Where enterprise tech goes next: With Microsoft and guest from Forrester

But how do we turn triple-A ambitions into reality? We got Mike Richter, Microsoft’s Principal Solution Architect, and Bernhard Schaffrik, Forrester’s Principal Analyst, to help. In a forward-thinking talk for Face Value, they discussed everything from knowing what you do best before getting on the AI train, to the demand for real insight into how people work if new tools are to be effective.

• Read our recap below, or watch on-demand: Anything is possible: where enterprise tech goes next with Microsoft and Forrester

Garbage in, garbage out: AI demands clean data

“Garbage in, garbage out.” It’s more than just a tech mantra, it’s a reality of making AI work for the enterprise. Just like sound human decision-making is nourished by varied and accurate understanding, AI craves a diet that’s organized, representative, and comprehensive. AKA “clean” data.

As Richter emphasizes, a weak data foundation may be the most common reason firms are behind on AI. No matter how sophisticated an AI system is, it’s ultimately a trained model. When fed garbage, it’s bound to produce low-quality outcomes that frustrate users and waste resources.

Some 89% of business leaders recently told a survey AI’s effectiveness hinges on being fed data that reflects how their business actually runs. For Schaffrik, a “clean, modernized data estate” will usher in the next frontier of AI models. “It has the potential to enable agents to traverse multiple systems and speak with each other,” he says. “As a result, their reasoning capabilities will be greatly enhanced.” While many of the agents we’re seeing today “don’t deserve the label,” interconnectivity and advanced reasoning capabilities will equip them with much-needed “arms and legs.”

“Data modernization, data cleanup is really becoming a priority,” echoes Richter. “Obviously AI is going to be a competitive advantage if you can implement it — but only if you’ve got good data.”

Crossing the automation efficiency threshold

Like AI, automation is fixed atop the tech agenda, but Schaffrik believes automation-mature enterprises have one thing in common: they’ve hit a ceiling of complexity at which robotic process automation (RPA) ceases to provide incremental value.

While the push to automate more complex processes has built up the promise of low-code and digital process automation (DPA), these advanced tools require firms to grasp what can go wrong — the exceptions, errors, and deviations from the happy path. As Schaffrik observes: “Many people started to give up. They just automated the happy path.”

Reaching the next threshold of complexity requires firms to look inward and assess “how work is really getting done” — i.e. how their processes actually unfold. “We are now understanding that we are living in these complex spaghetti processes nobody has a clue about,” says Schaffrik. “We need a deeper understanding of domain knowledge, not just what’s written down in standard operating procedures (SOPs) or work instructions.”

This is why he sees a slew of software automation vendors adding capabilities to their platforms like process mining and task mining. The breakthrough lies in bridging the gaps of “what had been laid down, written down, implemented in systems” and “how people really get the work done.”

• Read: What is process mining?

SLMs are the next AI golden child

While enterprises and consumers alike have been caught up in the promise of large language models (LLMs), Richter notes the rise of small language models (SLMs). Trained on smaller datasets than their larger cousins, SLMs can adapt to more specialized domains and bring AI features and functionality to new corners — like Internet of Things (IoT) devices. “With SLMs,” Richter explains, “you’ll be able to attach these models to your refrigerator and ask, ‘what do I need to get from the supermarket?’ That model doesn’t need to know quantum physics, it just needs to know how to identify things in a refrigerator.”

Schaffrik also adds that SLMs present a valuable opportunity for organizations to trim down energy consumption and costs. After all, small models are designed to be efficient, focusing on fewer things but doing them exceptionally well. “There will be a combination of models,” notes Schaffrik. “One for finding the right model for the right solution, but then also ones that we can run cost-effectively that will provide as much value as an LLM.”