Have you ever wondered how Amazon predicts your shopping preferences, Spotify curates your perfect playlist, or Netflix suggests just the right show for your mood? The secret lies in advanced pattern recognition and feedback and learning loops that adapt to your behaviour. These loops involve recognising patterns, making predictions, and adjusting based on feedback. The apps and tech we use daily rely on these AI systems to continually improve. At CloudCapcha, we harness the same machine-learning algorithms and technologies to transform time management. In our latest article, we describe how our new AI capabilities are revolutionising the way users interact with the platform to make time recording more intuitive, accurate and efficient than ever before.
What is AI and what do we mean when we refer to it?
Before explaining our AI module in more technical detail, it is worth putting AI in context: perception of what AI is, what it does, and what it can do, varies tremendously. Indeed, we’d go as far to say that a lot of what is described as AI is just automation that itself isn’t AI.
If ChatGPT is the ‘household name’ or ‘face of AI’, the lesser-known counterparts of deep and machine learning could be described as the “unsung heroes” or “behind-the-scenes architects” of AI. These technologies are the foundational elements that power and enhance the capabilities of many applications working quietly yet critically to advance their field. It is these sophisticated technologies that we have been using to develop our AI module to transform time recording in accounting.
Our desired outcome – and challenge!
WorkCapcha has already succeeded in capturing an individual’s work activities across numerous applications. Our next goal was to automatically categorise these activities by client, project, matter, task, or milestone. By utilising advanced machine learning algorithms, we wanted to define a user’s workday and suggest appropriate categories for time recording. The ultimate goal of the AI module is to eliminate the need for manual data entry and to provide a full day’s accurate prediction of the user’s time entries. The user would merely have to confirm the recommendation(s).
The complexity of the challenge lies in the inconsistencies of the available data and how that data is presented to the WorkCapcha DayBook listeners. No two applications are identical, even from the same vendor!
People are also infinitely variable in their approach to work. Some will work meticulously in a linear fashion – start a task, finish a task, on to the next task. Others will have multiple work tasks “on-the-go” concurrently. Throw the modern tendency for constant interruptions into the mix and the scale of the challenge for WorkCapcha to interpret these signals becomes clear.
Our approach
There are dozens of different data science methodologies, many of which we have used for this exercise. Some methods gave good results, some didn’t and were rejected. In some cases, the combination of different methods was the optimal route. In every case, there was extensive fine-tuning of how the various methods were used.
Essentially, there is no standard “out-of-the-box” solution for data science challenges of this complexity. The methods and results we have achieved represent valuable Intellectual Property for CloudCapcha, so the steps we describe are deliberately vague on some of the key details, but we can say the chosen AI module combines rules-based logic, clustering, non-generative NLP, and multi-layered neural networks to identify patterns and make predictions. It also improves its recommendations over time by learning from the user’s interactions and corrections.
Step-by-step
The first stage in the process was to identify where we have multiple signals for the same work item and de-dupe. In some cases, WorkCapcha can receive as many as 5 signals for the same piece of work. The user might open SharePoint to select a document, open up Word in the browser to review it, decide to open it on their Desktop for edits, and finally write it back to SharePoint. If the user had also been doing multiple saves, WorkCapcha would have received multiple signals for those too.
Here we use a mixture of rules-based logic and clustering techniques to aggregate the relevant signals to give us the one that a user would desire, and the AI would understand.
For the next step, we wanted to run some pre-processing to filter and categorise datasets where we already had high degrees of certainty about the correct recommendation. Here we used some rules-based models and NLP to recognise patterns and score the outcomes.
We then experimented with various techniques, in parallel and serial, including decision trees, linear regression, clustering and neural networks to arrive at the recommendations for each client, project and task. The composite mechanisms used for each of these 3 predicted variables/recommendations are quite different, mainly because the shape and size of the relationships between the users and these elements are quite different.
In practice, the chances of getting the right answer for all three of client, project and task every time are, of course, less than 100%. So, we don’t just pick a single “correct answer”. Instead, we are creating and storing a list in descending order of which clients, projects and tasks are considered the closest match together with our internal “score”. If the “scores” are strong enough, we present the top 1, 2 or 3 options to the users.
Training the model – the power of learning loops in action
For the AI to reach its full potential, it requires training. Users play a crucial role in this process by providing the AI with the necessary data through their everyday interactions. This data is then used to train the AI, ensuring that the recommendations it makes are tailored to the specific needs and patterns of the user.
Every time a user converts an activity to a time entry in WorkCapcha, the AI learns about what is “correct” for that user. We have been using that data for months to train the models.
With the AI in place, every time a user confirms a recommendation, the AI absorbs that information to reinforce the logic. Every time a user selects a different choice to the recommendation, the AI is driven to re-assess its logic and arrive at a better answer next time.
The training models are re-trained every week, so users should expect to see continuous improvement in the recommendations very quickly. There is a single model per Firm and each of these models are held within the CloudCapcha Azure tenancies in the Firm’s relevant jurisdiction.
Overcoming production challenges – new UI & fast performance
We wanted to retain the high-speed performance of the WorkCapcha DayBook. To achieve this, we chose to separate various steps from each other and away from the users’ sight.
The activity collection and aggregation steps are happening before the user sees the resulting activity in the DayBook. Each of the activities are being pre-scored in the nanoseconds before a user can select them for conversion.
This means that we always have 3 Clients, each with 3 Projects, each with 3 Tasks available for each activity available for selection before the user works through the time entry phase.
As the user selects a Client, if it’s one of the top 3, the Project options are immediately available. If it’s a different Client, we re-process the Project AI logic in nanoseconds to make a new recommendation, and so on.
The desired outcome – the ‘Low-touch Timesheet’
The AI module aims to revolutionise the time recording process: the ultimate goal is to eliminate the need for manual data entry.
Ultimately, we believe that we can make a full day’s accurate prediction/recommendation for each user’s time entries – without them having to intervene. They will be presented with their full picture and asked to confirm or edit as they desire (with more machine learning happening as above).
Once a user has been interacting with the CloudCapcha AI for a few months, we expect that the recommendation scores will reach a sufficiently high level to achieve this.
How confident are we?
Very! A glance at other industries that have pioneered techniques such as deep neural networks for their adaptations of recommendation engines is inspiring. Consumer-focused e-commerce sites would be unrecognisable without the “selected for you” choices being constantly presented to every user. Every time we make a decision (purchase or decline), we teach the models more about ourselves.
At WorkCapcha, we have built the logic and we’ve designed the new UI to present and achieve results.
Now we just need users to teach the CloudCapcha AI models more about their working lives and the “timesheet to end all timesheets” can be a reality even sooner than we may have dared hope.