Should you let employees build their own AI bot?

AI has democratised technology once exclusive to the IT team, but how far can such flexibility reach

Featured faculty: Prasanna Karhade
Written by Putro Harnowo

Earlier this year, an X post from a Meta engineer went viral after her artificial intelligence (AI) bot ran amok, deleted emails, and wouldn’t stop, even after being ordered to do so. The engineer was testing the agentic AI that can execute tasks on its own. This AI agent promises autonomous decision-making to solve complex problems without human supervision, but the accident may prove it’s not entirely safe.

While businesses ponder the pros and cons of agentic AI, a safer choice called intelligent process automation (IPA) has been around for years. IPA bots combine AI technologies with robotic process automation to create AI bots. Their strength lies in low-code/no-code (LCNC) toolkits that enable users without coding skills to automate specific tasks, making them more secure and reliable in corporate environments.

Microsoft Power Automate is an example. It can perform repetitive tasks like sending emails or generating reports seamlessly within the Microsoft ecosystem. Software maker SAP also offers Build Process Automation to automate tasks integrated with other SAP products, and IBM’s Cloud Pak for Business Automation caters to larger enterprises seeking cloud-based automation. Many other tech firms provide independent IPA bots for specific needs.

However, despite hefty investments in AI solutions, many companies still struggle to deploy IPA bots widely across their workplaces. Prasanna Karhade, Associate Professor in the Department of Decisions, Operations and Technology at the Chinese University of Hong Kong (CUHK) Business School, argues that companies are still in the dark about how to implement AI solutions effectively.

“AI technologies are rapidly growing, but companies are sometimes stuck with the traditional ways of governing technology solely in the hands of the IT team,” he says. “Employees without coding skills can now leverage IPA bots to create and refine their own automated workflows, bolstering the democratisation of technology development. However, this innovation must also align with company IT policies.”

Professor Karhade notes that this new dynamic has forced businesses to revisit whether centralised technology management remains relevant and, if not, what the best strategies are to roll out AI tools that can be widely accepted across the organisation.

When AI tools are failing, and why

In a study titled, AI governance and the decentralisation of technology production: An investigation of AI-based IPA bots, Professor Karhade and his co-authors examine 176 IPA projects at a Fortune 200 US multinational IT services firm, particularly in the banking, financial services, and insurance domains, to identify the critical factors in AI tool adoption.

“Companies naturally want all employees to use the AI tools, so they demand high utilisation. Apart from that, the AI solutions must be repeatable to ensure broader application within the company beyond the initial use,” says Professor Karhade. “Therefore, utilisation and repeatability are the two key factors in AI governance.”

Having investigated 24 highly underutilised and 54 highly unrepeatable IPA projects, Professor Karhade and the team find that more than 83 per cent of these failed projects are imposed by top management without employee input. Conversely, among the 46 and 35 IPA projects classified as highly utilised and repeatable, respectively, employees are responsible for over 91 per cent of them. These successful projects have low coding intensity, meaning employees apply their own knowledge to deploy AI bots using the LCNC toolkits.

A bottom-up approach turns out to significantly improve the utilisation and repeatability of IPA projects, while a top-down approach results in the opposite. This finding provides a robust starting point for computational experiments to further identify other influential factors in determining the success of AI solutions.

“AI users will be more empowered as they can do many things on their own, but a centralised IT team still plays a crucial role in enabling and overseeing AI infrastructure.”

Professor Prasanna Karhade

The elements of success

The key to the success of an AI solution lies in how it starts, or what the researchers call the “genesis”. IPA projects initiated by employees yield successful adoption, but further analyses find that how the project is deployed, the amount of coding skills required, and the complexity of the tools also contribute significantly.

IPA bots are highly used when deployed by users using the LCNC toolkits, so they require very little coding. Process intricacy, or how many steps an IPA bot has to do to complete a task, is also crucial. Intricate processes are harder to automate, but when users leverage their knowledge to create bots with the LCNC toolkits, the bots are highly likely to be used.

Deployment, or how AI tools work and interact with users, can be divided into three types: attended, where users trigger the process; unattended or fully automated; and hybrid, where AI works alone but sometimes needs human help. Users are still likely to use attended bots in a top-down manner, but they’re not widely accepted, as shown below.

AI bots may solve the problem at hand, but they are not necessarily repurposed widely. This is where repeatability matters. To create highly accepted bots for a wider user base, unattended AI bots are particularly preferable as they can handle processes autonomously.

Although an AI solution is all about democratising technology, when users implement an AI bot primarily for their own specific needs, the bot becomes too specialised to be reused by others. This contradicts the idea of decentralised technology, or, as Professor Karhade calls it, the limits of democratisation. When this happens, IT support is needed to ensure that the AI tools are reliable and flexible enough to be used across the company.

The evolving roles of the IT team

AI may have democratised technology adoption, but Professor Karhade underlines that the IT team is irreplaceable. “AI users will be more empowered as they can do many things on their own, but a centralised IT team still plays a crucial role in enabling and overseeing AI infrastructure.”

For companies, management should remain receptive to user-initiated AI projects and provide guardrails to enable seamless collaboration between employees and IT teams. This balanced approach will enable businesses to harness the benefits of decentralised technology while safeguarding operational integrity. Otherwise, an accident similar to what happened with Meta’s engineer could happen.

Professor Karhade believes the findings apply to broader contexts and industries dealing with large volumes of documents, such as retail, logistics, healthcare, and human resources. The core ideas revolve around how companies should rethink their technology management in the AI era and build a sustainable ecosystem that enables diverse employees to contribute to AI solutions.