“One more message and that inbox will explode”. Ever had a similar thought? Found yourself in a situation where reading all incoming email is physically impossible for a human being? Can we just put the blame on Ray Tomlinson for sending the first one ever in 1971?
Privately it’s a headache, professionally – real impact on the business. The root cause lies in our nature – we communicate using spoken or written (also sign) language and one sure thing to say about humans – change is hard.
Is there another way? Read along.
Why is this a business problem?
Nothing sinks in better than an example. Say hello to John Smith, your most faithful email-buddy:
It sounds like a simple five–minute task, and maintaining good business relations with John is important.
Now, multiply Mr. Smith by your entire client base and every customer-facing process (including internal ones like intercompany transactions). Suddenly, we’re back to the previous point – the inbox exploding. Let’s do the math: around 80 of these ‘quick’ 5-minute common business requests can fill an entire workday (1 FTE, assuming 6,5h productive time).
After years of transactional communication, we’ve reached a point where it’s common and globally accepted to receive unstructured information as input while relying on systems that demand strictly structured data.
Guess who transforms that data.
You might try fixing this first, and I genuinely encourage you to do so. But more often than not, you’ll hit a wall – organizational structure, procedures, SLAs, or the classic “we’ve always done it this way”.
Every now and then, you’ll spot a rescue boat in the form of structured inputs like forms or ticketing systems. But in our experience, those are rare – and even when they do exist, they often lose the battle to the good old, reliable email or start to resemble one by offering free-text fields.
How to fight the unfair fight
Before choosing our weapon of choice, let’s set the battlefield. We’re diving into Natural Language Processing (NLP) – a field of Artificial Intelligence (AI) and a vast, fascinating topic worth exploring beyond this article. Right now, it’s experiencing a Renaissance, driven by the rise of Large Language Models (like ChatGPT), but it took us a long way to get where we are now.
Getting back to our example, let’s define our objectives. In simple terms, we want to:
- Figure out what John wants – understand the intent of his message.
- Extract the important bits – pull out key data points from the text.
Why these two? Because that’s how (business) processes work: you take an input, process it through a series of activities, and produce an output. Recognizing intent (the message label) helps route it to the right process with a defined outcome, while the extracted data serves as the input. In this case, we could define our process as an “Order Query”, where the PO number acts as input, and the output is an order status update in response.
If you’re up for a deeper dive, these two goals align with well-known NLP tasks: classification (intent detection) and Named Entity Recognition (NER) (data extraction). There are multiple ways to tackle them, each with a dozen different libraries, tools, or (recently) LLMs to choose from. A more traditional approach might rely on Regular Expressions (worth checking out!), which work for simple cases. But we’re dealing with diverse messages and enterprise–scale volumes – think 5-15k emails per month – so we need something more robust.
The goal is to handle this at scale, in a repeatable way, so customer-facing employees can engage with important bits of communication, where AI falls short and human expertise is still necessary or where it matters most: strengthening client relationships, winning new customers and retaining those at risk of leaving (churn). Some speculate that in the near future, interacting with a human will become a premium service.
Plenty options to choose from, but we’ll be exploring UiPath Communications Mining.
The Basic Recipe
By way of introduction, Communications Mining is a cloud–based AI/Machine Learning (ML) service within UiPath’s ecosystem, built to automatically understand and process natural language data. There’s a lot to explore on the platform. We’ll cover the essentials, but if you’re hungry for more, I recommend checking out the official documentation and UiPath Academy courses for a deeper dive.
Most ML life cycles follow a few key stages: planning, training, and deployment – that’s exactly how we’ll approach this topic.
Planning
Communication is inevitable in any business – almost every process requires a conversation at some stage. But instead of tackling everything at once, start by asking yourself a few key questions:
- Where is communication an inseparable part of the process, e.g.: serves as input/output for the whole workflow or particular activities?
- Which communication channels are used in the organization: emails, chats, CRMs, ticketing systems, collaboration apps, etc.? (data sources)
- Who handles high message volumes? Which departments or teams are most involved?
- Are there recurring patterns or structured elements within these business communications?
And the most important one, often forgotten – why do this at all? Is the goal automation, analytics, or both? What KPIs define success? Simply saying “it works” isn’t enough for AI-driven solutions – more on that later.
Let’s keep orbiting around our example: the Order Processing Team handles 5,000 emails per month, most of them similar to John’s – new orders, modifications, cancellations, and inquiries. Asking the right questions and analyzing processes in scope will guide us toward defining a taxonomy for our model.
A taxonomy is a hierarchical structure of intents/categories (labels) which can be nested and the data points associated with them. Some fields are directly linked to specific labels (extraction fields), while others may appear in any message (general fields). A well-designed taxonomy should accurately reflect real communication patterns, as it serves as the foundation for model training and application.
While taxonomy can be modified, changes come with consequences – adjustments may extend training time, and some changes are irreversible.
Training
Basically any ML training is about providing it with enough examples to “learn” from. Under the hood AI uses mathematical and statistical techniques to detect patterns and relationships within the data – allowing it to make predictions on new, unseen cases.
How many examples are enough? There’s no universal answer, but here are some guidelines:
- Generally, more is better. Your training set should reflect real-world data as accurately as possible – reducing randomness and improving reliability. For a Proof-of-Concept in Communications Mining, expect to provide at least 10,000 messages; for a production-grade setup, aim for 60,000+.
- Capture seasonality. Include variations like monthly peaks, end-of-year slowdowns, or other cyclical trends in your dataset.
- Balance your dataset. Ideally, each category (label) should have an equal number of examples. While real-world business data rarely fits this ideal, do your best to minimize major imbalances.
Building and training a machine learning model typically requires programming skills, creating a high entry barrier. Fortunately, Communications Mining allows you to go through the entire process without writing a single line of code (it’s optional). Communication data can be uploaded via CSV (Comma Separated Values) file or pre built integrations with Microsoft Exchange Server or Salesforce.
Training is a super easy and user-friendly experience, almost like a coloring book. With just a few clicks, the business user (typically a Subject Matter Expert) assigns the correct labels (one or multiple) and highlights the relevant data points (fields) within the message text.
That’s it. The platform automatically detects changes and starts training. Rinse and repeat until you supply enough examples. How many? You guessed it right – no universal answer here either, but Communications Mining offers a guided training mode. It provides clear, instructive feedback to help you set measurable goals. Rule of thumb is to maximize model’s performance while minimizing time spent on training (work you need to put in).
As mentioned earlier, operational efficiency of AI-powered automation exists on a spectrum. In classical automation, such as Robotic Process Automation (RPA), you define strict requirements, and the bot follows a pre-programmed path, handling exceptions in a predictable, rule-based manner. In machine learning scenarios you need to face and accept some level of uncertainty.
Machine learning metrics are a broad topic beyond the scope of this article. Fortunately, Communications Mining interface presents key information in an intuitive, easy-to-understand way, while still offering a sufficient level of control. However, we will need to cover one essential concept of confidence (value), which leads us to the next chapter.
Deployment
Let’s bring everything together. Our scope and objectives are now reflected in the taxonomy, and we’ve decided the model’s training is complete – at least for now. Time for AI to do the heavy lifting.
We route John’s original email to Communications Mining, and in return, we receive the model’s prediction for that message:
The platform provides two key automation components we’ve talked about:
- Inferred labels (one or many), which determine which processes to run.
- Extracted fields, which serve as structured input required for that process.
Notice that each prediction comes with a confidence value ranging from 0% to 100% (this also applies to extraction fields, though not visible in this view). This value represents how certain the model is – based on its training – that a specific label applies to this message. It is up to us to decide what to do with that information. A good starting point is to set a threshold value as a cut-off filter. Before drawing any further conclusions, let’s explore a few hypothetical scenarios:
- Happy Path: the model predicts “Order Information Request > Delivery Status” with 74% confidence, surpassing a 70% threshold. The PO number is extracted with high confidence as well. The request is routed to an automation, which queries the ERP system and returns the delivery status. John receives an automated email with the requested information.
- Borderline Case: this time, 74% confidence isn’t enough because our threshold is set at 80%. We either let an employee process the case fully manually, or we implement a human-in-the-loop mechanism that will request human validation before executing the process.
- Unwanted outcome: the model misclassifies the request as “Order Cancellation” and its confidence exceeds the threshold. Another automation cancels the order in the ERP system and notifies John, who is now confused why his request resulted in an order cancellation.
These scenarios are simplified (and the third one is exaggerated), they highlight the many-sided nature of AI-driven automation. That’s why it’s crucial to carefully analyze processes and expected outcomes before deployment.
High-risk actions, like order cancellations, should have higher confidence thresholds, additional rule-based checks, or even mandatory human validation. This hybrid approach may not be as fast as a fully autonomous system, but it ensures that human intervention is limited to reviewing exceptions, while 90-99% of cases run automatically.
Make the best of both worlds
Another major advantage of Communications Mining is its reporting capabilities. Imagine being able to analyze an entire year’s worth of 60,000 emails and uncover facts like:
- The top 10 highest-volume request categories.
- That delivery status queries peak in December.
- That John Smith frequently orders products that are about to be discontinued – a perfect opportunity to offer him an alternative.
These are just simple examples of the endless analytical possibilities hidden within enterprise communication, which is no longer just gigabytes of unstructured data, but a goldmine of valuable insights waiting to be uncovered.
All of this is possible because every message uploaded to the platform receives a model prediction. When you combine metadata (like sender address, domain, timestamp, etc.) with inferred categories and extracted fields, you get a very powerful dataset. Single item can be treated as a training example or case to be processed, but when high volumes are in play, aggregated communication data starts telling stories.
Another key insight hidden in communication data is customer satisfaction. Communications Mining provides built-in sentiment analysis, automatically detecting Tone or Sentiment (positive or negative) for each message.
Beyond that, the platform allows you to configure a weight-based Quality of Service (QoS) parameter, assigning a score from -10 to 10 to each label for deeper context. For example, “Order Cancellation” might have a QoS score of -5 because losing an order negatively impacts the business – even if the message sentiment itself was positive.
Since the platform automatically re-trains and predicts each uploaded message instantly, the communication data is continuously updated, enabling real time monitoring. You can even set custom alerts to track key conditions – for example, if John Smith’s QoS score drops below a certain threshold, the system can send email alerts to his Key Account Manager immediately.
What are you going to build?
I hope you enjoyed this brief introduction to Natural Language Processing and UiPath Communications Mining. We barely scratched the surface, though, there’s a lot more to be discovered and learned, so I encourage you to try things out and stay curious. Technology is advancing at an unbelievable pace, turning yesterday’s impossible into today’s reality.
Feel free to reach out to us, if you’ve got any questions and happy automations!
