Does AI change the value of IFC data?
Short answer
For the balance sheet, hardly at all. In theory AI could rebuild an IFC model faster, which would lower its replacement cost. In practice the effect is small: AI can only rebuild what it already knows, accounting rules change slowly, and the amounts involved are modest.
The bigger gain from AI is in maintenance and operations, and it depends on whether the data underneath can be trusted.
How IFC data is valued
The value of an IFC model is based on the time it would take to redo it. That means rebuilding the data itself. The engineering thinking and design decisions behind the model are left out of the valuation.
This replacement-cost approach follows the accounting rules for intangible assets (IAS 38 internationally, RJ in the Netherlands, GASB 51 for US public bodies). Those rules decide what an asset owner may recognise and capitalise, and how high the carrying amount can be.
Why faster recreation by AI doesn't change the number
If AI can recreate a model in a fraction of the time, the replacement cost drops, and so does the value. That is partly true, but three things limit the effect.
- AI can only rebuild what it already knows. To recreate a model it needs the geometry, properties and relationships from somewhere, and for a real asset that information usually exists only in the model you are trying to value. So the shortcut stays theoretical for now.
- Accounting rules move slowly. The standards that set the carrying amount are not rewritten every time a new tool appears, so for the foreseeable future the official balance sheet value will be calculated the way it is today.
- The amounts are small. The balance sheet value of model data is always lower than people in the industry expect. Most feel their data is worth far more, but the accounting rules keep the number modest, and a few hundred euros or dollars more or less on the balance sheet makes no real difference.
It also helps to separate recreating a model from making it. Rebuilding a model when the end result is already known goes much faster than originating it, when every design choice is still open. The replacement cost already measures the faster of the two. An AI that speeds up the rebuild a little more shaves hours off a number that was modest to begin with.
Where the real value sits
The value of IFC data comes from how you use it once the asset is built. In maintenance and operations, owners use model data to plan inspections, schedule replacements, order parts and budget for the coming decades. That applies to buildings as much as to bridges, tunnels and roads, where the data can stay in use for up to 50 years.
AI can help with all of this. It can search the model for components due for maintenance, predict when parts will fail, or answer a facility manager's question from the data. The value that creates can be many times the number on the balance sheet.
AI in operations needs reliable data
AI bases its decisions on the data it is given. If a door is modelled as a wall, a pump has no type or a fire damper sits on the wrong storey, the AI will plan around those errors without noticing them.
Before an owner hands IFC data to an AI system, they need to know which parts of it can be relied on and which cannot.
Example: Waternet, Amsterdam
Waternet is Amsterdam's water utility. It supplies drinking water to about 1.4 million people, manages 5,119 km of sewers and 7 treatment plants, and looks after around 1,000 km² of waterways. It has explained in public how it is preparing its asset data for AI.
Waternet started where many asset owners are. Its data sat in separate systems (GIS, the Maximo maintenance system, BIM models, document management), some of it locked in proprietary applications, with missing and duplicate attributes and no single source of truth.
Waternet gave the data structure first. It built its own object type library and a knowledge graph on open standards, including IFC 4.3, NEN 2660 and GWSW, aiming for a fully digital asset information model by 2030. The AI uses are built on that structure: language models that answer questions from the graph without inventing facts, predictive maintenance that combines sensor readings with asset history, and leak and risk detection.
Waternet makes two further points. The reason data is missing matters more than the missing data itself. And drawings from the 1850s are still more useful to Waternet than some digital models made 25 years ago, because the old drawings are accurate.
forensicBIM examined one of Waternet's models, a critical pump station. The result is published as an example report with the client's permission:
| Pump station model | Amount | Compared with carrying amount |
|---|---|---|
| Carrying amount, if recognised today | €2,615 | 100% |
| Value in use today | €63k to €157k | 2,409% to 6,004% (24 to 60 times) |
| Savings potential in O&M over 20 years | €275k | 10,516% (105 times) |
The balance sheet value is less than 1% (0.95%) of what the data could save in maintenance and operations over 20 years. If AI shortened the rebuild by a few hours, the first number would move by a few hundred euros at most. How reliable the data is decides how much of the last number Waternet actually gets.
What makes IFC data reliable enough for AI
An AI system can only use what the model states clearly. Before relying on it, an owner should be able to answer these questions about each model:
| Question | What can go wrong |
|---|---|
| Is the file valid IFC, with author, organisation and date filled in? | Nobody knows where the data came from or how old it is. |
| Is every element in the right place in the spatial structure? | The AI cannot tell which storey, building section or bridge span a component belongs to. |
| Are elements modelled with the right IFC class? | A pump stored as a generic proxy without a classification is invisible to anyone searching for pumps. |
| Are elements classified, ideally with a URI to a shared dictionary? | The same component gets different names in different files, so queries miss half of them. |
| Do properties carry real values with a defined meaning? | Fields read “default”, “n/a” or are empty, and the AI fills the gap with a guess. |
| Is the geometry complete and consistent? | Quantities, locations and clashes come out wrong. |
| Is the data consistent across the file? | Duplicate IDs or negative quantities break the link to the maintenance system. |
The IFC version matters less than the content. IFC2x3 can be as useful as IFC4 or 4.3, although it cannot store full georeferencing.
Why forensicBIM matters more with AI
forensicBIM examines IFC models and reports how useful the data is: what is there, how complete and consistent it is, and what it is worth. Once owners put AI to work on their asset data, that check becomes a condition for using it safely.
The balance sheet value will stay roughly where it is. The gains from AI in maintenance and operations are far larger, as the Waternet figures show, and owners only get them when the data has been checked first.
What asset owners can do now
- Recognise the model data on the balance sheet at its replacement cost, and use that carrying amount to budget for keeping the data up to date. Don't wait for AI to change the accounting; it won't any time soon.
- Ask for IFC. Data in a proprietary format can lose all its value once the software stops supporting it, and it costs licence fees as long as it is supported. IFC stays readable, which matters for data that has to last decades.
- Have the models audited before an AI system works with them. Knowing which parts are reliable, and which are missing or wrong, tells you where AI can be trusted and where it will guess.
- Fix the gaps that matter most for maintenance and operations first, rather than everything at once.
- Keep checking. A model that is not maintained goes stale, and so does every AI answer based on it.