Menu

AI construction software development

Construction software development on AWS: site meetings spoken in each person's own language become reviewed, versioned minutes and tracked tasks.

A construction technology company · Client not named

At a glance

Key facts
MeasureValue
CaptureVoice or text, on phone, tablet or desktop
LanguagesSpoken in each person's language, issued in the project's official language
AWS servicesTranscribe, Bedrock, S3, Cognito
Task statusesTo do, in progress, waiting, completed, validated, cancelled
Contractor link actions"I have taken note" or "Add a remark"

The problem

Construction meeting minutes are slow to produce. A site manager collects information during the meeting, then rewrites it afterwards: picking out the decisions, assigning responsibilities, adding deadlines and checking by hand what was still open from the previous meeting. On an active project that happens every week.

Multilingual site teams add a second problem. A highly competent field professional may not be comfortable writing the language the official minutes are kept in, so someone else rewrites what they meant, which costs time and loses detail on the way.

And the minutes carry weight. They record who agreed to what and by when, and they can matter in later disputes. Output from an AI model could not simply be published. The workflow had to keep people in control of every word that becomes official.

What we did

We designed the product around one short flow: open the project, speak or type, review, send. Notes are captured by voice or text on a phone, tablet or desktop, because most of them happen on site rather than at a desk. Each capture records the project, the author, the time, the language detected and the project's official language.

The pipeline is designed for the way site teams actually talk. A foreman can dictate in Portuguese or Albanian, mix in French technical terms, and the draft minutes come out in the project's official language. Before anything is issued, the contributor can check in their own language that the structured result says what they meant, and an authorised user approves the version in the official language. One person can do both when they read both languages.

The AI follows one strict rule: it never invents. If someone says a subcontractor "must finish quickly", the system does not make up a deadline. Missing information is grouped under "To complete", uncertain transcription or translation is marked "To review", and the original wording behind any item is one tap away. Known names, such as companies, contacts, streets and buildings, are passed to the model as project context so they come out spelled correctly.

Tasks are persistent project records, not lines retyped into each meeting. Every task keeps its identity from the meeting or field note where it started, and every change of status or deadline is stored with who made it and when. Open tasks carry forward into the next meeting automatically. Completed, validated or cancelled tasks leave the agenda but stay in the history.

Information that turns up between meetings is captured as a field note, by voice or text with an optional photo. When the next meeting is prepared, the notes added since the last one are offered back, and the user decides whether each goes into the minutes, becomes a task, stays a note or is archived. Nothing the AI proposes becomes official without that decision.

Approval produces a version that cannot change. Version 1 of a set of minutes keeps the values it was issued with, even if a task changes later, and any correction creates version 2. Minutes are sent from the application with a history by recipient, and contractors without an account receive a secure link that allows only two actions: "I have taken note" or "Add a remark". Links use random tokens, expire and can be revoked. Each project sets its own approval rule, from tacit approval at the next meeting to no tacit approval at all.

The platform runs on Amazon Web Services. Voice is transcribed with Amazon Transcribe, AI-assisted extraction runs on foundation models through Amazon Bedrock, documents and their versions are stored in Amazon S3, and sign-in and roles are handled by Amazon Cognito. AI providers sit behind an internal interface, so a model can be replaced without rebuilding the product, and every AI run is logged with its provider, model, language, status and duration so failures can be diagnosed.

Around the workflow sit the parts a production SaaS product needs: customer workspaces with tenant isolation enforced on the server, roles from organisation admin to read-only, an audit log of sensitive actions, a data residency design that keeps production data in Switzerland, and a retention policy that deletes audio after review unless it is explicitly kept. The commercial layer is part of the product too. Usage is measured in AI credits rather than tokens, invoices are generated and emailed monthly, and an unpaid account moves to read-only after a configurable grace period, then reactivates as soon as payment is recorded.

What changed

The result is a meeting record a site team can defend. Every decision and task can be traced back to what was actually said, by whom and when, and an issued version never changes underneath the people who received it.

It also changes who can contribute. The person who knows the work can describe it in the language they are comfortable in, instead of waiting for a colleague to rewrite it, while an authorised reviewer keeps the final say over the official text.

And the minutes stop being the end product. Because tasks, decisions and field notes are stored as structured data with their history, every meeting adds to a structured record of the project rather than another PDF in a folder.

Common questions

How do you stop AI from inventing details in construction meeting minutes?

By making it a rule of the pipeline rather than a hope. The pipeline is built so the model never creates a date, a responsible party, a decision or a status that is not in what was said. Anything missing is left blank and listed under "To complete", anything the transcription or translation is unsure of is marked "To review", and the original wording behind every item can be displayed. Nothing becomes official until an authorised person has reviewed and approved it, and the approved version is then frozen.

Can site workers dictate in a different language from the official minutes?

Yes, that is one of the main design goals. The language of each note is detected and can be corrected, while the official language of the minutes is a project setting. Speakers can mix languages, such as French technical terms inside a Portuguese sentence. The contributor checks the structured result in their own language, and an authorised user approves the official-language version. That can be the same person if they read both languages.

What does AI software for construction need beyond the AI itself?

Most of the engineering. The model turns speech into structured content, but a product a project team can rely on also needs tenant isolation between customers, roles and permissions, an audit log, documents that cannot change after they are issued, secure links for external contractors, retention rules for audio and AI logs, monitoring of AI failures, and a billing layer that measures usage. Keeping AI providers behind an internal interface also means a better or cheaper model can be adopted later without rebuilding the product.

Services involved

Working on something similar?

Book a call and talk it through, or start with our Cloud Cost and Reliability Review.