AI in Construction: Automating Daily Site Reports with LLMs

Every morning you spend an hour copying photos, notes, and weather data into a report that nobody reads until something goes wrong. By the end of this article you’ll have a working pipeline that turns a folder of site pictures into a formatted PDF report, all driven by AI and no‑code glue.

Why Manual Site Reports Kill Profit

Most small contractors still rely on a shared Excel sheet or a Word template that gets filled out after the day is done. The process is repetitive, error‑prone, and eats up billable hours that could be spent on‑site or bidding new work. When you multiply that hour by five days a week and fifty weeks a year, you’re looking at 250 hours lost — roughly $12,500 at a $50/hour rate.

That’s not just wasted time; it’s a direct hit to your bottom line. The good news is that the pieces to automate this already exist and cost less than a monthly coffee subscription.

How do you turn a photo log into a structured report without typing?

You start with a simple trigger: a new image dropped into a designated Google Drive folder. From there you run two steps. First, a vision model extracts any visible text (like safety signs, equipment IDs, or handwritten notes) and describes the scene. Second, a language model turns that description plus the date, weather API data, and any voice memo transcript into a filled‑out report template.

Here’s the exact prompt I use with GPT-4o vision (the model that accepts images). I keep it in a text file so I can tweak it quickly.


You are a construction site reporter. Given the attached photo, list any visible equipment, safety hazards, and progress markers. Then write a short paragraph summarizing the day’s work in a formal tone. Include the date {date} and weather {weather}.


The {date} and {weather} placeholders are filled by a simple Make (formerly Integromat) scenario that pulls the file’s modified timestamp and calls a free weather API (Open‑Meteo) for the site’s zip code.

What Most Guides Get Wrong

Many tutorials tell you to start with a fancy no‑code AI agent framework and then try to force every step into a visual builder. The result is a brittle flow that breaks when the folder name changes or the image format is odd. They also ignore cost: they suggest using a paid OCR service that charges $0.02 per image, which adds up fast when you’re processing hundreds of photos a week.

Instead, keep the AI calls minimal and cheap. Use the free tier of Azure Computer Vision for OCR (first 5,000 transactions/month are free) and only call GPT-4o vision when you need a narrative summary. That keeps the per‑report cost under $0.003.

Concrete Named Example: Cost and Tools

Let’s put numbers on the table. I run this pipeline for a small renovation crew that produces about 20 site photos per day.

• Google Drive folder trigger – free (part of Google Workspace you already have).
• Azure Computer Vision OCR – $0.00 per image (free tier).
• GPT-4o vision call – $0.015 per image (based on $0.015 per 1,000 tokens, ~150 tokens per image).
• Make.com scenario – $29/mo for the Core plan (covers 10,000 operations).
• Weather API (Open‑Meteo) – free.
• PDF generation via Make’s built‑in module – included.

Daily cost: 20 images × $0.015 = $0.30. Monthly: $0.30 × 20 workdays = $6.00. Add the Make.com subscription: $29 + $6 = $35/mo. That’s less than half what you’d pay for a single seat of a generic construction AI platform that promises “end‑to‑end automation” but delivers a clunky UI.

I think $29/mo for Make.com is fair; the AI usage is the real variable and stays tiny if you limit calls to vision only when needed.

How to Debug When This Breaks

First, check the Make.com scenario log. If the Google Drive module shows “File not found”, verify the folder ID hasn’t changed after a Drive reorganization. Second, look at the Azure OCR output. If it returns empty text, the image might be too blurry or taken in low light; you can add a preprocessing step that rejects images under a certain brightness threshold (a simple ImageMagick command). Third, if GPT-4o returns gibberish, the prompt may be too long or the image size exceeded the model’s limit; resize to max 1024px on the longest edge before sending.

One concrete gripe: the Azure Computer Vision free tier throttles after 5,000 calls, and the error message is vague (“Quota exceeded”). I missed that once and spent an hour wondering why the OCR step suddenly returned nothing. Adding a retry‑with‑exponential‑backoff module in Make.com solved it, but the docs don’t mention the limit clearly.

One concrete love: the ability to feed the same image to both OCR and vision in a single Make.com bundle saves a duplicate download and cuts latency by almost half. It’s a small thing, but when you’re processing 20 photos a day it adds up to real time saved.

Putting It All Together: A Minimal Viable Pipeline

Here’s the step‑by‑step you can copy into your own Make.com scenario.

1. Create a Google Drive folder called “Site Photos”.
2. In Make.com, add a “Watch Files” module pointed at that folder.
3. Add an “Azure Computer Vision – OCR” module, map the file content, store the extracted text in a variable.
4. Add an “HTTP – Get Weather” module (Open‑Meteo) using the site’s latitude/longitude from a static map or a Google Sheets lookup.
5. Add a “GPT-4o Vision” module, attach the image file, pass the OCR text, date, and weather into the prompt shown earlier.
6. Add a “PDF – Create Document” module, map the GPT‑4o output into a pre‑designed template (you can design the template in Google Docs and export as PDF).
7. Add an “Email – Send Email” module to deliver the PDF to the project manager.
8. Set the scenario to run every 15 minutes or trigger on file creation.

That’s it. Once the scenario is live, drop a photo into the folder and you’ll get a polished report in your inbox within a minute.

We cover this in more depth elsewhere — deeper coverage of AI agent platforms.

If you’d rather skip the build and deploy a working version in an afternoon, we’ve packaged this workflow as a blueprint at deepusecase.com/vault.