Chemicals & Materials

From Pilot to Scale: What Separates AI-Materials Programmes That Progress From Those That Stall

Most organisations running AI in materials research have a pilot that works. Fewer have moved that pilot into everyday use. Closing that distance is the harder task, and it depends more on organisation than on technology.

The distance between a working pilot and a scaled capability is where much of the anticipated value is realised or lost.

Ask a materials or chemicals company today whether it is using AI in research, and the answer is almost always yes. Ask how much of that work has moved beyond a pilot into routine use, and the picture becomes more modest.

This pattern is common across the industry. It is worth understanding, because the gap between a successful pilot and a scaled capability is where much of the anticipated value is either realised or lost.

The Pilot-to-Production Gap

The pattern is not unique to materials. Studies of enterprise AI more broadly describe a persistent distance between the number of use cases organisations pilot and the number they operate in production.

The consistent finding across that research is that the barriers are mostly organisational rather than technical. In most cases the model itself works; the harder part is the work around it, namely data readiness, integration with existing systems, clear ownership, and skills.

In most cases the model itself works. The harder part is the work around it.

Why Materials Makes the Gap Wider

Materials research adds its own difficulties on top of the general ones. The data that AI models depend on is often sparse, inconsistent, and spread across decades of experiments that were never recorded with machine use in mind. Laboratory infrastructure is frequently older than the software being layered onto it. And the expertise required sits at the meeting point of two scarce disciplines, materials science and data science, which relatively few people hold together.

A further issue is specific to the field. Much of the promise of AI in materials rests on what researchers call the inverse problem: rather than predicting the properties of a known material, designing a material to meet a desired set of properties. This is genuinely valuable and genuinely hard, and the tools for it remain early.

A pilot that demonstrates property prediction on a clean dataset can look convincing without yet being ready for the messier, higher-value work of guiding real development decisions.

Three Structural Constraints Specific to Materials

Data

Decades of experiments never recorded for machine use

Sparse, inconsistent, and distributed across systems and formats that predate any expectation of model training.

Infrastructure

Laboratories older than the software layered onto them

Instrument estates and workflows that were not designed to emit machine-readable output at the point of measurement.

Skills

Two scarce disciplines, rarely held together

The work sits at the meeting point of materials science and data science, and relatively few people hold both.

None of these are model-selection problems, and none are resolved by a further pilot.

What Appears to Help Programmes Progress

Looking across the organisations that have moved AI from pilot to routine use, a few common factors stand out. These factors are organisational rather than technical.

Four Factors That Separate Progress From Stall

FACTOR 01

A defined problem rather than a general ambition

Programmes that begin with a specific, measurable question — a particular formulation to improve, a particular property to predict — progress further than those that begin with a wish to adopt AI.

FACTOR 02

Data treated as infrastructure

Organisations that scale tend to have made their experimental data consistent, accessible, and connected before expecting models to draw value from it.

FACTOR 03

Clear ownership

A pilot can succeed as a side project. Routine use needs someone accountable for it, with a place in the R&D workflow rather than alongside it.

FACTOR 04

Realistic sequencing

Leaders who separate what can be deployed now from what needs longer capability building tend to spend more effectively than those who treat every opportunity as immediate.

Matters of discipline and organisation rather than of algorithm selection.

The factors that separate programmes that progress from those that stall are matters of discipline and organisation rather than of algorithm selection. In most cases the technology is already capable enough to be useful; whether it becomes useful depends on conditions the organisation controls.

What This Means for Investment

For a company deciding how to invest in AI for materials, the choice of tools is only one part of the decision. The greater determinant of outcomes is often where AI is applied, how success is defined, and whether the organisation is prepared to integrate new capabilities into existing ways of working.

Before committing significant investment, it is worth considering a few principles that have consistently emerged from our work with leading chemicals, materials, and OEM organisations navigating these decisions.

Six Principles Before Committing Investment

01

Prioritise problems with clear business valueRather than broad AI ambitions.
02

Focus first where data quality and workflow integration can support routine useNot where the demonstration is easiest.
03

Distinguish near-term deployment from longer-term capability buildingAnd fund them on different horizons.
04

Define how AI-generated insights enter decisionsSpecifically, how they are incorporated into R&D and product development.
05

Build internally where it creates strategic advantageAnd use partnerships where they accelerate progress.
06

Establish clear ownership before expecting pilots to scaleAccountability precedes routine use.

The specifics will vary from one organisation to another. What appears consistent, however, is that programmes which progress beyond the pilot stage tend to be supported by clear priorities, defined ownership, and a practical understanding of where AI can contribute most effectively within the innovation process.