What AI should — and should not — do on a construction project

There is a lot of noise about AI in construction at the moment, and most of it is unhelpful in the same way: it treats the interesting question as a capability question. Can the model draft a response to an RFI? Yes, it can. That was settled some time ago.

The question that matters on a live project is different. Which decisions must a named person remain accountable for, and what does that imply about how the tool should be built?

Where AI genuinely helps

The honest answer is that AI is very good at the work nobody wants to do and nobody gets credit for.

• Reading and classifying incoming documents. Identifying what has arrived, what it references and where it belongs is tedious, high-volume and error-prone when done by a tired person at 6pm.
• Preparing a first draft. Editing a considered draft is a fundamentally different task from facing a blank page, and it is markedly faster.
• Surfacing patterns across the record. No human reads ten thousand documents looking for a trend. That is exactly what software should do.
• Answering questions of the record. “Which submittals have been with this consultant longest” should not require anyone to build a report.

Where it must not be trusted

Anything that creates or defends a contractual position. A notice, an instruction, a rejection, an approval — these carry consequence, and consequence requires a person who can be asked why.

This is why approval in Fortis Sync is mandatory and cannot be switched off. It is not a limitation we intend to remove in a later version. A platform whose output can be issued without review is a platform that has quietly transferred accountability to a system that cannot hold it.

The confidence score matters more than the draft

The genuinely useful feature is not the draft itself but the honest signal about how much scrutiny it needs.

A response drafted from an unambiguous specification clause and a clear precedent deserves a quick read. One drawn from contradictory correspondence deserves an engineer’s attention. Surfacing that distinction — rather than presenting every draft with equal confidence — is what makes the tool safe to use at volume.

A reasonable test

Before adopting any AI feature on a project, ask: if this output turns out to be wrong, who is accountable, and did they have a genuine opportunity to catch it?

If the answer to the second half is no, the feature is not ready, however impressive the first draft looks.