Heimstaden cuts the cost of virtual apartment staging by more than 99%

Property management company Heimstaden offers a rare and useful example of corporate AI adoption: not "we gave everyone a chatbot," but a breakdown of specific processes with clear costs before and after. The company uses ChatGPT and Codex for preparing listings, estimating rental prices, validating data and analyzing climate risks.

The most striking metric came from the Netherlands. Previously, virtually staging a photo of an empty apartment cost roughly €50. After generative tools were introduced, the cost of one image, according to Heimstaden's estimate, fell below €0.10, a drop of more than 99%. Potential annual savings in that market alone are estimated at €130,000.

But the cheap image is not the main point. The company is rebuilding the entire journey of a vacant apartment: the system helps suggest a price, prepare images and a description, check the published listing against internal standards and collect market feedback. An employee still approves the result and is accountable for the decision.

Not a chatbot, but a connected process

Imagine a vacated apartment. The manager must estimate the rent, order photos, write a listing, check its quality and then work out why the unit is renting slowly. If only the description text is automated, the gain is small: the other delays don't go away.

Heimstaden started from the process as a whole. Tools based on ChatGPT were created for employees, a central team coordinates the changes, and data from property management systems is gradually being prepared for machine processing. This is the boring part of adoption, but it is what separates a working system from an impressive demo.

The approach works beyond real estate. A product team should pick one repeatable customer journey, measure its cost and duration, find the manual handoffs between systems and only then connect a model. The metric should not be the volume of AI requests but the reduction in time, errors or cost per completed operation.

When a specialist builds the tool herself

Heimstaden's second case is even more interesting to developers. Sustainability analyst Julie Erlandsen needed to assess the climate risks of the portfolio. The source information was spread across different open sources, formats and datasets. Normally such a project would require a dedicated engineering team or consultants.

Using Codex, the analyst built a repeatable pipeline for searching, cleaning, standardizing and validating data. As a result, 20,112 properties were assessed and more than six million rows of useful data were produced. By the company's internal estimate, the workload shrank from an anticipated five people for a year to two people for two months.

This does not mean the model carried out reliable research on its own. The methodology and verification remained with the domain specialist. The practical scheme looks like this: the expert sets rules and control examples, AI helps write and connect the processing steps, and the human checks samples, exceptions and final calculations.

What the case has not yet proven

The material was published by OpenAI and is essentially a customer story, so the figures should be treated as Heimstaden's own data, not the result of an independent audit. Costs of licenses, integration, employee training, quality control and support of the built solutions are not disclosed. There are also no figures on errors in price recommendations or on the effect of automation on time-to-rent.

Nevertheless, the case provides a workable framework. Start not with choosing a model but with a process map and a baseline metric. Leave humans the right to approve financial and customer decisions. For data pipelines, add automated checks, a provenance log for sources and test sets with known results.

The main takeaway is simple: noticeable returns appear when AI gets access to prepared data and is embedded between the stages of real work. A corporate chat on its own saves minutes. A rebuilt process can save months, but only if the company measures quality and doesn't hand the model responsibility it cannot bear.