Enabling Email Automation in Shared Services

Email represents a huge cost centre for shared services and business as a whole. Just below phone calls, it is the most expensive channel to manage and process - with an average cost of £4.55 per email. If that doesn’t seem significant, consider that the average agent or employee spends three and half hours of the working day just in Outlook. Over the course of a year, the cost of manual email processing can be as high as £10,000 per every single agent.

Harder to measure, but more severe, is the indirect impact email has on value creation. The more time agents spend processing transactional emails, the less time they have to focus on the most important requests and valuable customers. Email constrains capacity and prevents shared services centres from rolling out new value-adding services. Indeed, nearly a third (30%) of workers say email is their biggest distraction from real work.

There’s only so much that agents and email providers can do to ease the burden of email. Ultimately, to free your agents from the inbox, you need to step in with automation.

What’s stopping email automation in the shared services centre?

The question is, can email actually be analysed and automated? Emails are processed today in much the same way they have been since the 1970s - manually, with a human on both ends of the workflow. It’s only very recently that a new generation of AI-based solutions has started to emerge which claim to automate the process of sending and receiving emails from end-to-end.

Just how realistic are these promises, and is a new solution even necessary? Many shared services leaders have recognised the hidden cost of email. The most common approach has been to migrate these emails to a case management or workflow system. However, while this has given the process more structure and accountability, it’s failed to reduce the amount of work required. In fact, such systems can actually increase manual effort where service agents are required to manually log case types and other forms of management information.

On the other end of the spectrum, a smaller number of services leaders have attempted to innovate, developing their own rules-based automation systems. Yet such projects more than often turn into costly failures before they can even be deployed. As a result, shared services has only become more sceptical of email automation. Meanwhile, the business decides that throwing more agents at the problem is the only solution to exploding email volumes.

The challenge faced by these and similar solutions is that successful email automation requires both understanding and execution. We have solutions that cover the latter process. An RPA bot, for example, could easily forward a message or respond with a predetermined canned response. The problem is there’s no guarantee that these actions will be the best ones for that message without some sort of underlying intelligence. Yet solutions for the crucial understanding phase have been much harder to find.

Natural language is the missing link. The majority of emails are communicated in freeform, unstructured and largely conversational natural language. This is fine for humans, but machines have traditionally struggled to comprehend it. Even when exposed to pre-trained language models, most solutions fail to appreciate the unique context and demands of a specific service type, leading to incorrect decisions and bad outcomes.

NLP: Enabling email automation

However, despite these challenges it’s important not to dismiss email automation. In the last few years, advances in natural language processing (NLP) have made the technology a game-changer in helping machines understand natural language. With today’s powerful NLP models, the end-to-end automation of low-value repetitive requests has become viable for the first time.

With the advent of Transformer-based architectures, we have seen the emergence of giant language models able to train on truly vast amounts of data. Alongside the expansion of computing power, this has led to the creation of models that are sophisticated enough to reliably understand email communications. In fact, the latest general language understanding evaluations show that NLP models now routinely and consistently outperform humans in reading comprehension tasks.

The practice of Active Learning also produces models that are more reliable and valuable within their business context. This describes the training of models when human subject matter experts – service agents in our case – are present to monitor and correct predictions before the model is deployed. By absorbing the knowledge and experience of these agents, NLP models gain an appreciation of the unique context of the service. This enables them to recognise the intent behind specialist language and jargon, helping them respond appropriately to messages and take the next best action every time.

Of course, process understanding is only the first part of successful email automation. Once you have detailed visibility into a comms-based process, you can see what parts of the process are redundant and what is repetitive and transactional. You should then find ways to eliminate the unnecessary procedures, removing them from the workflow, and to automate those repetitive parts of the process.

However, to execute the appropriate response to a message, an additional automation component has to be present to carry out the action. That’s why the latest NLP-based solutions allow for easy integration with automation solutions like RPA. When working in concert, these technologies allow shared services centres to understand and then correctly action requests and emails, at speed and at scale.

Yet, expectations need to be tempered. Not every message can be handed off to a machine. Most emails are routine and repetitive, but some of the most important ones will need special expertise, more detailed work and further engagement with customers both internal and external. Typically, between 10 and 50% of communications processes can either be automated or eliminated - leaving the remainder of the work to your human agents.

Efficiency and capacity liberation

There will always be messages your agents must read and respond to. However, the automation of low-level transactional emails is crucial to levelling up shared services and liberating its capacity for new and more valuable work. The shared services functions that take the initiative now, will help their organisations to be more efficient, more innovative, and more competitive.

Learn more about how shared services leaders are using NLP and Communications Mining to drive process improvement, eliminate inefficient requests and automate transactional conversational work.