When to use AI in construction estimating

Estimating is a deep practice. Many plates are spinning at once; planning what parts and pieces go into building your systems, reading and reconciling large documents (specs, plans, pricing narratives, vendor and sub bids), understanding what it actually takes to build the job in the field, and all of it balances in your head while making the bid and getting a number. The table stakes of this profession are high. Now that AI is in the mix, a new skill has surfaced that may be worth more than any of the rest: the skill of knowing what AI should do, not just what it can do.

When to be cautious

There are many situations where AI appears capable of performing a task, but tread carefully as one may often get caught in the "probabilistic-output-acting-as-deterministic" trap. Avoid delegating these kinds of tasks to non-humans:

• counting quantities (if using large language models)
• real-world interpretation of scope and spec requirements
• handling edge cases and one-off decisions
• adjusting field labor based on prompting
• generating pricing against project details, regional anomalies, and the macro-economic environment
• vetting and approving vendor and sub bids

Large language models may help get your head around many of these tasks, but actually performing these tasks and bidding off of the responses may not be the best idea in the long term. LLMs, being non-deterministic in nature, will attempt to make mathematical calculations even though they cannot guarantee an accurate outcome. While newer capabilities do have connection to the internet and other hard sources, there is still room for misinterpretation of data. In short, if it involves hard numbers or real-world context beyond what you can stuff into your prompt, use AI only for a starting point, but ensure you have real eyes looking over the work.

When to be adventurous

There is a limited but powerful intersection of work AI can and should take over from construction estimators. AI is excellent at digesting large volumes of text and recognizing patterns in text and images. Tasks such as:

• counting quantities (if using computer vision models)
• Generating spec summaries, and flagging common scope gaps or obvious missing information, typos, etc
• Detecting dimensions, symbols, and scheduled items
• Organizing and grouping sheets
• Surfacing recommendations from the information in context (recommendations, not decisions)
• repetitive data entry and sanitation tasks

You may have noticed we mentioned "counting quantities" in both sections, and this is an important point: much of the advantage in using AI lies in knowing which facet of the technology to use and when. While probabilistic large language models are not a trustworthy tool when finding visual quantities, computer vision models excel at recognizing repeated patterns of pixels in plans. Ensure your tools that claim to be AI are provisioned with these computer vision models.

Ensure your tools hit the mark

This awareness of task responsibility informs our philosophy of "AI when you need it, human when it matters". We built AI tools that assist the estimator instead of replacing them. We consider good data capture and management as the foundation of the product, not AI. We give the estimator the choice of how much or how little they reach for the AI lever. Completely manual is still an option (and faster still than legacy options).

Speed is a vanity metric if the destination is a 70% accurate, half-baked estimate with no downstream use that takes longer to validate than it would have doing the task manually. Anyone these days can wrap Claude in a shiny construction-application veneer—and many try. We took the road less traveled, and we think you should too.