Using AI Agents to Research Materials for Pre-Construction

Over the weekend, I decided to tackle a project would try to automate some part of the estimation process. To cut to the chase, it worked really well.

The aim was to see if AI could automate material research tasks that estimators usually perform on public tenders. Specifically, it involved using the specification and a search engine to get a list of acceptable manufacturer and links to suitable product pages. It’s not the most cumbersome part, but it adds time to an already constrained estimating workflow.

Here is the Airtable database I used to start the process.

What I Tried

I played around with AI agents to automate the search for specific construction materials: vapor barriers, anchor bolts, and medical support systems. The idea was for these agents to not just identify materials that fit the project specs but also to pull up product pages and pricing info.

I wrote a Python script that leverages Microsoft's Autogen. This script essentially ingests an Airtable database with materials and facilitates a group chat among three AI entities I set up for this task: an Estimator, a Director, and a Research Manager.

• The Estimator dug into the project specifications, used Google searches, web scraping, and summarization to get relevant information.

• The Research Manager took on the role of coordinating the search for pricing and URLs.

• The Director oversaw the entire operation, making sure everything was properly recorded and kicked off the process.

The results were surprisingly good. Our AI team found the relevant information in the specification easily and then found the corresponding product pages online shortly after.

Here is the appropriate section from the specification. The correct answer for Underslab Vapor Barrier is Permiator by W.R.Meadows.

And here is the output from our AI group chat. Here is the group manager creating tasks for the estimator.

Here are the results from the Estimators research.

And here we see the URL for the product page. https://www.wrmeadows.com/perminator-underslab-vapor-barrier-retarder/

For pricing, we did run into roadblock though. In fact, our AI seemed to appreciate this point. It said “As common in the construction materials industry, pricing information is typically not listed directly on the manufacturer's website and requires direct contact with the manufacturer or distributor for a quote. Therefore, for the most up-to-date unit costs, the Estimator should contact Stego Industries and W.R. Meadows directly.”

The Takeaway

When it came down to it this experiment worked. The economics were also interesting: it cost us about $0.12 to run our three agents (using GPT-4 preview) and took around 90 seconds. Putting this in perspective, a human doing the same job might take about 5 minutes and cost roughly $7, given an $80 per hour fully burdened rate. This also took less than an hour to set up.

While drawing takeoffs are a lot more complicated, I believe we will see a fully automated estimate performed within the year.

And while this was more of a proof of concept than a full-fledged system ready for prime time, the potential for saving time and reducing costs is pretty clear. It's a small step towards smarter, more efficient pre-construction phases, and I'm curious to see where this can go with a bit more development.