It’s time to rethink conversational AI

Andrew Moorhouse delves into the complex landscape of conversational AI, examining the efficacy of multi-channel approaches and the need for intent-level intelligence.

13
Chatbot Chat with AI, Artificial Intelligence. Man using technology smart robot AI
Contact CentreInsights

Published: August 10, 2023

Andrew Moorhouse

Andrew Moorhouse

It’s an omnishambles out there

Enterprise call volumes are not declining, despite multi-million-pound investments in self-serve technology, conversational AI, and the addition of channels like WhatsApp. The multi-social-omni-everything approach is just not working. Indeed, looking at a sample of 8 billion calls, at best, call volumes have plateaued. The spurious vendor claims of 10 to 20% OpEx reduction never materialised, and the addition of channels has only increased complexity.

You’ve all heard the vendor-propagated rhetoric, “Meet the customer in their channel of choice?” Well, how well is that working out for you? Not so good, eh? Do you really want 2,000 customers reporting they have no heating via a Facebook comments section? Or banking customers reporting fraud via LiveChat when you know they must phone to resolve? Or perhaps 3-day wait times on your new WhatsApp installation? One global consumer goods company receives an incredible 6% of all customer queries via its LinkedIn page!

Whatsapp message
Source: A ‘Big6’ UK energy company

I’d assert the creation of separate “digital teams” that bolted on additional contact channels has made things worse. Worse for both the enterprise and the consumer.

Due to the frequent failure of conversational AI efforts and the Death of Live Chat, enterprise firms now see the voice channel as the biggest opportunity to drive efficiencies via automation and self-serve deflection. But initial efforts are slow, tentative, and not generating significant returns.

The disconnect between the voice, omnichannel, and chat virtual agent systems means there is huge duplication of work, and until these silos are aligned, there is an inability to design functioning seamless journeys.

The rise and fall of Voice AI

We are about to witness the rise and subsequent failure of conversational Voice AI over the next 3 to 5 years. The intent of this paper is pure. As well as being a cathartic exercise, it shares the emerging approach being taken by a few firms that are leveraging a new breed of AI-powered conversation intelligence and AI-triage tools. Especially as they ramp up for Voice AI.

Don’t worry; this is not a thinly veiled whitepaper. In fact, it’s quite the opposite. To prevent the anticipated onslaught of badly designed Voice AI installations, this paper offers real-world case-study insights to provoke and challenge your thinking and, hopefully, help drive call containment and improvements in your digital deflection and Voice AI efforts.

Not all intents are equal or resolvable

A tier-1 bank was struggling to deliver any ROI from its conversational AI and live chat investments. On closer inspection, the bank had 180 pre-determined IVR telephony disposition codes; just 120 intents within its conversational AI platform, yet over 480 self-serve journeys online. None of them were linked up. The impact of this disjointed approach was severe.

Failed intent resolution graph
Source: ALITICAL Research

Digging into the minutiae, it was clear that 35.5% of NPS detractors were unable to have their intent resolved on live chat. Ever. Intents like reporting fraud, ATM disputes, and direct debit indemnity requests cannot be resolved via human agents on this channel. Moreover, complex queries like consolidating an ISA (a bit like a 401K) are too complex for even human live-chat agents.

…over one-third of NPS detractors are caused by a failure in contact strategy and conversation orchestration.

This means that over one-third of NPS detractors are caused by a failure in contact strategy and conversation orchestration. This isn’t an artificial intelligence problem; it’s a lack of intent-level intelligence and badly designed contact triage and routing.

Looking at the IVR disposition codes, you see that 37.92% of inbound telephony involves the customer reporting a problem. However, the chatbot /conversational AI team appeared to be unaware of these data. Indeed, one digital lead made a point of calling the in-house Verint telephony analysts, “The analog team.”

IVR customer intent graph
Source: Tier-1 bank

Upon deeper analysis, it turns out there are over 50 intents (dispositions) that cannot ever be resolved via the conversational AI route, even when escalated to a live human agent,

Banking on self-serve

The bank determined that 51.73% of its 14 million inbound calls could be deflected to a pre-existing self-serve journey (the transactional banking queries above). Also, it needed to create 1,200 intents to accurately identify all inbound customer contact across all channels (as opposed to 120 intents today). The most powerful step was aligning the naming convention across every channel. This is something termed unifying the customer contact taxonomy and is perhaps the most overlooked step today in conversational AI implementations.

The disconnect between the IVR telephony, self-serve and digital teams was so severe that the bank’s vision of seamless journies and becoming a leader in ‘conversational banking’ was impossible to realise until these silos were aligned.

…unifying the customer contact taxonomy is perhaps the most overlooked step today.

It’s also very important to discuss the “containable” intents when discussing containment. That is, don’t include the intents that can never be resolved via a bot. The bank is targeting 20% of “containable” intents on inbound telephony. That is, 20% of the 51.73%, which ‘should’ equate to approx 1.4 million calls per year being deflected to an existing self-serve journey.

So how can this insight be applied as a template for transforming your operation? Well, it’s time to go back to the fundamentals of mapping customer intents for all contact channels.

Rethink your approach to digital deflection, self-serve, and automation

A brand-new AI-driven approach to intent clustering and classification is proving to be the key to success at a select number of enterprise firms. This case-study-based article describes how they explore and curate intent taxonomies and topics to help build long-term digital transformation strategies and roadmaps. To ensure that automation, deflection, and Voice AI initiatives succeed, leaders should pursue the following three strategies:

Automation Strategies graphic
Source: Andrew Moorhouse

1. Unify your customer contact taxonomy

Unifying and creating one taxonomy for all customer intents is the bedrock for seamless customer journey design and getting customers to the single best place for resolution. This list must span every channel. The emerging technology (and your saviour) here is the AI-powered intent mapping tool. These tools strip out all the noise from the analysis of voice and chat conversations and automate the creation of an intent taxonomy.

Unifying and creating one taxonomy for all customer intents is the bedrock for seamless customer journey design.

This is a revolutionary step away from gleaning caller insight from your pre-programmed IVR menus. There may be multiple, complex intents in a single utterance. And the complexities of inbound contact are not reflected in the deterministic, IVR flow; one that was probably mapped out over 10 years ago.

The new breed of intent mapping tools offers incredibly advanced techniques that were previously the domain of Ph.D. data scientists. Leaders can convert a mass of data into structured hierarchies /taxonomies that scale across workspaces.

For the avoidance of confusion, in this context, taxonomy is a fancy term for a list. A list of intents. But a list that has a hierarchy of main intents and then nested (sub) intents. (I’m a biologist by trade, and sometimes these overly verbose terms creep in)!

Here’s a great example to show the type of tool available and the user interface:

Taxonomy tool example
Source: HumanFirst.ai

Harness the insights, align the silos

The foundation for any voice automation and deflection approach is understanding intent-level resolution and satisfaction across voice and text for all channels. Before you decide what to automate, you must understand the minutiae.

Below is a sample output from an intent-level mapping exercise. Both resolution rate and NPS (customer satisfaction level) were analysed from 10,000 live chat transcripts at a tier-1 bank. Understanding what can be pushed to self-serve and what intents require immediate escalation to a live human was a pivotal step in planning the self-serve and automation strategy.

Intent-level intelligence

Intent-level intelligence graph
Source: ALITICAL research. n= 10,000 intents mapped out for a tier-1 bank

It’s clear that only once you connect the silos can you deliver seamless customer journeys. But there’s an even more important step, one that requires more than just technology investments.

2. Define the single best way for contact resolution

The tech-vendor-led rhetoric of “meet the customer in their channel of choice” is well-intentioned but misguided. You see, meeting the customer on Twitter, within the Facebook comments section or even on Live Chat offers zero benefits if the customer is forced to phone for resolution. This is pushing the burden of resolution back to the customer. And it’s a serious and significant failing, often euphemistically termed ‘channel shift’. As we saw previously, ineffective contact strategy design led to 35.5% of detractors being unable to have their intent resolved via Live Chat.

The pivotal change we see here is that some enterprise firms are scrapping the old mantra and determining where the customer should have that intent resolved. They are pulling together cross-functional teams to figure out what is the single best, optimal way to resolve each specific customer intent. It’s not the technology. It’s the application of the technology that matters. Of course, meet the customer in their channel of choice; but resolve it in your place of preference. Or perhaps, the mindset shift required here is to meet the customer in their channel of choice and triage /route to the single best place for resolution.

…meet the customer in their channel of choice and triage to the single best place for resolution.

Case study insight: Emergency roadside services provider

Experiencing a breakdown on a highway or interstate can be a traumatic experience. In this situation, customers need certainty. They require precise timescales and continual reassurance that the recovery is on track. As you may imagine, this leads to multiple repeat contact within the call centre from distressed customers. As a cross-functional team, the CX leadership, operations and digital teams at an Emergency Roadside Services Provider decided that there had to be a better way; both in delivering a better experience and reducing impact on the contact center.

They replaced the legacy IVR system with an interactive voice assistant and the entire conversation is automated with conversational voice AI.

Customers then receive a SMS to pinpoint and confirm their location using GPS intelligence. Via an Uber-style app, the customer can track the progress of the recovery truck. Subsequent follow-up is via SMS, keeping the customer informed of progress. Most customers phone when they breakdown, but some do send a message. A select few even use FB messenger. But, irrespective of what channel the customer request originated from, the resolution is the same.

GPS mobile phone example
Source: Andrew Moorhouse

Irrespective of what channel the customer request originated from, the resolution is the same.

Within the first month, call containment (resolution without a human agent) achieved 48% and is now at 70% one-year following implementation. What’s essential to note, the breakdown assistance provider did not attempt to automate all intents. They chose one use case and executed it brilliantly.

Not all intents are equal

The key message is that not one size fits all. Each intent requires a different solution. The most important point to glean is that irrespective of where the intent started; resolution occurs in one single defined place. Some other examples that we’ve seen:

Example of intents
Source: Andrew Moorhouse

Getting surgical on containment

Understanding the intent and then routing it to the best place for resolution is the fundamental activity here. During the intent clustering activity, there will be clusters that the intent classification tool didn’t understand. The approach is to get very granular and disambiguate what customers want. These may be termed ‘edge cases’.

In this instance, it’s vital for the cross-functional team to decide with these topics where to send the call. Deflecting initial voice traffic to pre-existing self-serve menus can deliver results quickly, often within month one. The final element required to transform performance on containment is the application of intent-level orchestration.

3. Apply intent-level triage and routing

Once the intent resolution rules above are determined, then this is where the ‘intent-level orchestration layer’ comes in. Its sole purpose is to drive the intent to the single best place for resolution. It’s worth explicitly stating this is not customer journey orchestration; it’s quite different:

journey orchestration v intent level
Source: Andrew Moorhouse

Next available agent is the worst available idea

The current contact centre model is to stack customers up in a call queue. Perhaps the necessary agent skill has been identified, but today, there is no prioritisation given to any customers based on their call intent.

Imagine you are involved in a traumatic accident. You broke your femur and shattered your pelvis in a motorcycle crash. However, Clive, who visits the hospital A&E weekly, has stubbed his toe and currently sits in the queue in front of you. This is how almost all contact centres operate today, and it’s ridiculous (sorry, Clive). Curiously, many UK hospitals are lambasted for the fact that 24% of patients leave A&E without ever seeing a doctor. I think that’s a phenomenal achievement and shows that the 4- or 5-stage triage principles are being applied and working.

Triage is the most under-discussed element of contact centre strategy; because, until now, it has been incredibly difficult to understand the minutiae of customer intents across all voice and text channels to take decisive and determined automation decisions. Here’s a great example of Triage in action today:

Triage example
Source: Sentisum

Triage is the most under-discussed element of contact center strategy.

Intent-level orchestration (triage) tools are emerging that plug into existing CCaaS platforms, vastly improving performance. The evolution of intelligent triage allows leaders to intelligently prioritise support based on urgency, risk and customer effort and reduce response time to important conversations. The new breed of intelligent AI triage works as a contact orchestration tool across voice, email, chat, and all available channels.

Case study insight: Insurance provider

One global insurance company achieved a 17% reduction in inbound volume from a base of 120 million phone calls. But it took a methodical and data-driven approach. Across 12 different countries, they started with intent identification within its IVR to determine the possibility for end-to-end voice AI resolution.

The company initially collected intent data from the conversational IVR from 18 million customers and simply then routed the customer to the main agent queue. Only once they had this depth of insight did they run the intent clustering and identify what intents could be resolved by “one shot” self-service. That is, what basic intents could be resolved by looking up a simple database or API lookup request.

One specific use case is handling customers that repeatedly call in to check if their funds have been processed. If the customer already called one hour previously to check the payment status, just via the phone number identification, the insurance company now routes the customer to a payment status portal, showing them the required details. But this was not the starting point. Their advice is not to start with automation but to collect the data. And focus on the one-shot use cases coupled with deflection to pre-existing self-serve journeys.

Buyers are not aware of what’s possible

The amount of unstructured data within voice and text is so vast and so valuable, yet few organisations are using it today to drive call containment efforts. The following image depicts a fictitious, authenticated banking live-chat session. This is the future evolution of conversational AI. To achieve this, you need the right intent-level intelligence coupled with the best available AI triage and intent-level orchestration. Advising on ISA (401k) consolidation is a messy affair. Chatbot resolution rates are zero, and curiously, human live-chat resolution is less than 10%. Now armed with the right data, the intelligent triage system understands that resolution rates for ISA consolidation conversations are extremely poor, even via live chat. So, a trained ISA specialist is offered for the conversation via a co-browsing session.

Customer service text example
Source: Andrew Moorhouse

*An ISA (individual savings account) is a UK individual tax-free savings account. The Canadian equivalent is a Tax-Free Savings Account (TFSA) and in the USA, the closest is a Roth 401K.

Recommendations

Brand new AI technologies and contact strategy approaches are emerging that bolster existing installations. To ensure initiatives succeed, leaders should:

Unify the customer contact taxonomy across voice, live chat, bots and socials using a new breed of AI-driven Intent classification tool. Only once you connect the silos can you deliver seamless customer journeys.

Define the single best way for intent resolution. This needs to be done as a cross-functional team. Irrespective of the starting point, leading organisations are shifting to a new mindset: “What is the single best way to have this interaction?”

Apply AI-powered intent-level conversation orchestration to drive the intent to the single best place for resolution. This is not mere journey orchestration but a new approach using Intelligent AI Triage to deliver on automation /deflection /self-serve decisions.

In summary

We are about to witness the rise and subsequent failure of conversational Voice AI over the next 3 to 5 years. It’s not an AI – artificial intelligence problem. It’s a human intelligence issue.

Intent-level orchestration is the most significant missing strategy piece. Contact orchestration isn’t owned by voice; digital; CX leaders; the contact centre operation; social media teams, or marketing. Today’s stakeholders have zero bandwidth or time for the right cross-functional strategy decisions on what intent should be resolved. The narrative about “meeting customers in their channel of choice” is a horrific idea for intent resolution propagated by legacy tech vendors selling omnichannel consolidation platforms. Customers should be triaged in their channel of choice, and their intent routed to the single best place for resolution.

Customers should be triaged in their channel of choice, and their intent routed to the single best place for resolution.

The biggest risk is charging ahead on Voice AI without defining the correct intent-level orchestration. Without effective triage, prioritisation and deflection rules defined at the intent level, any Voice AI attempts will fail. The intelligent orchestration layer needs to sit across all digital, voice and self-serve options.

Before companies can effectively implement Voice AI as a channel, they must unify their customer contact taxonomy, align all silos and define the single best place for that intent to be resolved.

The most advanced enterprise firms will apply AI triage and intent-level orchestration to revolutionise their call containment and digital deflection performance. The technology exists; the challenge is the bandwidth, in-house capabilities and desire to make this happen. Understanding the nuances of resolution and satisfaction at the intent level is critical for future self-serve and automation roadmaps.

The evolution of intent-level conversation intelligence and triage tools offers unrivalled opportunities for leaders to fuel action and drive improvements. Rather than usurping existing providers with new vendors, leaders should augment and bolster existing investments. Tuning and honing the existing platform performance.

Leaders must objectively discuss all customer intents and consider, “What conversations do we want to have, with which customers, when, and in what channel? What’s the optimal, single best channel for resolution?” Your automation strategy isn’t a post-it note exercise. In a data-driven enterprise, it’s time to trust the data.

 

Artificial IntelligenceChatbotsConversational AIGenerative AIMyCustomerVirtual AgentVirtual Assistant
Featured

Share This Post