I have interviewed several startups that develop AI tools for the AEC sector. They argue that the industry creates data for human use and judgment, so you can’t count on it being systematized. Hence, LLMs are the way to go, as they can handle messy data.

However, experience shows that where data is structured and machine-readable, AI performs. Where it is not, LLMs specifically either fail or produce plausible but incorrect outputs.

This is not just a question of using AI as a productivity booster within a single company. It is about supply chain-wide use of data from early concepts to building operations, whether AI is involved or not. Is it realistic to assume that AEC firms will start paying attention to data quality and invest in improving it? According to Solibri‘s MEP webinar last Thursday (May 7), it is possible and surprisingly easy when the frameworks and tools are in place.

The effort is not what people think

At the webinar, Granlund’s Markus Järvenpää explained and demonstrated the use of Finland’s national data standard for MEP objects, system types classification, and property sets. It was developed during the RAVA2Pro and RAVA3Pro projects and has been added to buildingSMART Data Dictionary (bSDD) platform (in Finnish).

The Finnish standard covers the full scope of MEP: over 800 product type identifiers across HVAC, electrical, and building automation, plus system type classifications and data field definitions for each. It is publicly available and has been implemented in the major MEP design applications used in Finland.

The workflow in an application is a single action: select the product type from a dropdown built into the software (Markus used MagiCAD for Revit), and all required property sets export automatically with the IFC. There is no manual property entry, no separate data management step, and no additional coordination required. In a 10,000-hour project, Markus estimated the total added effort at approximately one hour!

Granlund has also released a free Solibri extension that checks IFC models against the standard, closing the loop from classification to verification without any additional cost or tooling investment.

What structured data unlocks

Structured MEP data means that every object in the model has agreed semantics, classifications, and property sets that a machine can read and compute without guessing. Markus illustrated this in Solibri Office.

When MEP objects are consistently classified, software can automatically calculate quantities, connect them to cost databases, simulate alternatives, and update estimates dynamically. The same structured foundation enables carbon calculations, logistics planning by installation zone, and procurement based on actual model data rather than summary estimates.

Markus made a pointed observation at the webinar: duplicate objects in unstructured IFC exports are common and are rarely detected. If you start procurement from a model containing duplicates, you may order sixteen electrical distribution boards when you need far fewer.

All of this applies whether or not AI is in the workflow. But as AI tools become more embedded in construction processes, the gap between structured and unstructured data widens further.

Some argue that LLMs are now sufficiently capable of extracting meaning from messy data, making structure unnecessary. There is partial truth in this; AI is genuinely good at converting unstructured information into structured information. But that does not make the structured output unnecessary afterward. The more AI agents automate procurement, logistics, and lifecycle management, the more valuable stable, machine-readable data becomes underneath them.

The commercial barrier is the real problem

Given that the tools are ready, the standard is published, and the effort for a designer is minimal, the obvious question is why structured MEP data remains the exception. The answer is commercial. Clients do not order it, so designers do not produce it. Without a contract requirement, the effort that enables everything downstream simply does not happen.

Markus noted at the webinar that he has seen project coordinators propose adopting the MEP standard mid-project, only to be told by the MEP designer that it was not in the contract. This is a collective action problem. Every party in the value chain benefits from structured data, but the cost falls on the designer, and the benefit is captured by contractors, owners, and facility managers who did not pay for it. Until clients specify structured MEP data in design contracts, the incentive structure will not change.

Finland introduced BIM-based building permitting this year, but the regulatory framework does not yet mandate structured MEP data, which would create a clear legal anchor for contract requirements. That may change. In the meantime, the practical path forward is straightforward: clients specify MEP compliance in design briefs, and designers who understand the internal benefits start applying it regardless.

What needs to happen

Designers can act without being asked. Markus made this point explicitly: the internal benefits in native MEP software are immediate. Consistent classification makes the model more usable within the design team before any IFC export happens. The designer who builds the habit now is also the designer whose models will be AI-ready when clients begin to demand it.

The systemic change requires clients. A simple contract clause referencing the standard is enough to shift the incentive. Clients who specify it will receive models from which quantities, costs, and carbon data can be extracted automatically. Those who do not will continue to receive models that require manual interpretation at every handover. As AI tools become more capable and more embedded in construction workflows, the gap between these two positions will widen.

Success with AI in construction depends not primarily on the technology chosen, but on an organization’s ability to connect AI to its own processes, data, and strategic thinking. Structured MEP data is precisely that connection point for the built asset. It is not a prerequisite for experimenting with AI; it is a prerequisite for AI delivering durable value.