From scattered data to a carbon assessment: how AI speeds up building analysis

Anyone who wants to calculate the CO₂ emissions of a building must first know what they are dealing with. How big is the building? What materials were used? How are exterior walls and other building components constructed? What building services is present? How much electricity and heat does the building consume? For new buildings, much of this information is already generated digitally today. For existing buildings, it often looks different. Data is spread across building plans, area calculations, documentation on renovations, energy performance certificates, and consumption data. Some documents are structured, others are PDFs, scans, or in non-digitized form. In addition, there are different qualities and designations.

This is precisely what complicates not only individual carbon assessments. Real estate companies, developers, or municipalities often have to evaluate entire portfolios with different building types. Opening each document individually, extracting relevant values, and then transferring them to a carbon assessment tool takes a lot of time.

Hitzler Ingenieure and elevait are therefore working together on an approach that aims to automate part of this work with artificial intelligence. Hitzler Ingenieure contributes experience from project management and sustainability consulting, among other areas. Founded in 2021, elevait develops AI solutions that read, structure, and provide information from documents for further processes. Together, both companies are working on the ai:cm solution.

Documents must first become data

A carbon assessment begins not with calculation, but with data search. For a holistic assessment, information such as location, year of construction and use, areas, components and materials, technical systems, and energy demand or actual consumption is required. Depending on the building, this information can be contained in very different documents.

The AI approach therefore starts one step earlier. Users first upload existing documents. The system unpacks the files, checks them for duplicates, and classifies the documents. Among other things, it distinguishes plans from other documents and categorizes information according to DIN 276. The software then searches the documents for the data it needs for the CO₂ calculation. For plans, it reads text both via OCR and from existing PDF layers, groups related text components, and extracts metadata from them. In doing so, the AI takes over a work step that previously required a lot of manual effort: it searches heterogeneous existing documents for relevant information and brings it into a structure that can then be used for further work.

A Munich residential project as a practical test

Elevait and Hitzler Ingenieure tested how well this works on the Stanigplatz residential development in Munich. The project in the Hasenbergl district comprises 49 apartments and a centre providing services for older people and was completed in 2021. For the test, a carbon assessment according to the DGNB method was to be created. The participants proceeded in two ways.

The AI first searched the available documents for the required key figures and used them to create a carbon assessment. In parallel, Hitzler Ingenieure's sustainability team calculated the balance using conventional methods. For the manual calculation, the experts evaluated, among other things, data from the building permit, determined areas based on calculations and floor plans, and identified component areas from detailed construction drawings. Since no complete bill of materials for the building services was available, they included an allowance for these systems. They extracted energy key figures from the energy performance certificate and consumption data from 2022. Both teams then compared the results and gradually adjusted the AI approach.

20 minutes instead of 40 hours

The comparison is initially significant. The AI calculated a total value of 20.43 kilograms of CO₂ equivalents per square meter of net floor area per year for the building. This took approximately 20 minutes. Hitzler Ingenieure's manual calculation resulted in 18.36 kilograms of CO₂ equivalents per square meter of net floor area per year. According to project data, this took approximately 2,400 minutes, or 40 hours.

There is about an eleven percent difference between the two results. However, this difference does not tell  the whole story yet. The calculations also differed methodologically. Especially in the building construction, different types of area calculation led to deviations, particularly for the exterior walls. This also changed the allowance for building services and the overall result. Furthermore, in the AI balance, the replacement of building components was not yet included at that development stage. The experts from Hitzler Ingenieure therefore manually checked the AI result. The project partners explicitly considered this review useful because it created certainty and showed where the calculation needed further improvement.

When is an approximation sufficient?

The test thus shows less that AI can fully take over expert planning. A different question is more interesting: How accurate does a result actually need to be for the respective application? For a certification or a final assessment, reliable and verifiable values are required. For the early assessment of a large building stock, however, a good approximation can already help to identify buildings that warrant closer attention and set priorities.

This is precisely where the potential of the approach lies. The AI is intended to enable a carbon assessment even when the data situation is not yet complete. Missing information can partly be supplemented with secondary data. At the portfolio level, the software could evaluate many buildings with significantly less manual effort and then make the results comparable via key figures, filters, and dashboards. This would change the scope of application. Instead of only undertaking detailed carbon assessments of individual buildings, owners could first analyze large portfolios and then look more closely where particularly high emissions or significant renovation potentials are evident.

AI is within the range of other methods in construction

A further plausibility check is provided by comparing the value for building construction with other calculation methods. The AI here came to 7.54 kilograms of CO₂ equivalents per square meter of net floor area per year. Comparative values ranged between 6.32 and 12.13 kilograms depending on the method. In calculations according to DGNB and QNG, the AI result was within or close to the values determined there.

Thus, the test shows, at least for this building, that the automated calculation was not outside the order of magnitude of established methods. However, a general statement about the accuracy of the system cannot yet be derived from this. The test refers to a specific project. The project partners also see further development needs and want to further optimize, among other things, the key figures and assumptions used.

The greater leverage lies in the data structure

The carbon assessment is just one possible application. If AI reliably recognizes and structures information from unstructured documents, the same data can generally be used for other tasks. The project partners see potential for variant comparisons, assessing the reusability of building components, pollutant information, and climate protection roadmaps, among other things. The decisive step is therefore not just faster calculation. First, a usable data basis is created from scattered documents. This can become particularly important for existing portfolios. Many companies have been aware of the problem not just since the increasing demands for carbon assessments: they have large amounts of information but can only evaluate it with considerable effort. AI can close this gap by reliably finding, classifying, and providing information for further calculations.

Automate, but do not adopt unchecked

The experiment at Stanigplatz also shows the limits of the approach. Completing a calculation in 20 minutes rather than 40 hours sounds impressive. But both values describe different working methods, and the calculation results were not identical. Experts had to analyze and explain the differences. Especially in complex sustainability assessments, it therefore remains crucial that experts can understand what data the AI uses, what assumptions it makes, and why results differ from each other.

The strength of AI lies primarily where a lot of time is lost today: in sifting, sorting, and evaluating large quantities of documents. If it reliably handles this preparatory work, experts can spend more of their time evaluating the results. The test at Stanigplatz provides a concrete example of this. However, it also clearly shows: automation does not replace expert review. It can provide a significantly better starting point for it.