Superintelligence Prep for Deck Builders: Designs, Permits and Changes

An AI agent for a deck builder is a bounded assistant that keeps the project's design revisions, permit status and change order records in one shared memory, so the estimator, the crew lead and the homeowner work from the same current facts. It drafts updates, flags missing approvals and prepares follow-ups, while a named person approves anything that changes scope, cost or code compliance. Setting this up now is mostly a matter of defining the records, their owners and the approval rules before the agent touches a live job.

Key takeaways

• Start with three record types: design revisions, permit items and change orders, each with an owner and a current state.
• Keep the agent's memory tied to the project file and its written approvals, not to a chat thread nobody can audit.
• Give the agent read access to current drawings and write access only to drafts, notes and status requests.
• Require a named person to approve any change that moves scope, cost or code compliance.
• Judge the setup by one question: can someone on site state the approved version and the open items without calling the office?

Why deck projects lose track of what was approved

A deck job moves through many hands. The first sketch often lives on a phone, the stamped drawing sits in an email thread, permit comments arrive through a portal or a letter, and the client changes the stair position during a walk-through and says it out loud. Each piece is reasonable on its own. Together they produce a job where the crew cuts from one version, the office prices another and the inspector reviews a third.

The problem is rarely lack of effort. The project has no single place where the current state is stated, dated and linked to its evidence. An agent can help, but only if the record it reads and writes is designed for that job. A chat history is not a project memory. A durable memory is a set of structured records that survives staff changes, busy weeks and a foreman who forgets to mention a verbal request.

Design revision tracking as project memory

A revision register is the backbone. Each entry should carry a revision identifier, the date it was issued, a plain description of what changed, the sheets affected, the person who approved it, and a status of draft, issued, approved or superseded. When a revision supersedes another, the older one stays in the record with its status changed, so anyone can see the history and nobody can cut from it by accident.

Illustrative scenario: a homeowner texts that the stair should move from the left side to the right after framing has been drafted. The agent logs the request as a pending revision, attaches the message as evidence, flags that the footing layout and the railing count may change, and drafts a note for the estimator. The revision is not marked approved until the estimator confirms the changes and the client signs the updated drawing. Until then, the crew sees pending next to the stair item.

The agent should never mark a revision approved because a message sounded agreeable. Approval is an event with a source: a signature, a written confirmation from the client, or a decision recorded by the person who owns it. A similar project memory approach is described in Superintelligent Agents for Engineering Firms: Project Memory.

A setup plan to run before the first live job

Work through these steps with the people who actually own each record. Most of the value comes from the decisions, not from the software.

• Define the record types: design revision, permit item, change order and inspection result.
• For each type, list the required fields and name the person who owns each field.
• Decide which sources count as evidence, such as signed drawings, portal confirmations and written client approvals.
• Set the approval boundary: the agent may draft, file internal notes and request status, while a named person approves anything affecting scope, cost or code.
• Create one current-state view the crew can read on site, showing the approved revision and every open item.
• Write the rule for superseded documents so an old drawing can never feed a cut list or a material order.
• Run a short trial on one live project, log every place where the agent and the office disagree, and fix the record or rule behind each disagreement.

Keep the first scope narrow. One deck type, one jurisdiction and one crew is enough to expose the gaps.

Tracking permit status without trusting memory

Permit status moves outside your control. Plan review, correction requests, inspection scheduling and final sign-off happen in a city or county system, in a letter, or on a phone call. The agent's job is to record each status with its source and the time it was observed, not to assume the answer.

Use a small, fixed vocabulary: submitted, under review, corrections requested, approved, inspection scheduled, inspection passed and closed. Silence from the authority is not approval. If the last confirmed status is old, the agent should flag it for a person to check rather than summarize it as current. When a homeowner asks whether the deck can be started, the agent answers from the stored status, cites the source, and says plainly when the answer depends on a check that has not been made yet.

This follows the same discipline as any retrieval setup: the answer comes from records you can point to. Anthropic's guidance on context engineering makes a related point, that curated and current context works better than everything that was ever said. For a deck job, curated means the approved revision, the open permit items and the open change orders, and little else.

Change orders: the record that protects the job

A change order is a decision with money attached, so it needs a clear lifecycle: requested, priced by the estimator, presented to the client, signed, scheduled and billed. Each change should link to the revision it modifies and to the materials or labor it adds or removes. When this chain is complete, the office can explain any invoice line and the crew knows whether a change is authorized before work starts.

The common failure is the verbal change on site. The agent can prompt the foreman to log the request the same day, asking for the location, the description and whether the client has accepted it in writing. If the change is not signed, the agent marks it open and includes it in the daily summary to the office. This does not make the agent a decision maker. It keeps open items visible so a person can decide.

If your business also prepares proposals for coatings or finishes, the approach in Superintelligence Prep for House Painters: Estimates and Color Records shows how a similar record can be structured.

Permissions and approval limits

The agent should have the smallest access that lets it do its job. That usually means read access to the project folder and approved drawings, write access to drafts and notes, and no ability to send a signed scope change, submit a permit file or alter an invoice. Access to email, storage and accounting tools should use scoped, revocable credentials, following the least privilege principle that OWASP describes in its authorization guidance.

Treat text from clients, neighbors, inspectors and vendors as data, not instructions. A message that says the deck is approved and the material should be ordered is a claim to record and verify, not a command to follow. Review permissions on a schedule, especially after staff changes, because access that suited one estimator can be too broad for a new coordinator.

Write the approval boundary into the agent's instructions and into the record design. If the boundary exists only in a prompt, it will drift over time. If it also exists as a required approval field that the system will not mark complete without a named person, it holds up much better.

How Be Human approaches this in a managed setup

Be Human starts with a first engagement that defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. For a deck builder, that usually means agreeing on the revision register, the permit status vocabulary, the change order lifecycle and the permission scope before anything is connected to a live job. Managed operations can then include monitoring of record quality, recurring evaluations that compare the agent's answers with source files, permission reviews, incident response when a record is wrong, and controlled changes when the business process changes. The goal is an agent whose memory and limits are visible to the people who run the business.

How to tell whether the agent is working

Pick a success criterion you can observe without trusting the agent's own report. Each week, sample a handful of open items and check them against the source: does the stored revision match the signed drawing, does the permit status match the portal, does the change order show the signature. Record each disagreement and its cause. Anthropic's guidance on evaluating agents recommends building tests from realistic tasks and reviewing failures rather than relying on a general impression, and that approach suits a small trade business well. The NIST AI Risk Management Framework is also a useful reference for documenting risks, owners and review points.

Good signals are unglamorous: fewer calls to the office to confirm the current revision, fewer cuts made from a superseded drawing, and fewer change order disputes that begin with I thought that was included. Review the criterion monthly and tighten it when the records change.

The free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

Can an AI agent replace the estimator or the project manager? No, and it should not be set up to. The agent keeps records current, drafts updates and surfaces open items, while the estimator prices changes and the project manager decides sequencing and scope. Keeping those roles separate is what makes the records trustworthy.

Should the agent file permits with the city or county? By default, no. It can prepare the packet, check that required sheets are attached and log the submission once a person has sent it. Submitting on the owner's behalf adds risk, so that step belongs behind an explicit approval.

What should happen when a client changes the design by text message? The agent logs the message as a pending revision, attaches it as evidence and flags the sheets and materials that may be affected. The revision moves to approved only after the estimator confirms the change and the client signs the updated drawing or gives written approval the business accepts.

Where does the project memory live, and who can read it? It should live in the business's own project file system or record tool, with access granted by role so the crew, the office and the owner each see what they need. Every entry should show its source and the person who approved it, which makes the memory useful during a dispute or an inspection.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.