A stress diagram cannot be scraped. A hardness indentation cannot be prompted. That is the entire moat.
For three years the enterprise AI conversation has been about the model. Bigger, better, cheaper. That conversation is finished. Every serious player has access to roughly the same frontier models. Switching between them is a configuration change. The question that matters in 2026 is the one Gartner put on the table. By 2028, eighty percent of the software tools used to build AI applications will be built around context engineering, not prompting. The model is the commodity. The context is the moat.
If you run IT or engineering at an industrial manufacturer, that has a specific consequence. The companies that win the next five years are not the ones with the cleverest prompts. They are the ones that pair AI expertise with deep industrial expertise. Three layers, three different kinds of skill, one stack. The test machine that produces the raw measurement. The lab data system that gives it structure. The conversational workflow agent that drives the next step. None of these layers can be replaced by a model.
Three reasons this is the new shape of the moat.
First, tokens are interchangeable, context is not. The same foundation models are available to every competitor in your category. What changes the answer is what the model knows about your specific lab, your specific parts, your specific failure modes, your spec sheets going back twenty years. McKinsey’s QuantumBlack team put it plainly this year. The durable advantage from AI is organizational intelligence and domain expertise that lives in people and operations. General purpose models do not carry that. They never will.
Second, the physical world is the hardest context to fake. You cannot scrape a stress diagram from the open web. You cannot prompt a hardness indentation into existence. The same is true for a fatigue cycle, a torque curve, a vibration trace, a thermal profile. These are signals generated at the machine by industrial OEMs that have spent decades engineering what the signal means. IBM estimates around ninety percent of the data produced by sensors and telemetry is never used. The data exists. It sits on a controller, in a log file, in a PDF report nobody opened. The reason most AI projects in industrial labs stall is not the model. It is that the model has no relationship with the floor.
Third, structure plus workflow turns the signal into a decision. A raw test result without sample lineage, method, operator, batch, and spec is noise. The lab data layer turns that noise into a queryable graph. The conversational workflow agent on top runs the loop end to end. Ask the question, get the answer, trigger the next step, write back to the system of record. This is the part that gets missed in most “AI inside LIMS” announcements. A chatbot inside one system answers questions about what is inside that one system. A workflow agent that spans the machine, the lab data, the ERP, and the engineering archive is a different category. Gartner has said where this goes. By 2028, most enterprises will abandon assistive AI for outcome focused workflow. The chatbot phase is ending. The agent phase is starting.
Two contrarian reads worth taking seriously.
The model camp argues that we are one or two generations from a frontier model so capable that the stack work I just described becomes irrelevant. Just wait. Two things make me doubt this. Scaling gains are flattening, and switching models is the cheapest line item in any AI program. And no model however large can read a stress diagram it has never been connected to, on a machine it has never been wired into, in a lab whose taxonomy it has never been taught. The physical floor is not a token problem. It is a wiring problem.
The lab data camp argues that many LIMS vendors are embedding AI chatbots in their 2026 releases, so why bring a third party in. The honest answer is that one chatbot inside one system is not the moat. The moat is the cross system stack. Machine plus lab data plus workflow agent plus the rest of the enterprise. That is three or four kinds of expertise no single vendor has, and no single model can substitute for.
This is the model Squirro has been building toward. We are not the LIMS. We are not the test machine vendor. The agent framework, the knowledge graph layer, the privacy controls, the on prem deployment. All of it exists because the moat sits in the connection between AI and the physical world. We work with industrial manufacturers like Bühler on agent workflows for manufacturing. We run knowledge systems for industrial groups like Henkel across more than forty five data sources. The same logic applies in the lab. When the machine vendor, the lab data vendor, and the workflow agent each bring their own expertise to the same stack, the result is something none of them could ship alone.
Three questions to ask before the next budget cycle. Where does the data from each test machine actually live, and how much of it never reaches a system of record. Which system of record is the source of truth, and what is missing from it. What workflow would you want to run across all three layers, if the conversation could span them. The answers are the shape of your moat for the next five years.
The companies that come out ahead are not the ones with the biggest model. They are the ones who pair AI expertise with the people who built the machines, the people who structured the data, and the people who know what a good answer looks like on the lab floor. That combination is hard. That is exactly why it lasts.
If you run lab IT or engineering at an industrial manufacturer, take thirty minutes with Squirro on what the three layer stack looks like for your specific test data. dorian@squirro.com.
