Superintelligence Prep for Kitchen Remodelers: Selections and Scope
Keep kitchen remodel selections, appliance orders and scope changes in one job memory so an AI agent works from settled decisions instead of reopening them.

Superintelligence prep for a kitchen remodeler means giving a capable AI agent one owned job memory: the signed scope, selection sheets, appliance orders, change orders and open questions, each with an owner, a date and a status. Without that record, an agent will reopen settled decisions, such as a cabinet finish or a range model, because the newest message in a thread looks like the current truth. The practical fix is a clean source of record and clear limits on what the agent may change, not a more ambitious model.
Key takeaways
- Treat each approved selection as a dated decision with an owner, not as a line buried in a chat thread.
- Link every appliance order, supplier confirmation and delivery notice to the selection it fulfils.
- Record each scope change as a change order that states what it replaces and what is still open.
- Let the agent read the job memory widely, but let it write only drafts and flags.
- Review the memory at each milestone so the agent works from what the client actually approved.
Why kitchen jobs keep reopening settled decisions
A kitchen remodel produces decisions over several weeks: layout, cabinet line, countertop slab, sink, faucet, range, hood, dishwasher panel, tile and lighting. Many of them are made by phone, by text, during a showroom visit or from a photo taken on a site walk. Weeks later, a client asks whether the pantry will have pull-out shelves, and the person answering scrolls back through the thread and finds a message about a different pull-out model from before the cabinet order was revised. Nobody made a mistake on purpose. The record simply has no hierarchy, so every message carries roughly the same weight.
An agent that reads everything will make the same error faster. It can summarize a long thread well, but it cannot know which message was superseded unless the job memory says so. If the most confident message is the most recent one, the agent will treat an old discussion as a current instruction and draft a reply, a supplier email or a schedule change based on it. The problem is not intelligence. It is missing structure around decisions.
What belongs in one job memory
A usable job memory for a kitchen remodel holds the facts that the business would have to defend later. Keep these records together, with each entry tied to a room or a selection:
- Scope baseline: the signed proposal with its inclusions, exclusions and allowances, dated at signing.
- Selection sheet: each product choice with model, finish, quantity, dimension check, approving person and approval date.
- Order register: purchase orders, supplier confirmations, deposits recorded as paid, and the selection each line fulfils.
- Change orders: the requested change, the estimator's impact in the project's own terms, the client decision and its status.
- Site facts: field measurements, access notes and photos tied to a specific wall, opening or fixture.
- Open questions: items waiting on a client, a supplier or a trade, each with an owner and the date the question was asked.
- Communication log: who said what, and whether the message changed a decision or only confirmed one.
The same logic applies in other trades that track specified materials against an approved baseline. A painting company, for example, needs the same separation between approved color records and draft estimates, as described in the house painters job memory guide.
How to build selection sheets that stop the argument
A selection sheet is useful only when it can be read without context. The goal is that anyone, including an agent, can tell in a few seconds what was chosen, who chose it and whether it is still in force. Use this sequence:
1. Freeze the baseline. Attach the signed scope and mark which selections were included at the time of signing, so later additions are visibly new. 2. Give each selection one line: item, model number, finish, quantity and the dimension it depends on. 3. Record who approved it, in what form (signature, email, showroom sheet) and on what date. 4. Assign exactly one status at a time: draft, approved, ordered, delivered or changed. 5. When a selection changes, do not overwrite the old line. Add a new line that references the old one as superseded, so the history stays readable. 6. Attach site verification to any item that depends on a dimension, such as a range opening, a hood clearance or a dishwasher panel height, and mark the verification date.
The sixth step prevents a common failure. A selection can be correct as a product and still be wrong for the opening it has to fit. Keeping the verification next to the selection lets the agent flag a mismatch before an order is placed.
Appliance orders and scope changes without rewriting history
An appliance order carries its own facts: model, finish suffix, supplier, confirmation reference, the delivery date as communicated by the supplier, and whether the order can still be changed. Supplier messages will keep arriving after the order is placed. The memory should record the confirmation as the source for that order, and every later supplier notice as a new entry linked to it. A notice should never silently replace the original line.
Illustrative scenario: a client approves a built-in oven during a selection meeting, and the order is placed the same week. Several weeks later, the supplier writes that the chosen finish has been discontinued and offers a close substitute. The memory holds three linked entries: the original selection, the order confirmation and the supplier notice. The agent can draft a question to the client with the photos and dimensions attached, but the status stays at ordered until the client approves the substitute in writing. The agent does not switch the selection on its own. A similar order-tracking discipline is used for rented equipment returns, as outlined in the equipment rental guide.
Retries deserve attention too. If an automated process repeats a purchase request because a response was lost, the ordering system should recognize the repeat as the same request rather than creating a second order. Payment platforms describe this as idempotency, and the same principle belongs in any order workflow connected to an agent, as explained in [Stripe's documentation on idempotent requests](https://docs.stripe.com/api/idempotent_requests).
Scope changes follow the same path. A client asks to move a sink wall or add a pantry pull-out. The request is logged, the estimator states the impact, the client decides, and only then do the selection and order lines change, each with a reference to the change order that caused them. The agent can prepare the change order summary and the client question. The estimator and the client own the decision.
Setting approval boundaries for the agent
An agent should have the access each task needs and no more. Standard authorization guidance, such as the [OWASP Authorization Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html), points to the same principle: grant the minimum access per action and check it at the point where the action happens. For a remodeling business, a practical decision guide looks like this:
- Read freely: the job memory, the signed scope, the order register and the communication log.
- Draft and flag: change order summaries, client questions, supplier follow-up messages and schedule conflicts.
- Requires a named person: any change to an approved selection, any new order, any change to contract scope.
- Never automatic: placing or cancelling an order, changing a deposit record, or confirming a delivery date to a client.
Write these boundaries into the agent's instructions and into the system permissions, so a rule that exists only in a document cannot be bypassed by accident. Keep the context the agent receives focused on the current job and current decisions. Long, undifferentiated histories tend to produce confident but stale answers, a risk that is well described in [Anthropic's guidance on context engineering for agents](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents).
How a managed setup keeps the record current
Be Human designs this kind of job memory as part of a first engagement: one defined role, the memory and tools that role needs, the approval boundaries, and an observable success criterion such as every open selection having an owner and a status. In managed operations, the memory is reviewed at milestones, permissions are checked against the current staff, and a sample of the agent's drafts is compared with the record so drift is caught early. The work is remote, available in English or Portuguese, and the business keeps every decision about what the client approves.
If you want to find the first task worth handing to an agent in your remodeling business, the free Be Human AI diagnostic is a plain place to start.
Frequently asked questions
Can an AI agent replace the project manager on kitchen remodels? No, and it should not be asked to. The agent can keep the record consistent, prepare drafts, flag conflicts between a selection and a site measurement, and answer routine status questions from the memory. Decisions about scope, orders and client approvals stay with named people, who review what the agent surfaces.
Should selection sheets live in the CRM or in a shared drive? The location matters less than having one source record that people update and the agent can read. Choose the system your team will actually maintain, then link documents, photos and confirmations to the selection they support. Avoid keeping the approved version in one place and the working version in another.
What happens when a supplier substitutes an appliance after the order is placed? The original selection and order stay in the record, and the supplier notice is added as a linked entry. The agent drafts the question for the client with the dimensions and photos attached, and the status remains ordered until the client approves the substitute in writing. Nothing changes silently.
How do we know the agent is not reopening settled decisions? Define that as a success criterion before the work starts, then check it. Sample the agent's drafts against the job memory at regular intervals, review the logs of what it read and proposed, and treat any answer that contradicts an approved selection as an incident to investigate. Evaluation practices for agents are described in [Anthropic's guide to evals for AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents).
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


