AI agents in construction: what they may decide
Every agent on this platform proposes. A named person disposes. The record says which of the two happened, on every single entry.
AI agents in construction are usually sold as autonomy: the software watches the project and acts on it. That is the wrong shape for this industry, and not because the models are weak. A construction decision has a name against it — an approval, a certificate, an instruction — and the person whose name it is carries the consequence of it for years. Software that decides on their behalf does not remove the consequence. It removes the trail back to whoever was supposed to carry it.
What can an AI agent actually do on a project?
Everything that is not a decision, which is most of the work. Read the specification and say which clauses have no verification method against them. Watch a programme and name the activities that have quietly lost their float. Read four subcontract returns and say which exclusions make them incomparable. Draft the pay-less notice with the ground and the calculation already set out. Each of those arrives as a proposal, with the records it read attached to it, and then it waits.
The waiting is the point. A proposal sitting in a queue is an hour of work somebody checks in thirty seconds. A proposal that executed itself is a change nobody checked at all — found later, usually by the other side, usually at the worst moment.
Where is the line between proposing and deciding?
At the event. Every change here is an event on the golden thread, and the catalogue that defines the event types marks which of them software may author and which it may not. A decision — an approval, an award, a certificate, a stage gate — carries a flag that refuses an AI author outright. That is not a setting an administrator can relax on a busy afternoon. It is a property of the event type, checked on the single write path every change in the system goes through.
The agents are bounded from the other end as well. Each holds a mandate, and the ladder stops at propose. No agent holds anything above it, in any capability area, on any subscription tier — so there is no upgrade, no enterprise exception and no flag that turns an agent into an approver. The ceiling is the design rather than a default.
How do you prove an agent did not decide something?
By replaying the record and reading who authored what. The ledger is append-only and hash-chained: each event carries the hash of the one before it, so an entry cannot be altered or quietly dropped afterwards without breaking every hash that follows it. A replay rebuilds the project from its own history and reports a root hash, and two replays that disagree mean the record was touched.
On top of that, every AI-assisted step names the engine that produced it, the records it read and the person who accepted it. So the question an adjudicator asks — who decided this, and on what — has a literal answer, and that answer separates the draft from the decision. Separating those two is the entire reason the flag on the event type exists.
An agent that cannot decide is an agent nobody has to defend.
What do agents cost to run on a project?
Metered, and refused outright when the meter is empty. Each run draws on a wallet the account funds, with pricing per call rather than per seat, and the platform will not call a provider against an empty wallet — it says so, rather than failing quietly and billing for it later. Every charge is written beside the output it paid for, so a month of agent work reads as a list of things produced instead of one line on an invoice.
Why the line matters more in construction than elsewhere
Because this industry settles its disputes on records, years later, in front of somebody who was not there. It is won by whoever can show what was known and when. Civil infrastructure and water treatment programmes run for years across dozens of firms, and the record has to survive every one of them leaving the job. An agent that had been quietly writing decisions into that record would not be an efficiency. It would be the liability at the centre of the case.
What you can check without taking our word for it
The demonstration environment runs a seeded programme end to end, with the proposal queue live and every agent’s output attributed to the engine that produced it. Nothing in it is a screenshot.
Any document the platform issues carries a content hash, so a stranger holding one — a solicitor, an insurer, a certifier — can check it against what was issued, with no account and no relationship to anybody involved.
The API is documented for anyone who would rather read the event catalogue than a brochure, and it sets out the eight engines and the stage gates they sit behind.
Platform status is a public page rather than a support ticket, which is the same argument in a smaller frame: a claim anybody can check is worth more than a claim anybody has to accept.
If the design of the ledger interests you more than the pitch does, the engineering notes on this site are where it is written down. If the shape of your own projects is the real question, talk to us and bring one — an agent that may not decide is easiest to judge against work you already know the answer to.
See the record being made rather than described.
Launch the demo environment