Ask ten real estate
professionals how they work with AI and the answers scatter. Hardly anyone
claims mastery. Many firms are experimenting and few have made AI part of their
daily work. For one person using AI means writing a good prompt. Another uses
AI every day for research and document review. A third understands the risks
well but rarely touches the tools. A board member reads more and more papers
that AI helped produce. Which of them is AI literate? Probably all of them in
part and none of them completely.
In Europe the question has become concrete. Article 4 of the EU AI Act (Regulation (EU) 2024/1689) requires providers and deployers of AI systems to take measures on AI literacy for their staff and for others operating these systems on their behalf. The obligation has applied since 2 February 2025.
In July 2026 it changed. The Digital Omnibus on AI (Regulation (EU) 2026/1744) was published on 24 July 2026 and has been in force since 27 July. It rewrote Article 4. A duty to "ensure a sufficient level" of AI literacy became a duty to "support the development" of it. The obligation itself remains. What fell away is any requirement to guarantee a specific level of AI literacy for any individual. National market surveillance authorities started supervising and enforcing Article 4 on 2 August 2026.
In our view the softer wording changes less than it seems. Article 4 still points to the same factors: the technical knowledge, experience, education and training of the people involved and the context in which the systems are used. It prescribes no fixed course, no certificate and no single level of competence for everybody. That leaves companies with a better question than how many people completed an AI course. What should our people actually know?
In Europe the question has become concrete. Article 4 of the EU AI Act (Regulation (EU) 2024/1689) requires providers and deployers of AI systems to take measures on AI literacy for their staff and for others operating these systems on their behalf. The obligation has applied since 2 February 2025.
In July 2026 it changed. The Digital Omnibus on AI (Regulation (EU) 2026/1744) was published on 24 July 2026 and has been in force since 27 July. It rewrote Article 4. A duty to "ensure a sufficient level" of AI literacy became a duty to "support the development" of it. The obligation itself remains. What fell away is any requirement to guarantee a specific level of AI literacy for any individual. National market surveillance authorities started supervising and enforcing Article 4 on 2 August 2026.
In our view the softer wording changes less than it seems. Article 4 still points to the same factors: the technical knowledge, experience, education and training of the people involved and the context in which the systems are used. It prescribes no fixed course, no certificate and no single level of competence for everybody. That leaves companies with a better question than how many people completed an AI course. What should our people actually know?
Literacy is not the same as being good at prompting
When we ran our first generative AI trainings most of the attention went to prompting. Understandably so. Better instructions produced better results. But prompting is only one part of using AI professionally. An analyst may be excellent at getting a model to produce a market summary without knowing whether its sources are reliable. An asset manager may draft a strong tenant letter without understanding what happens to the confidential information they typed in. A senior manager may know little about prompting yet spot immediately when an AI generated analysis makes no commercial sense.
All three forms of knowledge matter in our experience. AI literacy is not a level to reach. It is a combination of skills and the right mix depends on the work someone does. The person comparing leases needs different knowledge from the person approving an investment paper. Someone using AI in recruitment needs different knowledge again. Giving all three the same two hour introduction creates a common starting point but not equal preparation for the decisions they make.
All three forms of knowledge matter in our experience. AI literacy is not a level to reach. It is a combination of skills and the right mix depends on the work someone does. The person comparing leases needs different knowledge from the person approving an investment paper. Someone using AI in recruitment needs different knowledge again. Giving all three the same two hour introduction creates a common starting point but not equal preparation for the decisions they make.
Use it. Understand it. Judge it.
The first is being able to use AI. Give the system the information it needs. Describe the task clearly rather than assuming the tool will guess the context. Choose a suitable tool for the task and ask for output in a format the next step can use. Work efficiently instead of correcting the same result ten times. Poor use makes good technology look bad.
The second is understanding enough about what is happening. Nobody needs the mathematics behind a large language model any more than the code behind their accounting software. But some concepts matter. A confident answer can still be wrong. Information supplied to a model is not the same as information it retrieves and both are different from what the model asserts from memory. More context can help but does not automatically improve an answer. How company data is handled can differ widely between consumer tools and the systems a firm has approved. None of this requires a technical background. It requires knowing which questions to ask before trusting an answer.
The third is being able to judge it. That means holding outputs to the standards of the job rather than the fluency of the text. We find this the hardest to teach and the most important.
The second is understanding enough about what is happening. Nobody needs the mathematics behind a large language model any more than the code behind their accounting software. But some concepts matter. A confident answer can still be wrong. Information supplied to a model is not the same as information it retrieves and both are different from what the model asserts from memory. More context can help but does not automatically improve an answer. How company data is handled can differ widely between consumer tools and the systems a firm has approved. None of this requires a technical background. It requires knowing which questions to ask before trusting an answer.
The third is being able to judge it. That means holding outputs to the standards of the job rather than the fluency of the text. We find this the hardest to teach and the most important.
The same tool can require very different competence
Consider a lease review. You ask AI to extract rent, expiry, break options and indexation from twenty leases. The task looks simple and saves time. Literacy here means knowing what can go wrong. Is the break option conditional? Did a side letter somewhere in the file change a date the model lifted from the original lease? Which fields must be checked against the source before anyone relies on them? The extraction is not the work. The verification is.
Take contract summaries. They carry a subtler version of the same risk. AI may render the usual clauses precisely yet glide over the one unusual clause. A summary then reproduces the typical and misses the atypical. In that case, the reader sees a complete looking picture and has no way to tell what may have been smoothed out.
Market research demands a different skill. The user has to understand sources and dates and the line between evidence and interpretation. A market claim built on a source from five years ago may be perfectly traceable and still useless for a decision today. The date of a source is part of the evidence and not a footnote to it.
Now the board member reading an investment paper. Much of what reaches a board today was produced with help from AI. That takes its own kind of literacy. How was the analysis produced? Which parts were checked? Where did human judgment enter?
Part of that literacy is governance. In a forthcoming paper we describe this as a shift from automation to delegation. AI now takes on work that once went to a junior colleague and the old rule of delegation still holds. Responsibility does not transfer with the task. A blanket human check is not governance. We keep seeing the same pattern. In the first weeks of a rollout reviewers read every output carefully. As volume grows the queue lengthens and approval can turn into a stamp. The control survives on paper long after it has stopped working. Review capacity is finite and output volume is not. In our experience oversight has to match consequence. Light checks where mistakes are cheap. Strict verification where they are not.
Take contract summaries. They carry a subtler version of the same risk. AI may render the usual clauses precisely yet glide over the one unusual clause. A summary then reproduces the typical and misses the atypical. In that case, the reader sees a complete looking picture and has no way to tell what may have been smoothed out.
Market research demands a different skill. The user has to understand sources and dates and the line between evidence and interpretation. A market claim built on a source from five years ago may be perfectly traceable and still useless for a decision today. The date of a source is part of the evidence and not a footnote to it.
Now the board member reading an investment paper. Much of what reaches a board today was produced with help from AI. That takes its own kind of literacy. How was the analysis produced? Which parts were checked? Where did human judgment enter?
Part of that literacy is governance. In a forthcoming paper we describe this as a shift from automation to delegation. AI now takes on work that once went to a junior colleague and the old rule of delegation still holds. Responsibility does not transfer with the task. A blanket human check is not governance. We keep seeing the same pattern. In the first weeks of a rollout reviewers read every output carefully. As volume grows the queue lengthens and approval can turn into a stamp. The control survives on paper long after it has stopped working. Review capacity is finite and output volume is not. In our experience oversight has to match consequence. Light checks where mistakes are cheap. Strict verification where they are not.
Experience changes what literacy means
This industry adds a complication. AI can make inexperienced people look more capable than they are. A junior analyst can now produce a clean market summary and a sophisticated list of risks in minutes. That is useful. It also makes it harder to tell from the output alone how much the author understands.
Weakness used to show in the first draft. A thin grasp of a market produced a thin market section. AI can make that section read well. The quality of the writing and the quality of the understanding beneath it have come apart. A manager who once judged juniors by their drafts has lost that signal. Managers need different signals. Can the analyst explain the assumptions? Defend the sources? Say where the analysis could be wrong? Literacy therefore cannot only be about producing better outputs. It must include knowing when you are not qualified to judge the output you just produced.
Picture a situation that can arise in practice. A fluent AI written market section circulates and somewhere in it sits one clearly unrealistic assumption. An experienced professional is likely to notice it quickly. A junior colleague may simply improve the wording. Same tool and same output but a very different reading. We see that gap as a real risk in a sector where experience and critical thinking are crucial.
Weakness used to show in the first draft. A thin grasp of a market produced a thin market section. AI can make that section read well. The quality of the writing and the quality of the understanding beneath it have come apart. A manager who once judged juniors by their drafts has lost that signal. Managers need different signals. Can the analyst explain the assumptions? Defend the sources? Say where the analysis could be wrong? Literacy therefore cannot only be about producing better outputs. It must include knowing when you are not qualified to judge the output you just produced.
Picture a situation that can arise in practice. A fluent AI written market section circulates and somewhere in it sits one clearly unrealistic assumption. An experienced professional is likely to notice it quickly. A junior colleague may simply improve the wording. Same tool and same output but a very different reading. We see that gap as a real risk in a sector where experience and critical thinking are crucial.
Training by role rather than by tool
In the firms we talk to the tool is rarely the problem. What slows them down is scattered data, the way work is organised, workflows and missing skills. Once a technical basic fundamental understanding of AI has been obtained, training should therefore begin with the role. What does an investment professional use AI for? An asset manager, an investment manager? Someone in capital raising, fund management, development or research? The same question applies beyond the front office. HR, finance, legal, compliance and marketing work with AI on their own material and carry their own risks. Think of recruitment decisions or privileged documents or reported numbers. The material differs and so does the potential cost of a mistake. What decisions follow from those outputs and which mistakes matter in each case? For an investment team the priority is whether sources hold and assumptions make commercial sense. For asset management it is whether extracted numbers and clauses are right. For HR it is sensitive data and where human judgment must stay. Training built this way stays close to the work. Take a real lease or investment paper and walk through the decisions made with it. The Commission's guidance points the same way. It ties literacy to the knowledge and experience of staff and to the context of use rather than one course for everybody.
The question worth asking
The revised obligation makes it tempting to do less. We think that would miss the point. The legislature could have demoted the duty to a recital and chose not to. None of the risks in this article came from the regulation. A conditional break option. A market claim built on a stale source. Confidential data typed into the wrong system. An analysis that reads well and makes no commercial sense. None of these becomes cheaper because the compliance standard eased. They were never compliance problems. They are problems of quality and judgment and they live in the work whether or not an authority ever asks. The industry seems to sense this. What senior professionals keep telling us they want is proper training and real skills.
The better question is not what the law requires. It is what the work requires. Decisions are only as good as the judgment of the people who make them and that judgment now includes evaluating what the tools return. To us that is a question of capability rather than compliance and it is the more demanding of the two.
Which leaves one exercise for any team. Work it out role by role. The answer differs for an analyst, an asset manager and a board member. And if someone here says they are competent to use AI for their work, what exactly do you expect them to know?
The better question is not what the law requires. It is what the work requires. Decisions are only as good as the judgment of the people who make them and that judgment now includes evaluating what the tools return. To us that is a question of capability rather than compliance and it is the more demanding of the two.
Which leaves one exercise for any team. Work it out role by role. The answer differs for an analyst, an asset manager and a board member. And if someone here says they are competent to use AI for their work, what exactly do you expect them to know?
Created with
