Data Readiness

Why it matters in construction

AI in precon uses two kinds of data: the documents it reads today and the history it uses for comparisons and forecasts. Document features can work from day one. Historical data is where most GCs find gaps. Win rates cannot be computed if losses were never logged. Staffing forecasts cannot be calibrated if nobody knows how many superintendent hours the last hospital job consumed. Sub matching breaks down when the same sub is entered four different ways.

A model given thin or inconsistent history does not refuse to answer. It answers anyway, with the same confidence, and the errors look like insight. Data readiness is the work of making sure the answers rest on something real.

How it works

Ask four questions of each data source:

1. Is it captured? Pursuits, including the ones the firm lost or declined. Actual staffing hours by project. Final cost by CSI division as well as contract value. Sub performance history, not only contact info.
2. Is it consistent? One project type list. One sub master with deduplicated names. One definition of “start date.” The model cannot tell that “TI” and “tenant improvement” and “interiors” are the same thing unless someone decides they are.
3. Is it connected? A pursuit in the CRM should link to the estimate, which links to the awarded project, which links to actual cost and staffing. Each link is a place where a forecast can learn from an outcome.
4. Is it accessible? A model cannot use a spreadsheet on someone’s desktop. Data in a system with an API can be connected to the tool.

Most firms should pick the AI use case first, then fix only the data that use case needs. Projects often stall when teams try to clean everything before they start.

Example in practice

Imagine a $90M/year commercial GC wanting to use AI pipeline forecasting. The CRM has 600 opportunities over five years. On inspection, 410 are wins or active; the 190 losses mostly sit in a “closed” status with no reason, and 80 pursuits were never entered because the BD lead tracked them in a notebook.

The precon director spends three weeks with the BD team backfilling outcomes and reasons for the last three years, standardizing eight project types down from 23 free-text values, and merging 140 duplicate owner records. The resulting dataset covers roughly 350 completed pursuits with outcomes. That is enough for the forecasting model to produce win-rate estimates by project type and delivery method that the team can check against their own memory, and the tool goes live with numbers people trust.