AI for Glazing Contractors: A Practical Guide

Hey everyone - I'm Amanda, founder and CEO of Glazier Software.

Out here in San Francisco where I live, AI is all anyone lives, breathes, and sleeps. But what's been really interesting to watch is how it's started permeating every other industry, glazing included. Forward-thinking glazing contractors are experimenting with what AI can do for their businesses, and some of them are seeing genuinely cool results.

We're in an interesting position at Glazier - we have a lens on both the glass industry and the AI industry, and we get to see where they overlap. What the good use cases are, what the bad ones are, where the technology works, and where it falls flat.

So here's what I wanted to walk through:

1. The four kinds of AI tools glazing contractors will encounter - LLMs, connectors, systems of record, and agents. What each one actually is, and what it's good and bad for.
2. Why context is the real determinant of whether AI is useful - the concept that ties everything together.
3. Where AI is actually useful in a glazing business today - real use cases, and the right tool for each one.
4. How AI takeoff actually works - because it's one of the most common AI pitches right now, and worth unpacking a bit technically.
5. What's coming next for AI in glazing - the agents we're already building.

Let's go.

1. The four kinds of AI tools glazing contractors will encounter

There are four main flavors of AI tooling you'll hear about. They often get mashed together in marketing, which is part of why this space can be confusing. They're actually different things that do different jobs.

LLMs - ChatGPT, Claude, Gemini. Large language models. You've probably used one. You type a question, it writes an answer. Under the hood, they're extremely sophisticated prediction engines, trained on an enormous amount of text. The leaps in capability over the last couple of years have come from a mix of more training data, bigger models, and better techniques for teaching them to reason through hard questions instead of just pattern-matching. They're great at writing, summarizing, explaining, and brainstorming. They're not so great at anything that requires knowing your business.

Connectors - these are bridges. A connector lets an LLM reach into one of your existing tools (your email, your accounting software, your calendar) and either read data from it or take simple actions in it. If you've seen ChatGPT offer to read your email or pull up your calendar, that's a connector. Connectors are a real step up from a plain LLM - the AI can finally see some of your stuff. But they're limited: each connector sees one system at a time. You're still stitching fragments together. The AI doesn't have a unified picture of your business.

Systems of record - a system of record is "one place where all your operational data lives." For a glazing business, that means your bids, your jobs, your time tracking, your job costing, your photos, your submittals, your billing, your POs - all living in one connected place instead of scattered across eight tools and a filing cabinet. On its own, a system of record isn't AI. It's the foundation that makes AI useful on your business. Glazier is a system of record, and we'll come back to why that matters.

Agents - an agent is AI that doesn't just answer questions, it takes action on its own. It observes what's happening, decides what to do, and goes does it. Agents are already working in other industries - customer support, sales, finance - and we're actively building them for glazing right now. More on that at the end.

Those are the four pieces. Now the concept that determines whether any of them actually help you.

2. Why context is the real determinant of whether AI is useful

An LLM is only as smart as the information it has access to when you ask a question.

Here's a subtle version of the problem. Say you upload your job costing spreadsheet to ChatGPT and ask it which jobs look like they're underperforming. It'll give you an answer. The answer will be confidently written, look reasonable, and be partially wrong - because the spreadsheet alone doesn't capture the change orders that live in your email, the material increases you agreed to verbally, or the hours your foreman hasn't entered yet. The LLM can only reason about what you hand it. It doesn't know what it's missing, so it won't tell you when its answer is incomplete.

That's the real failure mode. Not "AI didn't know my business." It's "AI saw a sliver of my business, treated it like the whole picture, and gave me a confident-sounding answer off the partial view."

The smartness of the model is almost never the bottleneck. What the model can see is the bottleneck.

This is the lens to put on every AI tool you'll ever be pitched: what does this thing actually have access to? A raw LLM has access to general internet knowledge. An LLM with a connector has access to one of your tools at a time. An LLM running on top of a system of record has access to your whole business. Same underlying AI — wildly different usefulness.

And this is where glazing runs into a real problem.

Most glazing businesses still run a huge chunk of their operations offline. Phone calls with the GC. Texts to the super. In-person conversations at the job site. Paper work orders. A whiteboard in the shop. Handwritten notes stuffed into a project folder. Even the digital side is usually scattered - an accounting tool here, a spreadsheet there, an email thread over there, photos on somebody's phone, a time card app that doesn't talk to anything else.

No AI, no matter how advanced, can reason well about your business when the data lives in eight places. This isn't a problem you solve with smarter AI. You solve it by putting your operations in one place first.

That's why a system of record matters so much, and honestly it's a big part of why we built Glazier the way we did. Glazier is the single place all of that data lives — and once it does, an LLM running on top of it can actually reason across your whole business instead of guessing from a sliver.

With that foundation in place, let's look at where each of these tools actually earns its keep.

3. Where AI is actually useful in a glazing business today

Here's how I'd match tools to jobs in a glazing business right now.

Writing and communication - use a plain LLM. Drafting a response to a tough customer email, cleaning up a proposal, rewriting a tricky change-order explanation in plain English. ChatGPT or Claude will save you ten or fifteen minutes every time. You don't need anything fancier.

Summarizing long documents - use an LLM with document upload. A 50-page contract, a long spec section, a messy email thread. Upload it, ask for the key terms and anything unusual, ask follow-ups. You go from an hour of reading to five minutes of skimming.

Quick research - use an LLM with web search, but verify. "What are requirements for safety glazing next to a stairwell?" You'll get a fast answer that's mostly right. Go check the response for the exact citation before you send it anywhere.

Cross-business insight - this is where you need a system of record. "Which jobs are slipping on margin right now? Which customers pay late? Which bids haven't been followed up in two weeks?" A plain LLM can't help you here. A connector to one tool can't either — the answers live across your business, not in any single system. This is the work that only becomes possible once your operations live in one place.

Takeoff - this is a category of its own, and it deserves a closer look.

4. How AI takeoff actually works

"AI takeoff" is one of the most common AI pitches you'll hear as a glazing contractor, so it's worth going a layer deeper on how the technology actually works — because there are a few different technical approaches under the label, and they give you very different results.

OCR-based approaches. OCR - optical character recognition - is the technology that reads text off a page. It's been around since the 1990s. What it does is turn a picture of text into actual text characters. It's a genuinely useful building block, but on its own it falls short for takeoff. It can tell you what numbers and words appear on a drawing, but it doesn't understand what any of them mean. "12'-0"" is just six characters to it. It doesn't know that's a dimension, what it's measuring, or which elevation it belongs to.

LLM-only approaches. Other approaches try to use a language model directly on drawings. The problem is that LLMs are trained overwhelmingly on text, not on architectural drawings, so on their own they tend to see images imprecisely - miss small elements, hallucinate dimensions, struggle with dense or unusual sheets. They can work on clean, simple drawings and get inconsistent fast as real-world complexity goes up.

Computer vision plus document parsing. This is the approach we're building on at Glazier. Worth breaking down what each half does.

Computer vision models are trained specifically to understand images - in our case, architectural drawings. The model learns to recognize shapes, group related elements together, and distinguish drawing elements (an elevation, a panel, a mullion) from annotations (labels, leader lines, dimension strings). It understands spatial relationships - that this dimension sits next to that elevation, that this label points to that opening.

Language models then come in on top of that to reason about what's been extracted. Which label goes with which elevation. Whether a dimension is a width or a height. How the pieces fit together into a coherent takeoff. That kind of reasoning is what LLMs are genuinely good at - when they're given structured information from a vision model to reason about, rather than being asked to look at a raw drawing cold.

Layered together, computer vision and LLMs cover each other's weaknesses. Neither one alone can do takeoff reliably. Together, they can.

5. What's coming next for AI in glazing - the agents we're building

Everything I've described so far is AI that answers questions or extracts data. You ask, it responds.

The next wave is AI that acts - agents. This isn't science fiction or five years out. The technology is real. It's running in customer support, sales, and finance at other companies today. And here's the thing: we're actively building agents for glazing at Glazier right now, on top of the system of record that's already in place for our customers.

A few of the agents we're building:

A job performance agent. Watches every active job as it runs - tracking hours logged against the estimate, material usage, stage completion, and the pace at which each system on a project is actually getting installed. When a job's pace starts drifting in a way that would put its margin at risk, the agent flags it to the PM with the specifics - which system, how much slippage, what the projected finish looks like if the pace holds. Instead of finding out at close-out that a job lost money, you find out when there's still time to intervene.

A collections agent. Looks across your AR every morning. Flags the customers who are drifting past terms, cross-references their payment history, drafts the follow-up email in the right tone for each one, and surfaces it for you to approve. You approve or tweak, it sends. What used to take a half-day of chasing becomes a ten-minute morning review.

A bid follow-up agent. Watches your open quotes. When one has been sitting without a response for longer than your usual close cycle, it drafts a follow-up note with context from the original bid. Deals that would have gone cold get a nudge at the right moment.

None of these are chatbots. They're operational layers that run alongside your business - watching, deciding, drafting, and surfacing the work that needs your judgment.

Here's the part that matters for you today: an agent can only act on data it can see. Every one of these agents depends on the system of record underneath them. If your business is still running across eight disconnected tools and a filing cabinet, an agent has nothing to work with — no matter how good the AI behind it.

How to prepare your glazing business for AI

The glazing contractors who are going to do well over the next decade aren't the ones who buy the flashiest AI tool tomorrow. They're the ones who get their operational data into one place first - so that when the AI layer and the agent layer land in full, they're ready to use them.

The contractors still running everything across disconnected systems a few years from now won't be able to adopt any of this, no matter how much they want to. The foundation won't be there.

So if you take one thing from this, take this: start with the boring layer. Get your operational data into one system. That's the work that compounds. Everything on top of it - the AI chat, the takeoff, the agents - only works if the foundation underneath is solid.

I'm genuinely excited about where this is headed for our industry. If any of this resonates and you want to talk, you know where to find me.

- Amanda

‍