AI agents for construction companies, working the approvals backlog

The approvals admin that holds up the build

Ask a property developer where the program slips and approvals come up fast. A site can be ready and a builder lined up while the project waits on paperwork. Development applications run to hundreds of pages. Councils come back with requests for information that each carry a clock. Consultant reports, conditions of consent and contracts pile up, and every one of them has to be read closely because a single buried condition can change the cost, the design or the timeline.

The work is mostly reading, and reading does not scale with the team. A planner or development manager spends hours combing a DA pack or an RFI to find the few clauses that matter, then drafts a response, then cross-checks it against the conditions already imposed on the site. The conditions cross-reference each other, the language is technical, and the documents arrive faster than anyone can absorb them. Miss one obligation and it surfaces later as a variation, a delay or a dispute.

The setting is unforgiving. Approvals sit inside state planning frameworks and council processes, RFI responses have deadlines, and the conditions of consent become binding obligations that govern the whole build. Personal information turns up throughout the documents and is handled under the Privacy Act 1988. This is not a place for guesswork. It is a place where careful reading is the bottleneck, and careful reading is exactly the kind of work that piles into a backlog.

Why Claude, and what sits around it

The aim is to take the reading and first-draft load off people while leaving every decision with them. We headline these agents on Claude, Anthropic’s large language model, for one practical reason above the rest. Approval documents are long and densely cross-referenced, and Claude holds a large document in context and reasons across it accurately, so it can read a full DA pack or deed without losing the thread between a condition on page 12 and the schedule it points to on page 180. For documents whose whole problem is length and cross-reference, that capability is the build, not a nice-to-have.

Claude does the reading and reasoning. Retrieval-augmented generation (RAG) keeps it grounded. Rather than rely on the model to remember what a condition said, the agent retrieves the exact source clause and ties every extracted obligation back to it with a citation. The developer’s own library of conditions of consent and contracts becomes the source of truth the agent checks against, so a drafted RFI response is anchored in the actual wording rather than a plausible paraphrase. The whole service runs inside the developer’s Microsoft 365 environment, so the documents stay where the team already works and access follows the permissions already in place.

The agent’s job is deliberately bounded. It reads, extracts the obligations and changes, drafts the council response, cross-checks it against the conditions already on the site, and flags what a person must decide. Anything ambiguous goes to a human rather than getting a confident guess.

Building it, and where it got hard

The model was rarely the hard part. The friction lived in trust, and one failure stands in for the rest.

Early in testing the agent read confidently and wrote well, which turned out to be the danger. On a site governed by more than one consent it mis-attributed a condition, tying an obligation to the wrong consent while sounding completely sure of itself. In approvals that is the worst way to be wrong. A condition attached to the wrong consent can send a response down the wrong path, and because the prose reads cleanly, a busy reviewer can wave it through. Confident and wrong is more dangerous than visibly unsure.

The fix had three parts and all three mattered. First, grounding. Every obligation the agent extracts is now tied to its exact source document and clause with a citation, so a reviewer clicks straight to the wording instead of trusting a summary. Second, a flag-don’t-guess rule. Where a reference is ambiguous or a condition could belong to more than one consent, the agent stops and flags it for a person rather than picking. Third, a mandatory human review before any RFI response leaves the building. No draft reaches a council until a qualified person has checked it.

Two constraints shaped the rest. RFI deadlines meant the agents had to surface the time-critical items first, so triage sorts by due date and program impact rather than by the order documents arrived. And because the documents carry confidential commercial terms and personal information, everything stayed inside the developer’s tenancy, access followed existing permissions, and the material was not used to train the model. None of this is glamorous. All of it is the difference between an agent the team trusts with the backlog and one they quietly stop using.

What changed

In a representative deployment the agents triaged about four in five incoming approval documents into obligations, changes and open questions before a planner opened them, so review started from a structured brief instead of a blank pack. Draft responses to council requests for information came back in hours rather than the days a manual first pass had taken, each point cited to its source clause so the reviewer could verify it quickly. Grounding every obligation to its exact consent and clause caught cross-referenced conditions that a fast manual read had let slip, the kind that resurface later as a variation.

These figures are illustrative. They describe the pattern we see rather than a published result for a named developer. The shape is the point. The reading and first-draft load comes off the team, the time-critical RFIs surface first, and the people who carry the decisions get their attention back for the conditions that actually move the program. The agents clear the backlog; the developer’s team stays responsible for every call.

Where this fits

These approval agents are one application of our AI Agents service, built on Claude and grounded with retrieval, for the realities of Australian property development. It is a contained, high-return starting point, because the documents already exist and the value comes from reading them properly and getting the structured result in front of the right person fast. If approvals admin is where your program slips, the place to start is to map your DA, RFI and consent documents and decide where a person must stay in the loop.