AI Agents in a Construction Company

Most writing about “AI in construction” is about robots, drones and digital twins. The work that actually eats a construction company’s week is quieter: reading hundreds of pages of tender documentation, checking eligibility, drafting offers and appeals, looking up standards, writing site reports. That is language work — and language work is exactly what today’s AI models are good at.

We recently worked with a construction company from Iași — civil works and hydro projects, more than 25 employees, over €2.5 million in annual revenue. We built AI agents for three of their departments and trained their people to use them. The part we are proudest of is not the agents. It is that the team now improves them, and builds new ones, without us.

Why Claude, and why not a custom app

The company did not need new software. They needed their existing people to do their existing work faster, with the documents they already have. So instead of building an application, we built on Claude, which their team uses directly:

Projects — one per department. A project holds the department’s reference material (standards, templates, past documents) and standing instructions, so every conversation starts with the right context instead of a blank page.
Skills — reusable procedures packaged as instructions plus supporting files. A skill encodes how this company does a task: what to check, in which order, which format the result must follow. Claude loads the right skill when the task calls for it.

This choice kept the setup in the hands of the people doing the work. There is no codebase to maintain, nothing to deploy, and nothing that only a developer can change.

Tenders: the department with the most reading

Public tenders in Romania come with long documentation and short deadlines. The tenders team now works with agents that help them:

Analyze the tender documentation — summarize the specifications, pull out requirements, deadlines and award criteria, and flag what is unusual.
Check eligibility — compare the qualification requirements (similar experience, certifications, key personnel) with what the company can actually prove.
Prepare offer documents — draft the technical proposal, declarations and answers to clarification requests in the company’s own format.
Work through quantities and cost estimates — help extract quantities and cross-check bills of quantities.
Draft appeals — when the team decides to challenge a procedure, the agent helps structure the appeal, connect the arguments to the documentation and keep the wording precise.

The agent drafts and checks. The tenders specialist decides, edits and signs. That split is deliberate (more on it below).

Civil works and hydro: standards and site paperwork

The civil works and hydro teams have different projects but similar needs:

Standards and regulations — answers grounded in the standards and technical norms loaded into the project, instead of someone searching through PDFs.
Site documentation — reports, handover and inspection records, work progress statements and correspondence with the client, drafted in the company’s templates.
Reviewing technical designs — reading technical memos and specifications, spotting risks and inconsistencies before they reach the site.
Planning and offers — help with execution schedules, construction methods and requests for quotes to subcontractors.

Hydro work adds its own vocabulary and standards, which is exactly why it got its own project and skills instead of sharing one generic “construction assistant”.

The training: from users to maintainers

We could have handed over the agents and left. They would have been useful for a few months, then slowly drifted away from how the company actually works. Templates change, standards get updated, people find better ways to do things.

So the training covered four levels:

Using the agents in daily work — how to give context, how to ask, how to iterate on a draft instead of starting over.
Improving the skills — when an agent gets something wrong or a procedure changes, the team edits the skill themselves. A correction made once is applied every time after.
Creating new agents — once people understood how a skill is built, they could package their own recurring tasks into new ones.
Good practice and limits — what data may go into the AI, how to verify results, and when not to trust an answer.

The fourth level matters as much as the first three.

The rules that keep people in charge

An AI agent that writes a tender appeal or a site report is powerful, and that is exactly why the rules have to be clear. The principles we trained on:

The agent drafts, a person decides. Every document that leaves the company — an offer, an appeal, a report — is reviewed and owned by the responsible person.
Check against the source. An answer about a standard or a requirement must point back to the document it came from, and the person checks that reference.
Know what goes in. Clear rules about which documents and data can be used with the AI.
Errors fix the skill, not just the document. When something is wrong, correct the output and the skill, so the mistake does not come back.

This is the same principle we apply to our own software work, where AI drafts the code and senior engineers review and own every release. The tools differ; the accountability does not.

What we learned

Start with the work that is mostly reading and writing. Tenders and site paperwork are where the value is easiest to see, and early wins build the trust everything else depends on.
One project per department beats one assistant for everyone. Context is what makes the answers good.
Skills are the real asset. They capture how the company works, in a form the team can read and change.
Training is not an extra. It is what turns a set of agents into a capability the company keeps.

Want agents like these for your team — and the training to keep them improving? See how we approach AI product development or tell us about your workflow.