All four defining features of AI are present in the Intelligent Email Assistant: the system “senses” the content in the email by receiving the email and attached documents and transforms them into workable text. It then reads and “comprehends” the content of the email using Natural Language Processing (NLP) techniques, and defines the required follow-up actions based on its understanding of the email. The execution of those actions, the “act” component of the AI set-up, is often passed on – either to another robot, to an existing application or to a human operator. Finally, the Intelligent Email Assistant typically embeds machine learning components to continuously “learn” from past events to improve its future accuracy and correctness.
Ability to read & understand incoming emails
NLP uses natural, human language in interactions between computers and people, adhering to the new paradigm of systems adapting to humans instead of the other way around. NLP systems can undertake one of two tasks:
- Natural Language Comprehension: Extracting meaning from spoken/written language.
- Natural Language Generation: Generating messages and text in a natural, human language, for example, drafting contracts or meeting minutes.
Using Natural Language Comprehension as part of an Intelligent Email Assistant can take different forms:
- In its most basic form the system will ‘read the email’ and classify it against a list of categories. This can be a predefined list of categories – often an existing list that has a clear business logic and is already known to human operators. The list of categories can however also be derived from historical data using machine learning algorithms to detect clusters – often resulting in a technically cleaner and better list. In practice, a combination of the two approaches is usually applied, starting from an existing list that is then gradually challenged and evolved. The outcome of this kind of classification set-up is the fast and continuous categorization of emails, which can then be used, for example, to route to the right team or person for handling.
- In a more advanced set-up, an Intelligent Email Assistant can go further and extract specific data from the email and/or the attached documents, hence collecting all data fields needed to facilitate or automate the handling of the email.
- A third set-up variant is that of an Intelligent Email Assistant detecting the prevailing sentiment in an email, looking at choice of words, sentence structure, punctuation, emoticons, etc. providing valuable input, for example, to guide handling prioritization and routing criteria.
Automating vs. enhancing email handling: Enabling people to do more instead of replacing them
As with any AI system, the objective of the Intelligent Email Assistant is to complement and empower people to do a better job in serving customers – processing incoming emails faster, with higher consistency and fewer errors, offering 24/7 or at least extended office hours, handling multiple languages, while at the same time increasing internal efficiency and reducing employee frustration over repetitive tasks.
In case of simple requests and emails, this objective can be reached through the full automation of the email handling. However, a lot of cases will continue to require the involvement of a human operator for at least some of the follow up actions. In the latter, the Intelligent Email Assistant will not only ensure that the email reaches the correct human operator faster, it will also make sure that the human operator has all the necessary data at his/her finger tips to define and execute the optimal actions.
Raising performance of the Intelligent Email Assistant through Machine Learning
Like humans, an Intelligent Email Assistant can’t deliver 100% accuracy. Their performance depends on the linguistic complexity of the emails, the breadth of the categories and data fields they need to extract, as well as the quality of the written texts they are dealing with, e.g. very short extracts can be harder to understand and language can often be ambiguous.
The use of supervised Machine Learning is a powerful means to achieve a sound initial performance both in terms of ‘recall’ or automation (% of emails handled automatically) and ‘precision’ or correctness (% of emails handled with the right outcome – for example, classified in the correct category, or all required data correctly extracted). However, this requires a sufficiently large and clean historical data set (a ‘reference corpus’) – something that often proves to be a challenge for a lot of companies, forcing them to opt at least initially for a different, rule-based set up.
In any case, it is advisable to include a continuous Machine Learning component in every Intelligent Email Assistant solution, allowing it to improve its accuracy and correctness based on past events. Choosing supervised or observational machine learning puts the responsibility and accountability of the changes made in the system firmly in the hands of the human operators – as a means to stay in control of the customer experience.
With time and continuous training, an NLP empowered Intelligent Email Assistant can achieve up to 80%-90% ‘recall’ or automation and 70%-90% precision’ or correctness, with the trade-off between recall and precision tailored to the company’s preferences.
Insights from the field
Experience in the delivery of AI solutions, including Intelligent Email Assistant solutions, to our clients in Belgium and the rest of Europe has taught us some valuable lessons, which our experts are happy to share below.
“The availability of a sufficiently large, relevant and clean reference corpus (defined as a project-specific collection of data samples that are representative of future input email and documents expected to be processed by the Intelligent Email Assistant) offers significant advantages in the set-up and training of an Intelligent Email Assistant. It is therefore worth the effort, even if assembling this reference corpus can prove to be challenging in practice.”
Francesco Cartella (Data Scientist)
“Aligning ambitions in terms of benefits including (but not limited to) the target level of automation and quality at the start of the set-up is key. While the first test runs will need to confirm the feasibility of those ambitions, having agreed KPI’s and KPI targets as well as a solid view on current base line performance enables an objective evaluation of the solution and hence supports management decision-making.”
Aram Mamian (NLP Business Analyst)
“A specific challenge we are facing with our Belgian clients is the occurrence of multiple languages in the same email. For example, we often see emails written in French or Dutch, with an English signature or attached documents in English. Supplementing semantic NLP techniques with advanced statistical NLP techniques that are language independent have proven to be very powerful in these kinds of situation and allows us to create and train Intelligent Email Assistants that achieve excellent performance in an efficient way (for example, avoiding the need to retrain by language).”
Michiel Neels (NLP Business Analyst)
“Given the maturity of the AI technologies underlying an Intelligent Email Assistant, an initial production-ready solution can be delivered in a very short timeframe for a very reasonable investment. Applying an agile approach with a mixed skill team of around three people, a first proof of concept (PoC) can be realized in as little as seven weeks. Key to such a fast delivery is an upfront agreement on the scope to be covered in the PoC, and continuous discipline in sticking to that scope. Their first exposure to the strength of AI solutions often inspires people to come up with new ideas, adding complexity and scope to the AI set-up. Resisting the temptation to overload the PoC and opt for a multi-sprint approach is the best way to achieve fast and consistent results.”
Toon Lybaert (Project Lead)
Want to discuss how an Intelligent Email Assistant can create value for your business?
Feel free to contact us for a chat!