Last week I was reviewing a maintenance proposal that looked ready to send. It was clean, branded and well written. The pricing table showed $30,000. A paragraph farther down promised the client the work would not exceed $28,000. Both numbers couldn't be right.
An AI assistant had helped assemble the proposal, and it had done exactly what makes these tools so useful and so dangerous at the same time. It produced something polished enough to trust before it had earned my trust.
That proposal is a pretty good summary of my first year running a landscape company with AI. Some of it gave me back entire evenings. Some of it created work I had to undo. And almost none of it looked like what I was worried about after watching the news.
I've written in this column about buying Geared for Growing, a 40-year-old landscape company in Northern California. I spent my first year getting the business out of people's heads and into systems everyone could see. There was a second experiment running in the background at the same time. While our crews were building landscapes and handling maintenance, I was in the office testing whether AI could do anything useful for a midsize landscape company like mine.
Proposals Were the First Domino
Our estimator is excellent. He builds his cost breakdowns in a spreadsheet with unit costs, labor hours and material lines. The spreadsheet is mostly right, but you wouldn't want to put it in front of a client. Turning that internal estimate into a clear proposal used to involve a full morning of copying, pasting, formatting and proofreading so a homeowner could understand it.
AI earned its keep anywhere the real thinking had already been done by a human.
Now I give an AI assistant the spreadsheet, our proposal template and examples of proposals we've actually sent. I ask it to flag inconsistencies, explain the scope in plain English and format everything to be client-facing.
I’ve found that AI is very good at that type of translation work. It can turn an estimator's line item into client language, organize a large-scale job, and find a quantity that doesn't match somewhere else in the document. But it also misses things, like two different prices in the same proposal. Every number, date and promise still gets checked before it goes out.
AI hasn't become our estimator. It's that a finished estimate spends less time sitting in my drafts folder while the client is still excited about the project.
The pattern became clear after a few months. AI earned its keep anywhere the real thinking had already been done by a human.
Building a World Model of the Company
The second place I see real potential is bigger than any one proposal. I'm trying to build what I think of as a world model of our company: a referenceable record of what has happened across the business and the context behind it.
A world model built on incomplete information is just a more sophisticated way to be confidently wrong.
Every client interaction, proposal, CRM entry, time entry datapoint, and meeting note creates another piece of that picture. If those records live in a shared drive and are organized so an AI assistant can reference them, the tool can sit alongside me as the owner and help me see the full story instead of one document at a time.
Before a client call, it could pull together the last proposal, recent meeting notes, open follow-ups and the history in our CRM. If two records conflict, it can point that out. If we discussed something three months ago and never closed the loop, it can bring it back to the surface. That doesn't sound as exciting as a robot designing a landscape, but for an owner with a dozen decisions each morning, it may become even more valuable.
The real bottleneck in a growing company is often not a lack of information but hat the information is scattered across inboxes, paper documents, software and people's memories. I don't need AI to make every decision for me. But the more I can digitize, the clearer decision making I can have.
This only works if the company gets their information online. We use a tool to track project schedules, handle time entry, invoicing, and the CRM. We even record our crew meetings. We know that if it’s online, then it’s good data — because AI can't reference a meeting that was never documented. A world model built on incomplete information is just a more sophisticated way to be confidently wrong.
Where it Doesn't Work
That brings me back to the proposal with two prices. Every AI failure I've seen has one thing in common: the tool sounds exactly as confident when it is wrong as when it is right.
Nothing with a price, date, scope or promise goes to a client without a person checking it against our own records.
If I give it an incomplete estimate, it doesn't lean into my office and say, “Lee, I need the missing information.” It tries to be helpful. It fills the gap. The invented answer may be perfectly written, which makes it more dangerous than an obvious mistake.
It also doesn't have the judgment of the people doing the work. It isn't standing on the property looking at plant layout, checking the soil or noticing that the stairway access will slow the crew down. It doesn't know from a client's tone that a small detail matters more to them than the largest line item. It can organize documented knowledge, but it hasn't earned the field judgment of our foremen.
Nothing with a price, date, scope or promise goes to a client without a person checking it against our own records. I treat AI like a new office hire in their first week. It's fast, surprisingly capable and able to absorb a lot of information, but not yet totally trusted.
From my seat, I'm very pro-AI. I think it will help landscapers respond faster, preserve knowledge that used to live in one person's head and understand our businesses more clearly. But speed and context are not the same as judgment. AI doesn't carry the consequences of a low price or a decision that puts the crew in a difficult position; I do.
Start with work where the human thinking is already done. Give the tool organized records, ask it to translate and connect what you already know, and build in a real review before anything becomes a commitment. AI can sit beside you and help you see more of the business, but it shouldn’t sit in your chair.
