Industrial AI starts with what you already have

AI has never been more accessible. The tools work, examples exist in almost every sector, and vendors are ready to help any organization go further on their AI journey. Despite this, industrial companies often find themselves stuck anyway. The technology is mature enough, however, there are simply too many options and too many reasons to wait.

Getting started with AI in industry is one challenge, building something durable with it is another. The first takes courage, the second takes discipline. Both draw on something industry already has in abundance but rarely uses well: data and operational knowledge accumulated over decades. A large processing plant typically runs on tens of thousands of tagged parameters recorded at second-level frequency, alongside maintenance records, ERP data, and years of written notes. The raw material is already there but not used to its full potential.

Starting with what you already have

The most common reason industrial organizations give for postponing AI acceleration is that the foundations are not ready. Data is messy. Systems are old. Processes are undocumented and often vary from shift to shift, with each team working from its own beliefs about what does and does not work. All of this is true and rarely resolves on its own. Waiting for perfect foundations has historically meant waiting forever.

What has changed is that language and multimodal models can now work with production and service notes, handwritten logs, manuals, and text fields that were never designed to be analyzed by machines. Maintenance reports describing the same event in many different ways can now be classified consistently, helping surface patterns that were previously hidden in the noise. The unstructured layer of an industrial company, long considered too messy to use, is now within reach as a source of insight, context, and untapped potential.

A good first project fixes more data quality problems than any abstract governance initiative ever will.

Starting is less about getting the foundations right and more about choosing one real use case with enough operational weight that the data and process questions become concrete.

Finding the right use case

Two approaches tend to work better than open-ended brainstorming. The first is to work from known archetypes of industrial AI and ask which of them actually fit your production and commercial reality:

• Soft-sensors and quality or grade prediction from upstream process data
• Condition monitoring and anomaly detection on critical equipment
• Throughput, yield, and energy optimization at set-point or schedule level
• Natural-language copilots for plant operators, engineers, and maintenance teams
• Production planning connecting demand forecasting, pricing, and capacity constraints

The second is to walk the plant section by section with production, maintenance, operations, electrical, sales, and management, asking each group what their real challenges are in that part of the process. The conversation stays grounded in operational problems rather than drifting into technology, and it surfaces issues that different departments perceive in different ways. Both approaches shift discovery from what the organization could build to what actually matters.

Discipline that respects what industry already knows

Once early experiments deliver results, the pressure flips, with ideas multiplying, vendors arriving with new angles, and teams exploring in parallel. Momentum builds, which is healthy, but only up to a point.

Scattered momentum is not the same as progress.

Industrial leaders have been through this pattern before. Advanced process control, MES, IoT, predictive maintenance, digital twins. Each wave had its hype and its winners, and each taught industry something about separating the two. That instinct is often under-leveraged in the broader AI conversation, which leans heavily on technology-first perspectives and on vendors without deep operational experience.

Direction in practice looks more like a compass than a detailed map. A small number of priority areas tied to real operational or commercial value, clear boundaries around what is out of scope for now, and enough stability that teams can act without escalating every choice. Saying ”not yet” to an interesting idea is how the priorities that actually matter get the room to develop.

The capability is tacit knowledge

Individual AI projects can deliver value, but only some organizations turn a series of such projects into a durable capability. In industrial settings, operational knowledge lives in people. That has always been true, but with AI it is suddenly actionable.

Examples of such tacit knowledge are various: Operators recognize problems by sound or vibration long before instruments raise an alarm. Metallurgists read flotation cells the way experienced clinicians read patients. Maintenance engineers know which equipment lies, and under what conditions. Planners carry mental models of plant dynamics that are nowhere in the SOPs.

What AI needs most is precisely what industry has captured least.

Building AI capability in industry is therefore partly a technical exercise and partly an organizational one. Here are a few practical ways to uncover what people already know:

• Record recurring meetings and feed the transcripts to AI. Production planning, shift handovers, maintenance reviews, and quality meetings contain most of what actually drives day-to-day operations. Transcripts of these meetings, made searchable and analyzable, turn routine conversation into an organizational memory.
• Instrument the decisions, not just the outcomes. Most plant systems capture what happened. Far fewer capture why an operator intervened, which option was considered and rejected, or what the on-shift judgment call was. Small additions to existing workflows, such as short, structured notes at key decision points, make reasoning visible over time.
• Make existing unstructured content usable. Shift logs, service reports, incident write-ups, and free-text comments are often already there. Giving AI tools access to this content, and extracting structure or patterns from it, turns years of accumulated notes into something the organization can actually learn from.
• Run structured knowledge conversations with experienced people. Targeted interviews with operators, engineers, and planners (recorded and transcribed) capture tacit judgment before it leaves with retirement, in a format that can be directly used by AI systems.

Companies that take this as seriously as they take model development tend to pull ahead. The tacit layer is very hard to copy from the outside, which makes it a real source of competitive advantage in industrial AI.

Forward energy

Industrial leaders are used to making decisions under uncertainty, and AI is not a special case. The companies making the most real progress are rarely the ones with the most sophisticated technology or the largest data science teams. They are the ones where leadership pairs the courage to start with the discipline to hold direction, and where everything the organization already has and knows is treated as the strategic asset it is. That combination is hard to buy and hard to outsource. It is also something industry has been quietly building for decades.

Jussi Järvinen is the CEO of Brillian and one of its founders, with over 25 years building digital services for industry, including in demanding leadership roles at Metso and Outotec. Drawing on these experiences, he brings insight into how digital intelligence can be used to create value in business and management, highlighting what this requires in practice—from people, organizations, and leadership.

Brillian is a leading digital intelligence company working with companies that want to make AI a real part of how their business operates. We combine strategic advisory with full technical delivery, building exceptional products and operations in the AI-native world.