Scaling industrial AI through digital workers

Somya Kapoor, CEO of IFS Loops, speaks about the expanding portfolio of digital workers, the role of multi-skilled AI agents in enterprise workflows, governance and orchestration, and the measurable efficiencies customers are achieving through real-world deployments.

IFS has expanded its portfolio of digital workers. Is the current lineup comprehensive, or can we expect further expansion?

You’re going to see more. What we showed on stage (at IFS Unleashed) was roughly 20 digital workers across different domains, supported by 150 agentic skills.

Each digital worker is a multi-orchestrated agent with multiple agentic skills designed to deliver an outcome or complete a task. I don’t think you’re going to see 100 digital workers, although you might see another 50 next year. What you’re going to see is a lot more agentic skills that customers can pick and choose from to create digital workers that drive specific outcomes. All 20 digital workers are multi-skilled. They’re not single-task agents.

Are these digital workers designed to operate horizontally across industries, or are they becoming more industry-specific?

We’ve mapped them across the major processes that IFS is known for: asset management, field service, finance, and procurement. Projects is coming in Q4, which will complete those core workflows. From there, the digital workers are becoming increasingly industry-specific as well.

With Agent Studio enabling customers to build their own agents, how does IFS guide the development process and prevent agent sprawl or duplication?

Now that we’ve released Agent Studio, we have recommendations built into the process. When you go through process discovery, Agent Studio and the meta-prompter, and identify something you want to build, the system can tell you that there’s already an agent available for that purpose. It asks whether you’d like to use that existing skill or create a new one.

We also have a certification process in place for deploying agents in marketplaces, where customers can build their own digital workers. We’re going to refine that moving forward to ensure there’s no unnecessary duplication. Of course, you can’t completely control what customers do within their own environments, but we seed a lot of recommendations directly into Agent Studio to help guide those decisions.

Can IFS also help customers validate whether the agents they’ve developed are meeting their intended objectives?

Meeting the objectives is really about the outcomes that customers want to drive.

There’s a dashboard in the system where you can track things like how many purchase orders have been processed, how many parts have been replenished and other operational activities. All of that information is fed into the system, so customers can track the value they’re getting from each agent as they deploy it.

Can these agents be modified while they’re running if customers find that they aren’t performing as efficiently as expected?

You can provide feedback and test the agents before deploying them. Once they’re in production, you can’t make changes directly. You can always bring an agent back into your development workspace, make enhancements, test it and then relaunch it into production.

Who typically takes ownership of custom agent development within customer organisations? Is it primarily IT or the business teams?

It depends on the organisation. In the mid-market, we’re seeing mainly IT teams taking ownership. In large enterprises, IT is more focused on governance, while the process teams are actually building these agents.

IFS has also introduced Industrial AI Harness as a governance and orchestration framework. What capabilities does it bring together?

Industrial AI Harness brings together security, governance, planning, memory management, context and model switching. It’s essentially a framework that provides everything you need to support and orchestrate the LLM.

How does model switching work within Industrial AI Harness, and what impact does switching between different LLMs have on operational outcomes?

You can bring your own model, so you can switch the LLM. In our environment, if you’re using the out-of-the-box capabilities, we provide the orchestration of the LLM. But if a customer says they want everything to run on OpenAI, Claude or open-source models, that’s absolutely possible. In that case, they provide their own GPUs and the infrastructure required.

We offer the capability to orchestrate different tasks across different models.

You can switch models, but the outcomes and accuracy will be different, of course.

Looking at deployments over the past year, which use cases have delivered the strongest results for customers?

What we’ve targeted most broadly is manual data entry into ERP systems. For example, converting service requests into work orders, purchase orders into downstream work orders, invoices into work orders and similar processes. Essentially, wherever there’s manual entry into the ERP system, that’s where we’re deploying agents.

Then there are additive workflows, such as material replenishment, inventory replenishment and service closure management. These are activities that add value further downstream in the operational process. We’ve seen some very strong efficiency gains from these deployments.

Can you share some of the measurable efficiency improvements customers are achieving?

KLN Family Brands is seeing a 60% efficiency gain, while Kodiak is seeing 87%. We’ve also seen Ependion achieve 82%. These are real production deployments. We run webinars with customers only when they’re in production. We don’t put out customer success stories until the solutions are actually deployed and delivering results.

With AI expected to unlock significant untapped workforce potential, how far can these efficiency improvements realistically go? Is 100% achievable?

You can aim for 100%, but the question is whether any LLM can be 100% accurate. I don’t think you’ll ever reach 100% accuracy. But if you have that 1% error, can you bring a human into the loop to handle it? Yes, absolutely.

What best practices would you recommend to enterprises developing and deploying their own custom AI agents?

I’ve actually published a lessons-learned white paper on this. The first recommendation is to pick a small process with high impact and low risk. That’s how you start building trust within the organisation. Treat an agent as an intern coming into your environment before it becomes a full-time employee. Think about the escalation path, security and guardrails before you even start building or using an agent.

You also need to be very clear about the outcome you want to drive. What are the wins you’re looking to achieve? And then keep improving it. It’s something that continues learning within your environment.

Are digital workers also helping improve collaboration between frontline employees, operational teams and management?

They are, actually. Some of our customers, like Kitron, have said that deploying digital workers brought their teams together. They’re now collaborating more across the organisation.

We’ve seen something similar with Pedro Buhigas, the Chief Information Officer (CIO) of Kodiak Gas Service. He brought his entire business unit together, and they came up with 25 use cases. He then let the team decide which use case they wanted to deploy. They selected one and started with that.

This approach brings the organisation together because people are involved in identifying the opportunities and understanding the value of deploying the technology.

How well do IFS digital workers integrate with third-party enterprise applications, including systems from competing vendors?

I don’t think we need to look at it purely from a competition perspective. What we’re looking at is how to extend the workflow.

For example, if documents are stored in SharePoint, how do I connect to that? If customer data is in a CRM system, how do I connect to it? If the finance system is somewhere else, how do I connect to that?

It’s not about whether the system belongs to a competitor. It’s about where the data resides and how we can extend the workflow through agentic AI. And yes, we’re connecting to Maximo and other systems as well.

What is the most important message you would like enterprises to take away from the progress IFS is making with industrial AI and digital workers?

This technology is very real. It’s not something that’s just being talked about. But it does require change management, and that’s a significant cost element.

You have to consider how you’re going to upskill your people, what process changes you need to make and where the data gaps exist. The sooner you start deploying the technology, the better, because you can begin that change management process within your organisation. If your competitors start doing this before you, they’re going to have an advantage.