How an Estimating Team Cleared a 40-Project Backlog with AI

The client's estimators built estimates in claims estimating software, but projects ran in separate construction management software. An analyst moved each estimate across by hand, mapping every line item to the right cost code, and more than 40 projects had queued up behind that step. Keeping pace would have meant hiring two more business analysts.

BaseForge built an AI classification layer that maps estimate line items to the client's cost codes automatically. A smaller Gemini model summarizes each line item, and a larger Gemini reasoning model chooses the best match from the client's full corpus of cost codes. Results are cached, so the many repeat line items don't call the models again. A web application, signed in through the client's Azure tenant, pushes the coded estimates into the construction management system, Salesforce and the other estimating platforms the team uses.

BaseForge started with a two-week proof of concept. The team sat with the analyst doing the mapping by hand to capture the business logic, wrote a script to transform the estimate data, and layered in the AI step that classifies each line item into a cost code. That proof of concept cleared the 40-project backlog. Gemini was chosen after it performed best in internal evaluations, with a smaller model handling summaries and a larger reasoning model making the final cost-code choice, and caching added for the line items that repeat across jobs. The team then built a web application on the client's Azure single sign-on that pushes estimates into every system that needs them, and later connected Salesforce and other estimating platforms. The work ran across the year, but the build itself took about eight weeks: four for the web app, one for testing and three for new systems and logic.

Contractors, restoration firms and estimating teams that move estimates between claims or estimating software and construction management systems, where line items have to be coded by hand and the backlog grows faster than analysts can be hired.