Where construction actually sits in the AI adoption curve.

In March 2026, Anthropic published new research mapping roughly 800 US occupations against the way people are actually using AI in the wild. Construction and the skilled trades landed in the least AI-exposed 30 percent of the workforce, alongside cooks, mechanics, and bartenders. That is the right answer for the field work itself. It is the wrong answer for everything else inside a construction business.

01The Anthropic study, in brief

The paper is Labor market impacts of AI: A new measure and early evidence, by Maxim Massenkoff and Peter McCrory, published March 5, 2026. It introduces a metric the authors call "observed exposure," which combines three things: the O*NET database of around 800 US occupations and their constituent tasks, prior theoretical estimates of how exposed each task is to large language models, and Anthropic's own data on what people actually use Claude for in production.

The intuition is that earlier work measured what AI could in principle do. The Anthropic paper measures what AI is actually being used to do, occupation by occupation. That distinction matters. The gap between theoretical capability and observed usage is enormous, and the paper quantifies it.

02Where the data places construction trades

When the authors rank US occupations by observed AI exposure, the highest-exposed 30 percent of the workforce includes computer programmers (with about 75 percent of their tasks covered by Claude usage), data entry keyers (around 67 percent), customer service representatives, and financial analysts. These are the jobs LLMs have actually moved into.

The least-exposed 30 percent looks completely different. It includes cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers, and dressing room attendants. Construction trades sit in this same bucket. The pattern is consistent: jobs that require physical presence, manual skill, and in-person service are the ones LLMs have not meaningfully reached. They cannot be reached by a chat window.

For the field work itself, this is straightforward. AI does not pour concrete. It does not frame a wall. It does not run conduit, install a fixture, or inspect a foundation. By Anthropic's data, the trade jobs themselves are among the least automatable occupations in the US economy, and there is no signal in the labor market that this is changing.

03But a construction business is not just trade work

A construction business is not a single occupation. It is a portfolio of occupations stacked under one roof. The owner is doing project management, customer service, financial analysis, document handling, scheduling, and data entry, on top of running jobs in the field. The bookkeeper, the office manager, and the estimator are doing those same tasks full-time.

Look at that list against Anthropic's most-exposed occupations again. Customer service representatives. Financial analysts. Data entry. Computer and math work. Those are exactly the categories where AI is doing real work today. Inside every construction business is a quieter set of office tasks that map directly to the highest-adoption use cases in the entire labor market.

The implication is the opposite of "construction is safe from AI." The field work is. The office work that surrounds it is not, and is in fact one of the highest-leverage targets for AI in any small business right now. The story for construction SMBs is not whether AI is coming. It is which side of the business it is changing first.

04Augmentation, not automation, is the actual pattern

Anthropic's methodology weights two kinds of AI use differently. Tasks where AI completes the work end-to-end count for full credit. Tasks where AI assists a human count for half. The framing is deliberate: augmentation is the more common pattern in the data, not full automation.

The labor outcomes back this up. The paper finds no systematic increase in unemployment among workers in highly exposed occupations since late 2022. The closest negative signal is a roughly 14 percent drop in job-finding rates for young workers entering exposed occupations, and the authors describe even that finding as "just barely statistically significant." Three years into widespread LLM availability, the pattern is not job loss. It is workers in exposed roles getting more done.

Translated to a construction business: AI is not, today, replacing the bookkeeper or the estimator. It is making each of them faster at the parts of their job that match LLM-friendly task profiles. The same headcount handles more jobs, more quotes, more invoices. The hours saved come out of the evening and weekend office work, not out of payroll.

05The capability-to-usage gap is enormous

One of the most striking numbers in the paper: Claude currently covers about 33 percent of tasks in the Computer and Math category, despite a theoretical feasibility ceiling closer to 94 percent. AI usage is running at roughly a third of what is technically possible, in the category most exposed to AI in the first place. The gap between what AI could do and what AI is doing is huge.

If that gap is wide in software and finance, the gap in the office side of construction is wider still. Construction office work has historically been served by enterprise project management suites that small contractors cannot afford and do not need, leaving most of the actual paperwork in spreadsheets, email, and a phone camera roll. The technical capability for AI to handle the bulk of estimating, quote drafting, invoice generation, expense categorization, and document summarization has been there for two years. The tooling that delivers it to a 4-person to 50-person construction business has not.

06Why SMB construction adoption specifically lags

Three barriers, none of them technical:

Pricing was built for enterprise general contractors.

Most construction technology, AI-enabled or otherwise, has been designed for top-of-market GCs with hundreds of millions in revenue and dedicated technology teams. The per-seat pricing, multi-month implementation timelines, and integration consulting that come with those tools are non-starters for a 10-person business.

Setup overhead is the hidden cost.

An owner of a small construction business does not have a procurement team, an IT manager, or three weeks to implement a new platform. The tool that wins is the one that delivers value in the first hour of use. Most enterprise construction software does not.

Generic AI does not speak construction.

A general-purpose LLM does not know the difference between a quote and an estimate the way a contractor does. It does not know your local permit conventions, your typical payment cycles, your trade-specific line items, or the difference between retention and a holdback. Useful AI in this industry has to be built for the work, not adapted from a general chat interface.

07What good AI in construction looks like

The pattern across the construction SMBs we work with is consistent:

• Built into the workflow, not a separate chatbot. The AI shows up where you are already working, drafting a quote, generating an invoice, summarizing a job, instead of asking you to context-switch into a chat window.
• Specific to the trade, not generic. The system knows what a takeoff is, what a change order looks like, what a 1099-NEC is for, and how a typical payment cycle moves. Specificity is what makes the output usable instead of approximate.
• Operates on your data, not somebody else's. Your jobs, your customers, your historical pricing, your documents. Not a generic model trained on the open web pretending to know your business.
• Reviewable, undoable, controllable. The owner stays in charge. The AI proposes; the owner approves. This matches Anthropic's data: augmentation is what is actually happening in the labor market, not full automation.
• Priced for the size of the shop. If a tool costs more than a part-time admin, it has to deliver more value than a part-time admin, quickly and obviously.

08How we think about AI in Sitetraq

The Anthropic data lines up with what we have seen building Sitetraq for construction SMBs. The trade work in the field stays where it always was, and the data suggests it will stay there for a long time. The office work behind it, quoting, invoicing, document review, scheduling, change orders, expense categorization, is where AI is doing real work today and where the time savings show up in a contractor's week.

Sitetraq treats this exactly the way the data suggests it should be treated. AI is embedded inside the office workflows that already exist, not as a separate AI mode bolted onto the side. The estimator gets a draft to review, not an answer to accept blind. The owner keeps control of every quote, invoice, and document decision. The hours saved come out of the night and weekend office work, not out of the field, and not out of payroll.

That is also the bet at Shinpo Capital more broadly. The most useful AI for small businesses is not dramatic. It is invisible. It is the difference between a quote that takes two days and a quote that takes ten minutes, the difference between a Sunday spent reconciling invoices and a Sunday spent doing literally anything else.

Read more

• Data before AI: how small businesses make AI actually work
• Top operational challenges for small business owners
• Construction quote guide: how to write quotes that win
• Construction invoice template (2026): 11 elements every contractor needs
• Sitetraq, the construction operations dashboard
• Our mission, software that runs your business