Don't let data scientists ruin your RPA project

author

Robotic Process Automation (RPA) is the fastest-growing enterprise software category, with large companies scrambling to automate manual tasks while avoiding costly projects to renew their legacy systems.

Mainstream RPA today still mostly covers simple tasks, not processes. There is nothing wrong with the simplicity that gets the job done and brings savings without massive development costs. That is absolutely where the first opportunities of RPA lie for many companies. At the same time, RPA teams I’ve spoken with are pondering the question “what next”.

In real life, processes have judgment-based decision points, they deal with unstructured data, have exceptions, and they evolve. These complexities are the kryptonite of RPA.

Superman pic. You should really see this.
Photo by Esteban Lopez

There is a limit how far you can get with rule-based logic, until hitting a brick wall of impossibility and realising you created something so challenging to maintain that it counters the benefits of automation.

AI solves this, right? Bring in the scientists!

Yeah sort of, but we are not going far if all talk is about OCR and document understanding, which seems to steal the majority of the attention with the RPA platform providers AI initiatives. There is so much more. For example, validating and fixing input data, matching datasets between systems that lack shared identifiers, detect anomalies to prevent errors and automate repetitive or “cognitive” decisions made by humans with attended bots.

What do the RPA leaders say?

While my sample of discussions with RPA leaders is hardly statistically representative, I think it is a fair assumption to say there is a problem. Here are three quotes, generalised to protect the identity of individuals.

We have had this automation on our RPA backlog for more than a year, and we tried something but the machine learning part didn’t come together. — RPA team lead

It’s really difficult to fit the data science project into our agile 3 week bot development sprints. — RPA team lead

Our data science team is like a black hole. You put an idea in, and after six months you have a data pipeline built. — CEO

Machine learning is undeniably needed when RPA program grows in complexity. Yet, the common theme here is that data science doesn’t meet the needs of RPA. Availability is minimal, tooling drives towards lengthy projects and introducing more technical specialists adds communication challenges. The mindset of data scientists drives towards exploratory projects, whereas RPA engineers need quick solutions that move items on the backlog. As a result, costs stack up too high for many RPA use cases.

There is a 10–100x gap in “AI to RPA fit” at the moment.

While I believe data science as a skillset is very much needed and there will “always” be a place for that, I also think there are so much more can be done without bringing the science into every project. Don’t let this mismatch ruin your automation project.

Solutions are out there already.

I got very excited when in the recent Sofa Summit RPA Conf 2020 Ericsson’s Director Automation & AI Transformation Kanda Kumar presented their thinking of “extend the capacity with democratisation”. In practice, there are two ways to get from the status quo towards the full potential:

  1. Augmented data scientists with tools like AutoML and MLops platforms, to help them achieve more.
  2. Domain experts equipped with the tools that allow them to get to their goal without spinning off a separate project for “ML parts”.
My manual reproduction of Kanda’s slide from Sofa Summit on May 20th 2020
My manual reproduction of Kanda’s slide from Sofa Summit on May 20th 2020

I believe there is a massive opportunity in especially the latter. A domain expert having the capability of solving the problem as a part of his/her daily workflow is an undeniable benefit. Apart from allowing cost-effective development of more complex automation workflows, it also enables agile development that is free of the constraints of an external team, their ways of working and backlogs.

Instead of throwing your automation project into the data science black hole and crossing your fingers, see what you can get done yourself first. You’ll be amazed. Efficient and easy to use tools are already available. Here is my list of six features to consider.

  • UX serves the target audience: great API for developers and integrations to RPA platforms, or a no-code experience for citizens developers.
  • It takes away the burden of data wrangling and feature engineering from the user.
  • Works without the user explicitly selecting an algorithm, and especially optimising it.
  • Is clear and transparent about prediction accuracy, allowing confidence to determine the next bot action.
  • Scales past initial dataset: automate new data ingestion and updated predictions in production.
  • Comes at a cost that doesn’t ruin the RPA budget.

Aito ticks the above requirements. What is your go-to tool fulfilling some or all of the above requirements? Or how would you improve the list? Continue the discussion at our community Slack!


This post was first published at Towards Data Science by Tommi Holmgren, a Chief Product Officer at Aito.ai. This re-post has been slightly modified from the original.

Back to blog list

New integration! Aito Instant Predictions app is now available from Airtable Marketplace.