Introducing AI in a construction company starts with people and AI being able to find the same current document. The drawing revision, document status, responsible person and access permissions need to be clear before we expect AI to give a reliable answer.
On 22 October, I attended the Estonian Construction Conference 2026. One of the clearest presentations of the day came from Dmitri Garbuzenko, Head of Innovation at Endover. He spoke about construction data, introducing AI and checking the answers a system produces from your own documents.
Most of these principles were familiar to me. Still, it was good to see a speaker who wants to move things forward within his company and explains the necessary foundations clearly. His presentation left me with the impression that he carries much of the responsibility for driving this change. For me, it was one of the stronger talks of the day.
There was also a point where I had a perspective of my own: if testing leads to moving to a different AI tool every month, it is worth asking what the company gains overall. More on that below. First, the data, because the value of organising it does not disappear when the next model arrives.
Why does construction data quality affect an AI answer?
Take a simple situation. There is a drawing in the project folder, a newer version in an email attachment and a file called “final_corrected_new” on someone’s computer. Each looks plausible when opened on its own. If AI receives the wrong one, it may produce a fluent answer based on outdated information.
In that case, the problem is not limited to the model’s capabilities. The tool was given the wrong source, or nobody specified which document was current. An employee would also struggle to make the right decision in that folder.
Construction document management should help answer at least three questions: what is this document, can it be used for this task and where can I check whether it is current? Access is another part of the picture. Having a project document platform does not mean every employee or connected AI tool should be able to see every file.
For example, Bauhub’s file management features already include version control. That helps when the team uses it in an agreed way. If the latest drawing still travels only as an email attachment, a person or AI working from the project platform will miss that change.
What should you organise in construction documents first?
I would start with one project and one recurring task. For example, a project manager needs to find the current drawing for a particular activity and the requirements associated with it. You do not need to rebuild the company’s entire archive before running the first trial.
The following checklist is a simple example of how to organise the work. It does not describe a client’s project and is not a universal file naming standard.
- Document identifier. Which project, discipline and document does the file belong to? The identifier should make it possible to find the file unambiguously.
- Revision. Which version is current? An older version can remain in the history when needed, but it must not silently replace the current one.
- Status. Is the document being prepared, under review or approved for a particular use? The most recently uploaded file may not be approved.
- Responsible person. Who uploads a new version, checks the information and approves the document for use?
- Location. Where does the team find the agreed current information? Email attachments and local copies should not become separate, unmanaged sources.
- Access. Who may view, change or share the document? Which user’s permissions does the AI connection operate under?
- Relationships. Which drawing, meeting record, certificate or change does the document relate to?
The distinction between revision and status matters. A new revision tells you that something has changed. It does not, by itself, confirm that the change has been approved or issued for construction.
The UK BIM Framework’s guidance on the common data environment also explains document states, revisions and information exchange workflows through the principles of ISO 19650. The practical value is traceability: people can understand the stage the information has reached and what it can be used for. This is implementation guidance; the checklist above is my simplified starting point.
An audience question also raised a concern about giving AI project information marked AK, the Estonian designation for information intended for internal use. That highlights a separate check: a file being organised and accessible does not automatically mean it is permitted to use it in an AI service. Before connecting the systems, clarify the conditions for using the data and the company’s agreed rules. Technical access and permission to use information are separate questions.
An AI connection does not replace rules for the project document platform
I have written separately about Bauhub’s MCP connection and AI in construction. The connection allows AI to interact with the project environment through permitted tools. It does not decide on the company’s behalf who must add a missing document or whether work can continue using an old drawing.
That is why I would keep setting up the connection and organising the data in the same work plan. If an assistant needs to answer a question about a drawing, it must be able to access the right file, know its revision and status, and show the source alongside the answer. If the source is missing or the documents contradict one another, that needs to be clear in the response.
Simply asking it to “be accurate” will not resolve unclear document management. A better task description specifies which source to use, which fields to check and when to identify missing information instead of providing an answer.
Permissions need to be checked in the specific system. For example, Microsoft’s guidance on SharePoint agents states that permission to use an agent does not automatically give a user access to all its sources. The behaviour of a different connection needs to be checked separately.
What did I take away about testing AI?
Dmitri also covered evaluations, often called evals, in his presentation. In plain language, this means asking a system questions relevant to your work and comparing its answers against verified source information. It helps you see where the system makes mistakes, fails to find something or provides an answer without a source.
For me, the important question is which decision the test helps us make. If we move to Claude one month, Copilot the next and then ChatGPT, we also spend time learning new ways of working, checking connections and helping the team adjust. A better result in one test does not yet show the overall benefit of switching.
That does not mean we can stop checking. Models change, and today’s ranking may change with them. A new version may be better at some tasks, but we cannot promise that in advance for our project’s documents.
I would want to keep the task, the test questions and the quality requirement stable. I would switch tools when the benefit outweighs the work of reconfiguring the setup and learning to use it. When the model, connection, search or source information changes significantly, it is worth running the same check again.
For this article, one simple trial is enough: can the assistant find the current drawing, show its revision, cite the source and clearly say when there is no approved document? A thorough approach to testing AI answers deserves an article of its own. When organising data, even this trial helps reveal whether the obstacle lies in the source information, permissions, search or the way the answer is produced.
The NIST AI Risk Management Framework also emphasises evaluation in conditions similar to the intended use, along with documenting tests and metrics. My construction example is a practical adaptation of that principle, not a universal ranking of models.
Who will keep the construction data organised next week?
Organising the folders once is a start. The process begins to work when the next document also reaches the right place.
That is why I would assign a named person to each trial and agree on a simple process: who adds a new revision, who changes the status and who is notified when a source is missing or contradictory. The person leading AI adoption can help establish the workflow, but keeping documents in order day to day also requires the people working on the project to contribute.
At the end of the first trial, I would not only ask whether the answer sounded intelligent. I would ask whether the employee found the right document, could check whether it was current and knew what to do if something was missing. That is where the benefit starts to emerge, and it can then be measured.
Where should your construction company start?
Choose one task where searching for documents or confusion over versions is a recurring problem. List the files and people responsible for that task. Agree on where to find the current information, then try AI assistance on that clearly defined task.
For me, that was the most important reminder in Dmitri’s presentation: introducing AI still depends on the foundations of work organisation that people rely on too. Well-organised construction data helps the team now and provides a better starting point for the next AI model as well.
If you want to work through this with your team, I can help choose a suitable first task and teach people how to find information in documents with AI and check the result. AI team training can start from the way your team works. Tell me what your team needs.