Your AI Is Smart. It Still Doesn't Know the Industry.

3 November 2026 · By Oleg Gavrilov

Why boring is an opportunity

Industrial components rarely make headlines.

O-rings are a multibillion-dollar market. These small seals prevent leaks in machines, pumps, valves and pipes, often working for years without anyone noticing.

The industries that rely on these small parts have a few things in common that make them worth a closer look when building AI tools:

Knowledge lives in people's heads. The best selection rules are held by experienced engineers and sales staff, many of whom are nearing retirement.
Data lives in documents. Catalogues, datasheets, certificates and approval lists, mostly PDFs, many of them scans.
Mistakes are costly, with delayed consequences. A wrong choice may only show up months later, as a leak, a stoppage or a failed audit.

None of this is glamorous. That may explain why so little good software exists here. The gap is real, and so is the opportunity.

What we found when we looked closely

We started with the less exciting job of reading manufacturers' catalogues, datasheets and approval documents.

We went through them line by line, noting every value along with its source and test conditions. Here is some of what turned up. I've left out brand names because this affects the whole industry, not one company.

A catalogue that contradicts itself. In one recent catalogue, the approvals table lists two drinking-water approvals for a compound. A few pages earlier, the same catalogue's compound table credits it with drinking-water approvals in four countries. An older brochure from the same company lists five more.
Same compound, different numbers. A manufacturer's flyer states that a food-grade compound has a hardness of 70 Shore A. It also shows a compression set of 35% after 24 hours. Their newer brochure gives 74 Shore A and 32% after 70 hours. Same product code.
Limits that outrun the tests. A datasheet shows a minimum service temperature of −50 °C. However, the low-temperature test only goes down to −42 °C. That may be fine for a static seal, but the reader cannot tell how the number was set.
Numbers that can't be compared. Compression set is one of the most quoted properties of a seal. One datasheet shows a lower value after 72 hours than after 24 hours. That's simply because the longer test ran 20 °C cooler. Without time and temperature, the number on its own means little.

None of this comes from carelessness. It comes from documents created over many years by various teams for different markets.

But it means that anyone, human or machine, who reads these documents at face value will sometimes be confidently wrong.

Why a language model alone is dangerous here

Ask a general-purpose chatbot which O-ring material to use for hot drinking water, and it may suggest EPDM. That can be a useful starting point.

But EPDM is a family of materials, not a complete specification. You still need to identify the manufacturer's specific grade and check that its drinking-water approval covers your application and temperature and is still current. A material recommendation is not the same as a verified selection.

In many fields, plausible is good enough. In industrial sealing, plausible and correct are not the same. The gap between them can lead to leaks, recalls or failed inspections later on. A model that is confidently wrong does more harm than good. It takes away the doubt that might make someone double-check.

How we're approaching it

Early on, we decided that at Seallect, AI is the interface to verified facts, not the source of them. In practice that means a few simple rules:

Every value has a source. Each hardness, temperature limit or approval points to its original document, noting the edition and date.
Every value keeps its conditions. Compression set is stored with its time and temperature, and each low-temperature limit with the method behind it. Values tested differently are never compared as if they were the same.
Claims are not certificates. A tick in a brochure is recorded as a claim. It becomes "verified" only when we see the certificate, its scope and its expiry date, or find it in a public register.
We flag conflicts rather than average them. When two documents disagree, we keep both. We mark the conflict and ask the manufacturer for clarification, instead of just choosing one.
If it doesn't know, it asks. If a key condition is missing, like temperature, media or approval, the system will ask a question instead of making a guess.

This is slower than letting a model read everything and summarise. But it is the only way we know to build something an engineer can actually rely on.

Lessons for anyone building AI for traditional industries

I'm a sealing specialist, not an AI researcher. But a few things have become clear to me, and I suspect they apply to many "boring" sectors:

Domain knowledge beats model choice. Understanding that a compression set number means little without its test conditions matters more than which model reads the document. Partner with people who know where the problems hide.
Data provenance is the product. In regulated or safety-relevant fields, "where did this number come from?" is the first question a serious customer asks. If you can't answer it, nothing else matters.
"I don't know" is a feature. Users in these industries forgive a tool that asks a question. They don't forgive one that is confidently wrong.
Start narrow. We began with O-rings only, not "all seals". A narrow scope lets you go deep enough to be genuinely useful.
The mess is the moat. The work of reading, checking and reconciling technical documents page by page is hard to copy. That is exactly what makes it valuable.

Where we are

Seallect is at an early stage. We're building the dataset one manufacturer at a time, starting with O-rings, and talking to distributors and engineers about how they choose seals today.

If you work in AI and you're curious about applying it to traditional industry, or if you've built something similar in another "boring" sector, I'd love to compare notes. And if you work with seals every day and recognise the problems above, I'd especially like to hear from you.

The next wave of AI won't only happen in chat windows and code editors. A lot of it will happen in pump rooms, workshops and spare-parts warehouses.

It just needs someone to read the fine print first.

If you’d like to follow our progress or join the early access list, you can find us at seallect.com