Before You Ask AI About Margin, Fix the Records Behind It

Ask an AI assistant which homes are losing margin and it may produce a convincing answer in seconds. The harder question is whether it used the right sales price, the current scope and the complete expected cost.

If accounting has one version of the job, purchasing has another and construction is working from a revised plan, the analysis begins with a reconciliation problem. A fluent explanation can make that problem less obvious.

Builders should be demanding about the evidence behind an answer. “Margin declined” is useful only if someone can trace the decline to specific changes and determine what to do about them.

Define the record before defining the prompt

Each home needs a consistent identity across the contract, selections, budget, purchase orders, schedule and accounting. Dates and statuses need clear meanings. Approved changes should be distinguishable from proposals. Actual costs, remaining commitments and uncommitted estimates should be separately identifiable.

That doesn't require every employee to use the same screen. It requires the systems and people involved to agree on which record governs each fact and how updates move through the business.

Preserve history, too. To explain why a forecast changed, the system needs the previous forecast and the intervening decisions. Overwriting yesterday's number with today's number leaves little to analyze.

Begin with a verifiable management question

Choose a narrow question such as: which homes had a material increase in expected final cost this week, and which approved transactions explain it? Require the response to identify the home, amount, date and supporting records.

Have the responsible manager check the answer against known cases, including incomplete records and unusual changes. The system should flag missing information rather than quietly interpret a blank as zero. Keep the distinction between a recorded fact and an inferred explanation visible.

Measure whether the analysis helps the team find a problem sooner or reduce review work. An impressive paragraph is a weak success criterion if the manager still has to reconstruct every calculation.

Put approval around consequential actions

Reading and summarizing records is different from changing them. A proposed budget revision, vendor payment or customer commitment needs permissions appropriate to the action. Define when human approval is required and retain a record of the proposal, approval and resulting change.

The same principle applies to access. An assistant should operate within the user's authorized information and actions. Convenience isn't a reason to expose another company's records or remove an existing approval requirement.

The opportunity is substantial: less time collecting information and more time resolving exceptions. Start with a trustworthy operating record and a question the business can verify. That gives the builder a basis for expanding AI's role when the results justify it.