Failed pilots and poor results are all too common when it comes to many of the past year’s AI efforts across the world. Most struggle because of a rush to jump on the bandwagon without the tedious job of building up a digital foundation.
Bucking the trend, Canadian software firm OpenText started mapping out the entire company’s job processes just under two years ago to bolster its AI efforts. This meant understanding the job functions of each of its 21,000 employees by building up a sort of “AI genome”.
“We mapped our entire organisation, function by function, job by job… to see what humans can do and what agentic AI can do,” said Shannon Bell, OpenText’s chief information officer (CIO).
Looking through the standard processes and mapping out the rules needed, the company found many processes it could automate – usually, 80 per cent were common across enterprises while 20 per cent were unique to an industry, she told reporters in Singapore during a customer event earlier this month.
OpenText, which helps organisations manage their data for AI and other uses, ended up mapping 1,700 job descriptions among its employees.
This enabled it to build an AI model to understand where AI agents can be deployed, with precisely what inputs and outputs they needed to deliver the outcomes desired.
In the end, its most compelling use cases included basic “Level 1” helpdesk tasks, which can be automated, say, to address calls with similar requests. Code assistance for developers is another big use for agentic AI that’s taken off of late.
Such “classically low hanging fruit” may not realise the promise of AI that many had expected but they offer a good starting point to build and scale up, said Bell.
“There is a mismatch with the outcomes (expected) and the length of time it takes, she noted, adding that many CIOs are under pressure to deploy some AI – any AI – because they are under pressure from the board or shareholders.
Companies that begin their AI efforts with the most complex legacy processes will run into challenges very quickly, often derailing their early efforts, she noted.
These pilots fail because they often don’t break down a business process – something that takes a lot of human effort, she said. “Edges cases” are another issue, she added, because an AI agent has to deal with a lot of exceptions to the rule to get the right answer.
OpenText built the foundations early because it understood the effort required to get things right, she added. “We invested in our data, curating data for our AI processes; we understood our business processes and started small.”
“Some use case for agents are so simplistic… where’s the real value? We saw each one as a building block. You see a little more gain and it’s over a period of time.”
To be sure, no organisation is going to wait for a “Big Bang” transformation before kickstarting their AI efforts. Yet, OpenText’s experience shows that incremental increases can come more easily with a more patient foundation building than a scattershot approach to AI.
Another big factor is people. The team at OpenText started with colleagues who were the most open to the technology – the network operations centre staff and site reliability engineers (SREs), for example, used AI to reduce the number of major incidents by 16 per cent.
“They don’t want to be on incident bridges (calls); they want to drive innovation,” said Bell.
Another key factor is ensuring that there’s a person bringing their expertise into the equation, especially in more complex and high-value tasks.
“We firmly believe in having a human in the loop,” said Bell. “Can I use agentic AI to run test cases and assist in code? Yes, task-based functions that are repetitive in nature should be the first pieces for AI.”
“However, the higher you go up value chain, you absolutely need domain knowledge and expertise,” she stressed.
