3 Lessons Learned in 2024 & 7 Industrial AI Predictions for 2025

3 Lessons Learned in 2024

1. Amongst clear (and understandable) AI fatigue, there are rough diamonds being polished all over
2. User Experience (UX) still remains an afterthought, resulting in great technical PoCs that just don’t find users
3. Real-world scaled digital programs remain few and far apart. Still.

7 Industrial AI Predictions for 2025

1. The gap between AI Hype and AI Fatigue will reach its peak, with true value thereafter

2025 will be remembered as the year when the first truly operational, truly transformative—and above all—real customer-validated Gen AI solutions stole the thunder from previous demoware.

Just like most disruptive technologies, the early Gen AI excitement of 2023—supercharged by abundant promising demos (not to be confused with production-ready solutions!)—resulted in mainstream hype in 2024. Yes, this is very much referring to what Gartner calls the Peak of Inflated Expectations. With this came also AI Fatigue: the general tiredness of everything being AI. Everything. And Everyone.

As we start 2025, hype and demos are giving away to more grounded timelines and reality-conscious product architectures. Through the fog of AI for AI’s sake noise, singular examples of truly transformative embedding of Gen AI into actual workflows will stand out:

2. From systems of record to systems of engagement

2025 will be remembered as the year of data fabrics. Done well, these are true systems of engagement for cross-data source insights across one or multiple data domains, offering intuitive search, data exploration, and application development interfaces. Above all, they offer simple access to business users, not only IT.

As technology investors know well, data has gravity. This means the tendency of large data sets to attract other data, applications, and services. It's similar to how a planet's gravity pulls objects towards it, with the more data that accumulates, the more it attracts.

This is why, over decades, software solution providers have all aspired to become the system of record for something: customer data, inventory data, product data, financial data, time series data, any data—and most recently, all cloud data. In economic terms, the resulting lock-in from being the ‘master copy’ is perceived as more predictable and safer than actual customer value creation.

In an era marked by a proliferation of AI-powered solutions and agents interacting with data, accessing the right data, regardless of where it's located or in what format it's stored, is disrupting conventional wisdom around data gravity being tied to systems of record. Data gravity is, in fact, shifting to systems of engagement.

A system of engagement can help businesses unify data across multiple touchpoints, which can help create more actionable insights. A system of engagement differs from a system of record, which is a centralized data storage and retrieval system. Systems of engagement are designed for employees to interact with in their daily work lives, with value growing with data connections and context provided, rather than being the master copy of a particular data set.

3. From centralizing data storage to AI-enabling open data models

In 2025, the focus will shift from using raw data to power AI to building strong data models that make AI trustworthy and reliable. It’s not just about having lots of data anymore—it’s about organizing and structuring that data to ensure that AI delivers dependable solutions, even in complex environments.

At the heart of a data fabric—a system of engagement that is—is its ability to model data, resulting in one or multiple high-performance (low query latency, high parallel query support) data models that represent data in business-understandable language. To simplify, data modeling serves as the bridge between raw data and actionable insights.

The diversity of data types in the industry—from sensor readings to maintenance logs—requires optimized handling that only sophisticated data modeling can provide. It is in data models that efficient search capabilities and reliable data provision for AI agents equally happen—or fail to happen. By standardizing data formats and relationships through precise data modeling, we can reduce latency and improve the performance of AI-driven tools and agents, yielding interactive-level user experiences with data that speak the business language.

While most organizations have already implemented a datalake (i.e., centralizing data storage in the cloud) by 2025, it is only through data models that these become valuable to AI and human users.

4. From edge-to-cloud to operations-to-enterprise

2025 will be remembered as the year the edge-to-cloud Industrial IoT (IIoT) Platform category finally sunset, being replaced by vertical cloud platforms spanning — and more importantly, connecting — operations-to-enterprise for data and business solutions while leaving edge to specialist partners.

Let’s face it, no one—no customer and no solution provider, that is—has seen success with delivery on the Industrial IoT (IIoT) Platform approach as defined since 2018 (the inaugural release of Gartner's Industrial IoT Platform Magic Quadrant report).

Perhaps because of insurmountable product development scope (solution provider perspective), perhaps because of lack of alignment on the customer side on deployment and scaling (customer perspective), perhaps because of out-of-sync velocities at which edge and cloud markets have developed and keep developing, perhaps because of data gravity favoring convergence between operations and enterprise over edge and cloud both in operations convergence, perhaps all and more.

However, the overwhelming conclusion is that there is a shift in the market away from edge-to-cloud (IIoT Platform) to operations-to-enterprise (let's call this Digital Industrial Platform, as some are starting to refer to it), which forward-thinking CDOs and digital transformation leaders at industrial enterprises should add to their radars.

5. From bigger LLMs to specialized tools for agents

2025 is the year of Agents and Agentic AI. More so, 2025 is the year agentic tools will march into the AI solution limelight."



"By the end of 2025, buyers of Agentic solutions will not ask which or how many models are supported but which and how many agentic tools are provided.

For some time, all we had to do to make our dream AI solution work was wait for OpenAI, Anthropic, Meta, or a similar foundation model provider to deliver the model that would finally make it possible. After all, LLMs seemed to improve exponentially in capabilities, correlating more or less to their training parameters—or so we wanted to believe.

By the end of 2024, we know better. Bigger LLMs will not turn LLMs into non-LLMs or anything more than LLMs. Large language models are great at many things — just like we humans with our human brains are—but they are simply not the universal solution. Enough hope on LLMs; time to take destiny into our own hands. The answer lies with agentic tools.

Not one large tool, but many tools. Simple and more advanced, general and domain-specific. In fact, by the end of 2025, buyers of agentic solutions will not be asking which or how many models are supported but which and how many agentic tools are provided.

To complete the analogy with LLMs + agent tools and the human brain + tools we’ve invented (such as the laptop I’m using to write this predictions blog), our advance as a species did not stop because our brain stopped growing larger (and thus more capable). In fact, recent research shows that the average human brain size has slightly decreased over the past 10,000 years. As humans, our capabilities have developed exponentially over the past 10,000 years without more grey matter, we’ve simply learned to augment ourselves with tools. The same path lies ahead on agentic AI.

6. From omnipotent copilot mirages to 100s of nimble agents in production

2025 will see the first industrial customers with 100+ agents in production, supported by dozens of agent tools to make them work. Many agents will be invisible to the workflow user.

With AI development already shifted in 2024 from building impressive (but not actually useful) demos to solving actual needs that are small in scope but actually useful—as well as useful as building blocks of multi-agent solutions—it’s a great time to allow universally capable Copilots to fade away gracefully as retired first-wave artifacts of Gen AI introduction.

AI needs to be brought to where users already are, into their workflows, that is, not trying to move workflows into a Copilot interface unsuited to the end user’s needs. This will happen by introducing agents that are invoked from the context of a business workflow, making the workflow itself more efficient—and having contextual awareness of the workflow by virtue of the context of invoking, rather than the user needing to prompt the agent with everything.

It may even be that the best agents are those users hardly see explicitly but whose outputs they benefit from in their regular business workflows.

7. From proven agents to more complex levels of multi-agent orchestration

At the end of 2025, agent orchestration frameworks will allow for rule-based triggering and sequencing of agents with agent tools, providing a glimpse of what the new AI ecosystem will look like by 2028.

The agentic orchestration layer is the latest entrant to the contemporary AI stack. With early maturity in agents and agentic tools in 2025, ambitious leaders will forge their own frontrunner paths on multi-agent orchestration as a route to a disruptive new application development paradigm.

Depending on appetite (and budget) for innovation, there will be an even greater bifurcation in the market between those partnering with solution providers to co-innovate at the cutting edge of technology and those comfortably awaiting for the winning solutions to emerge. Expect interesting discussions on IP ownership, as well as more strategic partnering activity.