A builder AI permission list states what each connected tool may read, draft and check, and which actions need a person. Record the prohibited approvals, payments, field directions and commitments next to the account permissions. Write the list before you connect project records. Then test that a draft cannot become an issued instruction without the responsible reviewer.
Which decisions go on the permission list?
The ones that stay with the role that already owns the consequence. The superintendent decides how to respond to a field condition through your operating path. The estimator or principal approves pricing work. The design or engineering professional settles technical matters. The warranty owner weighs a request against the governing documents and the facts.
An agent helps those people see information faster. It compares document versions, summarizes an intake, drafts clarifying questions. The help stops short of the decision, the approval and the external commitment.
Why is a confident AI answer not enough?
A model produces a complete-sounding answer when the source is incomplete, old or contradictory. It has not stood on the slab, seen the concealed condition or heard which conversation became an approved change. Confidence is a writing quality, not proof.
| Decision area | Permitted assistance | Human authority |
|---|---|---|
| Contract and scope | Locate supplied clauses and compare wording | Interpret obligations and approve scope |
| Safety and field direction | Organize approved procedures or reports | Assess conditions and direct work |
| Engineering and code | Collect questions and source references | Qualified review and current determination |
| Pricing, draws and payments | Assemble documented inputs | Verify, approve and release |
| Selections and changes | Track status and flag missing records | Approve item, scope, effect and release |
| Warranty | Classify intake and draft questions | Decide coverage, priority and response |
Why does AI never make schedule promises?
A schedule depends on current field progress, inspections, weather, material status, trade coordination, buyer decisions and project-specific constraints. The agent sees only what it was given. It can draft an update from approved notes. It cannot verify what happened after those notes were written.
Prohibit language that turns a target, an estimate or a trade expectation into a commitment. The same goes for a spec home completion date on a listing page. If a date matters, the responsible person checks the current schedule and conditions before the message goes out. The AI weekly buyer-update guide describes the reviewed drafting path.
Can AI send routine messages on its own?
Routine messages still carry risk when they touch timing, cost, selection status, completion, inspection results or warranty responsibility. Start with drafts. If you later automate one narrow notice, define the exact trigger, the permitted text, the failure path and the person who watches it.
Where does AI stop in estimating and document work?
It can sort quote lines, compare category labels and flag a missing document. It does not choose a quantity, assume trade scope, fill an allowance, accept a substitution or release a draw. Those actions depend on current records and accountable review.
Specifications get the same boundary. AI assembles an initial document from controlled sources and shows the conflicts. It cannot decide that an old plan note overrides the current schedule. See AI specification-sheet drafting for the source-control method.
What about codes, permits and platform rules?
Use agents to find possible sources and prepare questions. Verify the current controlling record with the relevant authority or a qualified professional. A model's remembered rule never becomes a permit assumption, a site instruction or a published claim.
How do you enforce boundaries instead of stating them?
Give each connected tool a permission row: source access, draft output, prohibited action, human approver and shutdown control. Compare that row with the real account settings. Keep send and approval permissions with people. Require outputs to show their sources and their unresolved fields.
- Name the decision owner. The role authorized to approve the final action.
- Restrict the task. What the agent may read, draft, compare or classify.
- List prohibited outputs. Commitments, approvals, payments, technical judgments and invented facts.
- Require visible uncertainty. Missing or conflicting information is marked for review.
- Separate draft from issue. The working output looks nothing like an approved record or an external message.
- Test failure cases. Incomplete, conflicting and superseded inputs. Confirm the workflow stops.
- Keep a manual path. The team can do the work when the agent or the integration is down.
The human review checklist turns these boundaries into a repeatable approval gate.
What happens when an agent gets something wrong?
Stop the affected output. Compare it with the source. Work out whether the problem came from bad input, unclear instructions, missing context or a model error. Correct the project record through your normal process. Do not quietly patch the sentence and leave the workflow as it was.
Update the instruction, the source boundary or the review checklist. Retest the failure. If the error shows the task needs judgment the agent cannot reliably supply, take that task out of the workflow.
Who stays accountable for the output?
The person who approves and uses it. A checking agent catches formatting, missing fields and inconsistent labels. It does not move responsibility away from your authorized reviewer.
How does permitted work appear in the runbook?
Copy each permitted task from the permission list into its runbook: organizing meeting notes, preparing follow-up drafts, flagging selections with missing status, structuring warranty intake, checking content against approved sources. We run a fleet of Claude and Codex AI agents at AgAI Labs that research, draft, build and check inside assignments like these. People set direction and review every output.
The AI operations stage defines access, review, failure rules and ownership around one bottleneck. The AI guide for home builders holds the broad operating model. The remaining tactical workflows are in the AI for builders library.
Which AI permission questions come up most?
Can AI approve a change order if the numbers match?
No. Matching fields do not confirm scope, authority, project effect or governing approval. An authorized person reviews every change order.
Can an agent answer warranty questions from our warranty document?
It can draft a response or find the relevant wording. A person decides applicability, priority and the final communication.
Can AI tell a superintendent how to correct field work?
No. Field direction needs current site knowledge, approved documents and the right technical and operating authority.
Does a second AI check count as human review?
No. A checking agent surfaces possible errors. The accountable person still verifies sources, judgment and consequences.