do when something unforeseen happens will help you respond decisively.
Best Practice #8: Measure the Benefits of Your Contact Center AI Pilot
Our eighth and final best practice recommends you quantify the benefits and celebrate the successes.
Like all business actions, it ultimately comes down to seeing a sufficient net benefit. A bot will cost resources to develop and support. That cost should be offset by sufficient benefits which cover initial and ongoing costs.
With most IT projects, the development and operational costs are more easily quantified. In addition, operational benefits will be easier to determine if you started with a clear set of use cases and baselined your measures of success.
In the case of our AI bot, we determined that our objective was to lower contact center labor costs using an AI bot to deliver self-service options, thereby reducing the need for live agents. Our measures of success were call deflection, first call resolution, abandon rate, and CSAT.
For simplicity's sake, we will a) assume average operating costs AND b) that abandon rate and CSAT are neutral. This will allow us to focus solely on call deflection and FCR. If our AI bot can reduce the number of calls requiring live agent assistance by 25%, then on a monthly volume of 25K calls, we have reduced agent-assisted calls by 6,250. If the labor cost per call is roughly $6.25 then this will yield a labor savings of nearly $465K annually (other related savings are also possible).
If the incremental cost of development was $500,000 and the ongoing annual incremental operating cost is $150,000, then it will take a little over 18 months to pay back the investment and then produce $300K + in annual labor savings.
When the benefits are measurable and impactful, then it may become easier to expand the use of AI in the contact center.
Introducing artificial intelligence into the contact center will be a new experience for many organizations. It’s likely you may not have all the needed skills in-house. Therefore, using these best practices, begin by identifying a candidate AI project. Next, identify a partner you are comfortable working with. Together you can start creating a strategy that is just right for you.
Special thanks to the following who contributed to this article:
Graham Allen, Sr. Director, Product Management
Lori Britt, VP Consulting Office
Wes Chipman, Principal Integration Software Engineer
Derek Humphrey, Software Integration Engineer
Matthew Riccardi, Principal Implementation Manager
Eric Shelley, Principal Implementation Manager