Why Most AI Pilots in Chemicals Never Become Production Systems

Why Most AI Pilots in Chemicals Never Become Production Systems

Why Most AI Pilots in Chemicals Never Become Production Systems

Justyne Smagacz

0 mins

Share

Most AI pilots do not fail dramatically.

They work just well enough to create excitement.

A small team runs a controlled test. The data set is limited. The users are engaged. The implementation team is close enough to explain, correct, and guide the system when needed. The demo works. The answers are promising. The business case looks possible.

Then the organisation tries to scale it.

That is usually where the real problem appears.

The issue is rarely that the AI model was incapable. Foundation models are now remarkably powerful. The harder question is whether the organisation has built the conditions required for AI to be trusted, maintained, governed, and used at production scale.

A prototype proves that AI can work in a narrow environment.

A production system has to work inside the real organisation.

Those are not the same test.

What pilot success actually measures

A successful pilot usually measures whether a system can execute core operational tasks and deliver performance under favourable conditions.

The data set is chosen carefully. The scope is manageable. The users are often early adopters. The use cases are usually representative, but not exhaustive. There is a project team watching the results closely and filling gaps manually when needed.

That is valuable. It tells you something.

But it does not tell you whether the same system can support thousands of products, multiple divisions, conflicting document versions, regional variations, legacy naming conventions, different permission levels, and users who have neither the time nor the patience to experiment.

In specialty chemicals, this gap becomes visible quickly.

A global product range is rarely clean and uniform. The same product may appear under different names across markets. Technical Data Sheets may vary by region. Safety Data Sheets may be current in one country and outdated in another. Product attributes may be inconsistent, incomplete, duplicated, or organised around historical business processes rather than AI usability.

A pilot can often avoid that complexity.

Production cannot.

This is why “we connected AI to our documents” is not the same as building an AI system that a sales team, technical service team, product manager, or distributor can rely on every day.

The model may be strong. The interface may be easy to use. The first answers may look convincing.

But production AI depends on what sits around the model: structured product data, clear taxonomies, document version control, expert feedback loops, data ownership, permissions, integrations, and governance.

Without those foundations, the pilot eventually reaches a ceiling.

It’s not worth to spend a lot of effort on data without application use cases already in use. The right consulting partner will support you by meeting you wherever you are on your data clean-up journey, and will help fill the gaps to support the scale to production. 

The first ceiling: data architecture

Many organisations underestimate how much architecture is required before AI can reason usefully over chemical product data.

The instinct is understandable. If the company already has documents, product records, brochures, TDSs, SDSs, sales decks, and application notes, it feels logical to connect an AI layer and ask questions.

But AI does not magically resolve years of inconsistent structure.

If product attributes are duplicated, incomplete, or named differently across business units, the AI inherits that confusion. If SKUs, base codes, grades, regional names, and application segments are not mapped properly, the system may retrieve information, but it will struggle to understand the relationships between products.

That matters because chemical queries are rarely simple.

A user is not always asking, “What is the viscosity of this product?”

They may be asking, “Which product would you recommend for this substrate, in this region, with this curing profile, under these regulatory constraints, and what is the closest alternative if the preferred product is unavailable?”

That question requires more than document retrieval.

It requires structured chemical intelligence. A combination of product data, expert knowledge, trust governance, and workflows that transform general-purpose AI into a reliable enterprise production system.

A general-purpose AI tool may be excellent at language, but it will not design the product data architecture for a specialty chemical company. It will not decide which attributes matter across adhesives, coatings, personal care, construction chemicals, or electronics. It will not resolve the operational reality that one business unit may classify a product by chemistry, while another classifies it by market, application, or sales channel.

That architectural work has to happen deliberately.

In one implementation pattern we have seen repeatedly, hundreds of product properties exist across legacy systems, but only a smaller, standardised set can actually support scalable AI use. Reducing that complexity is not a cosmetic clean-up exercise. It is the work that allows AI to understand the business in a way users can trust.

The second ceiling: knowledge completeness

Even with good structure, most AI systems are only as useful as the knowledge they are given.

This is where many pilots overestimate what formal documents can do.

Technical Data Sheets are important, but they are not the full truth of how products are selected, sold, substituted, or recommended. They are often designed as customer-facing documents. They may contain core technical properties, but they rarely capture the full decision-making process of an experienced application engineer or technical sales manager.

Safety Data Sheets add another layer, but they also introduce complexity: country versions, language versions, regulatory updates, and the need to know whether a document is still current.

Then come brochures, catalogues, sales training materials, application guides, case studies, competitor comparisons, troubleshooting notes, and market-specific positioning.

And above all of that sits institutional knowledge.

That is the layer most organisations need most, and the layer they usually have documented least.

It is the product manager who knows why one grade performs better in a specific edge case. The technical service lead who knows which alternative works when a customer’s preferred chemistry is unavailable. The regional sales expert who knows that a product is technically suitable but commercially unrealistic in a particular market. The application engineer who can answer in 20 seconds what it would take someone else six months of testing to discover.

That knowledge does not usually live in a neat folder.

It lives in people.

This creates a practical problem. You cannot ask an expert to sit down and write everything they know. It is too broad, too time-consuming, and too unnatural.

But people are very willing to correct something.

That is a critical design principle for production AI.

The system should not only answer questions. It should expose gaps, capture corrections, and turn expert feedback into reusable organisational knowledge. Every “that answer is missing a caveat” or “I would recommend a different product here” becomes part of the company’s intelligence layer.

A static knowledge base gets old.

A production AI system should get better through use.

The third ceiling: trust

The most important question in enterprise AI is not whether the system can generate an answer.

It is whether people trust the answer enough to use it.

Trust is built through many small mechanisms.

Can users see where the answer came from? Can they identify whether the source document is current? Can a subject matter expert correct the response? Can the organisation control who is allowed to edit, approve, generate, or publish information? Can the system handle contradictions between documents? Can it flag uncertainty rather than pretending to know?

This is where production requirements become very different from prototype requirements.

In a pilot, uncertainty can be managed informally. Someone on the project team knows the data set. Someone can explain why an answer appeared. Someone can manually correct the gap before the next demo.

At scale, that informal support disappears.

A distributor may be using the system while trying to respond to a customer quickly. A new sales hire may be relying on it to understand a product range. A technical service team may be using it to compare options across regions. A product manager may need to know whether an outdated document is still being surfaced.

In those scenarios, trust cannot depend on goodwill.

It has to be designed into the system.

That means version management. Permission levels. Data ownership. Review workflows. Error handling. Expert approvals. Clear escalation paths. Reliable integrations. Visibility into what users are asking and what answers they are receiving.

Without that, usage data can become misleading.

A team may appear to be using the AI tool, but the organisation may have no visibility into whether the answers are good, whether users trust them, or whether incorrect information is being repeated.

For industries built on technical credibility and customer relationships, that is not a minor issue. It is the difference between a productivity tool and a risk.

The fourth ceiling: adoption

Pilots usually attract champions.

Production includes everyone else.

That distinction matters.

The first users of an AI pilot are often curious, motivated, and forgiving. They want the system to work. They understand the context. They are willing to experiment with prompts, tolerate rough edges, and provide feedback.

A production rollout includes people who are busy, sceptical, resistant, or tired of new systems. Some will worry that AI threatens their expertise. Some will assume that if the first answer is not perfect, the system is not useful. Some will avoid logging in altogether unless they are trained, supported, and shown a relevant use case.

This is not a technology problem.

It is change management.

Good AI adoption requires strong executive sponsorship from the top and practical enablement from the ground. Leaders need to show that the system matters, that they trust the process, and that using it is part of how the organisation works. At the same time, teams need hands-on training, real examples, internal champions, and clear guidance on what good usage looks like.

Teams also need to see leadership driving the behavioural change. Visibility is needed of leaders using the new system themselves by opening the system in meetings, asking questions in real time, challenging questions and displaying use cases and enthusiasm. 

The strongest adoption moments often come from inside the organisation.

A vendor can explain value. But when a respected leader or sales manager says, “This saved me half a day,” or an application expert says, “This helped a new hire find the right answer faster,” the message lands differently.

This is especially important because AI is not replacing experts in these environments.

It is preserving and distributing their knowledge.

That distinction changes the conversation. The purpose is not to remove people from the process. It is to make their expertise available to more colleagues, in more markets, more consistently, before that expertise walks out of the organisation through retirement, turnover, or restructuring.

Production AI is a living system

The mistake many organisations make is treating AI like a software installation.

Connect the data. Launch the tool. Train the users. Done.

That model does not work.

Production AI is closer to a living knowledge system. It needs to be fed, corrected, governed, maintained, and improved. Every new document can strengthen it. Every outdated document can weaken it. Every expert correction can make future answers better. Every unanswered query can reveal a gap in the organisation’s knowledge.

This is why the final percentages of accuracy are so difficult.

Getting a system to answer basic questions across a small product set is not the same as supporting complex, multi-constraint queries across 17,000 products and 88,000 documents. It is not the same as identifying that one division may not have the answer, but another division does. It is not the same as connecting technical recommendations to commercial opportunities across countries, business units, and customer segments.

As the system improves, the questions become more ambitious.

The bar moves.

That is a good thing. It means users are beginning to trust the system with more valuable work. But it also means there is no final state where the AI is “finished.”

The goal is not perfection.

The goal is a system that continuously identifies what is missing, routes it to the right people, captures the correction, and improves the next answer.

In that sense, the best production AI systems become a kind of self-cleaning organisational brain: one that builds more connections over time, not because the model magically understands the business, but because the organisation has designed the workflows that allow knowledge to accumulate.

The real reason pilots do not scale

Most failed AI pilots are not really AI failures.

They are failures to understand the difference between a prototype and a production system.

A prototype asks, “Can this work?”

A production system asks much harder questions.

Is the data structured enough? Is it complete enough? Is it fresh enough? Who owns it? Who can change it? What happens when the AI is wrong? How do experts correct it? How are corrections reused? How do users learn to trust it? How does the system improve six months after launch? How does it handle the complexity of the real organisation?

Those questions are less exciting than a demo.

But they are where the real evaluation happens.

The companies that succeed with AI will not necessarily be the ones that choose the most impressive model. They will be the ones that build the strongest knowledge infrastructure around it.

In specialty chemicals, that means treating AI not as a chatbot project, but as a serious enterprise system: one that combines product data, technical documents, expert judgement, governance, and adoption into a single operating model.

The pilot is only the beginning.

Production is where the real work starts.

Share

Recent Posts

Company News

September 30, 2026

Kimia

AI & Technology

September 23, 2026

Anna Keeley

AI & Technology

August 27, 2026

Sadra Boutorabi

Company News

September 30, 2026

Kimia

AI & Technology

September 23, 2026

Anna Keeley

Solutions
Platform
About