Superintelligent Agents for Kitchen and Bath Remodelers: Selections
How kitchen and bath remodelers track selection sheets, change orders and client approvals in one memory layer, so each job improves the next estimate.

A superintelligent agent for a kitchen and bath remodeler is a governed assistant that reads selection sheets, change orders and client approvals from one shared memory, then tells the team what is pending, what changed and what the next estimate should check. Its usefulness depends less on clever reasoning than on a single record of decisions that outlives the project manager's notebook. Be Human builds that record first, then sets exactly what the agent may draft, flag and route.
Key takeaways
- Selections, change orders and client approvals are the three records most often lost between the showroom, the site and the office.
- The memory layer should store each decision with its source, approver, channel and the date it became binding, not only the final invoice.
- The agent drafts, flags and routes. A person approves money, scope changes and any alteration to a finished surface.
- Each closed job should leave reviewed notes on allowances, lead times and accepted substitutions for the next estimate.
- Start with one job type, one owner and one success test before connecting more tools.
Where selections and change orders get lost
A kitchen or bath job creates more written decisions than most projects of its size. The client chooses a cabinet door style in a showroom, sends a photo of a tile from their phone, and later asks by text whether the niche can move. The plumber mentions on site that the valve needs a different trim kit. Each of these is a decision, and each lives in a different place: a supplier portal, an email thread, a voice note, or the margin of a paper drawing.
The cost of that scatter shows up late. A selection is ordered against an outdated sheet, a change order is approved by someone who was not in the original conversation, or an allowance is spent before anyone notices the bathroom now has a second vanity. None of these failures needs a dramatic cause. They need a record that is easy to search and clear about who decided what.
What the memory layer should hold
Memory is not a dump of every message. It is a set of typed records that the agent can retrieve and cite, so each answer points back to a document someone can check. Keeping only the records relevant to the current question is the same discipline described in Anthropic's guidance on context engineering, and it matters more in remodeling than in most fields, because a wrong finish code costs more than a wrong sentence.
- Selection record: item, model or finish code, vendor, source document, approver, and the date the selection became binding.
- Change order record: requested change, reason, trades affected, cost review status, and client signature status.
- Approval log: the client's exact wording, the channel, the date, and what the approval covers and what it does not.
- Site note: access conditions, measurements confirmed on site, and photos referenced by filename rather than copied into the record.
- Job review: what differed from the estimate and why, written once the job closes.
A practical test for any record: could someone who was not on the job answer 'what did the client approve, and when' in under a minute, without calling the project manager? If not, the record is incomplete. Painting and finish-heavy work follows the same logic, and the color record approach in house painter estimates and color records applies directly to paint selections.
A step-by-step plan for the first workflow
Choose one workflow where the loss is visible. For most remodelers that is selection sign-off before an order is placed, or change orders on bathroom work. Then work through these steps in order.
- Name the job type and the owner who will answer questions about it.
- Collect the documents from recently closed jobs of that type, including the ones that went badly.
- Write the record fields for selections and change orders, using the terms your office already uses.
- Map who may write each record: the coordinator, the estimator, the field lead, and the client through a controlled form or a confirmed message.
- Set the approval boundary in writing: what the agent may draft, what it must route, and what only a named person may approve.
- Define one success test, such as every ordered selection having a source document and a named approver before the order goes out.
- Run the agent on past jobs in a test environment and check its answers against what actually happened.
- Enable live use for that single workflow only after the test passes, then review errors on a fixed schedule.
The same discipline applies to other trades that depend on site records. Route and bid notes in landscaping follow a similar sequence, as described in landscaping routes, bids and crew notes.
Change orders: what the agent drafts and what a person decides
Change orders are where a remodeling business gains the most from memory and where it can lose the most from a careless automation. The useful split is simple enough to print and post in the office.
- The agent may log the request, attach the message or photo, and link the affected selection.
- The agent may draft the change description and check whether the work sits inside an existing allowance.
- The agent must route to the estimator any cost difference, any effect on another trade's schedule, and any change touching waterproofing, structure or venting.
- Only a named person may send a price to the client, issue a change order, or authorize work that alters an item already ordered.
- The agent must not reverse a binding selection on its own, even when the client's new message seems clear.
Illustrative scenario: a client writes after the plumber has set the drain location, asking to replace a single-bowl vanity with a double-bowl model that has a wider footprint. The memory layer shows the original selection, the rough-in note with the drain position, and the allowance remaining. The agent logs the message, links the three records, drafts a change description stating that the drain location must be checked, and routes it to the estimator. The estimator confirms the cost and schedule effect, and the project manager sends the revised price. The agent did not price the change or promise the new vanity.
Client approvals that hold up later
A thumbs-up emoji or a message that says 'sounds good' is not a dependable approval for a change in cost. The approval record should store the exact wording, the channel, the time, and the scope it covers. When the team asks what the client agreed to, the answer should quote the message and point to the change order it refers to.
Client messages also need a boundary. A message may contain text pasted from another contractor's quote, or instructions that try to change how the assistant behaves. Treat client text as data to record, not as commands to follow. The OWASP guidance on prompt injection prevention describes this separation between instructions and untrusted content. Access should follow the role as well: the homeowner sees their own job, the subcontractor sees the schedule and the site notes they need, and only the estimator edits pricing, consistent with the OWASP authorization guidance for systems that hold sensitive records.
How each closed job improves the next estimate
The compounding part begins after the job closes. A short review, written by the project manager, captures what differed from the estimate: an allowance that was exceeded, a lead time that slipped, a substitution the client accepted, or a measurement that changed between the survey and the install. The agent turns these into candidate notes attached to the item type, for example the allowance behavior of a particular tile category.
A person reviews each note before it enters the estimating template. Unreviewed notes stay in a holding area, so one unusual job does not quietly reprice the several that follow. The same caution applies to plumbing work, where callback history is only useful when it is checked before it changes a template, as shown in plumbing dispatch and callbacks.
Be Human designs this memory layer as part of a managed service. The first engagement defines one useful role, the records it needs, the tools it may call, the approval boundaries and an observable success criterion. After launch, operations cover monitoring of the agent's answers, recurring evaluations against sample jobs, maintenance when a supplier portal or scheduling tool changes, permission reviews when staff join or leave, incident response, and controlled improvements to notes and instructions. The review rhythm follows the lifecycle thinking in the NIST AI Risk Management Framework, so the owner can see what the agent knows, where each fact came from, and who approved each step.
If you want to find which of your selection, change order or approval steps is the best first task for an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Do we need to replace our estimating software to use this? No. The memory layer sits beside the tools you already use for estimates, schedules and supplier orders, and the first engagement defines which systems it reads and which it may write to. Connections are scoped to the records the role needs, so the agent does not become a second estimating system.
Can the agent sign a change order or send a revised price to a client? No. The agent can draft the change description, route it to the right person, and track its status, but a named person sends prices and issues signed documents. The approval boundary is written during the first engagement and reviewed whenever roles change.
What happens to client details stored in the memory? Only what the role needs is stored, and retention, access and deletion rules are agreed in writing before records are loaded. Personal details that do not affect the work are left out. Where clients or data subjects are covered by specific data protection rules, the handling is checked against those obligations before go-live.
How do we know the agent is useful and not just one more place to check? The success criterion is set before launch, for example that every ordered selection has a source document and a named approver, and it is measured on real jobs. Recurring evaluations compare the agent's answers with what actually happened, and a workflow that misses its criterion is paused and corrected rather than expanded.
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.


