You've probably heard a hundred people throw around "AI" and "LLM" and "large language model" without ever explaining what any of it means. So let's fix that. No buzzwords. Just what the thing actually is and how you'd use it to write a better estimate this afternoon.

Strip the jargon first

An LLM is a next-word predictor. That's it.

It was trained on a huge pile of text from the internet, and from all that reading it learned one trick really well: given a chunk of words, guess what word probably comes next. Type "the customer called back and said the job" and it figures out that "wasn't" is a likely next word. Then it guesses the word after that, and the word after that, until you've got full sentences.

It's not thinking. It's not a brain. It doesn't know anything about your business. It's pattern-matching against everything it's ever read.

Here's why that matters: the quality of what comes out is tied directly to what you put in. Feed it a thin, vague request and it has nothing to pattern-match against except a thousand generic examples. Feed it real details and it has something to work with.

Why "just asking ChatGPT" usually produces garbage

Most people open ChatGPT, type "write me an estimate," and get back something useless. Then they decide AI is overhyped and close the tab.

They're half right. The output was useless. But the prompt was the problem.

Think about what the model has to work with when you type "write me an estimate." Nothing. No business name. No customer. No address. No idea what you're estimating, what you charge, or what the job involves. So it does the only thing it can: it invents all of it. You get a template with [Company Name] brackets and made-up line items at made-up prices.

That's not the model failing. That's the model guessing, because you gave it nothing to go on. A one-line question forces it to fill in every blank, and the blanks it fills are often flat wrong.

The LLM isn't a vending machine where you press a button and good work falls out. It's more like a sharp new hire who's never seen your business. Tell that hire nothing and they'll produce nonsense. Tell them exactly what you need and they're genuinely useful.

How to actually use it in a service business

Three things turn a bad prompt into a good one: a role, some context, and a constraint. Tell it who to be, give it the actual details, and tell it what the output should look like.

Say you wrap up a service call and jot down four lines of field notes:

  • main shutoff valve corroded, replaced with ball valve
  • water heater is 11 years old, showing early rust at the base
  • kitchen faucet dripping, replaced cartridge
  • customer asked about full replumb — house is 1970s galvanized

Drop that into an LLM with a real prompt: "You're a service writer for a plumbing company. Turn these field notes into a clear, friendly summary I can send to a homeowner named Marcus. Group it into work completed today and recommended future work. Keep the tone plain and skip any technical jargon he won't understand." Leave the prices for you to fill in.

What comes back is a customer-facing summary that reads like a person wrote it. "Today we replaced a corroded shutoff valve and fixed the dripping kitchen faucet." Then a separate section flagging the water heater and the galvanized pipe conversation. Five seconds of typing turned into something you'd actually put your name on.

Same trick works for the emails you hate writing. Take a stiff collections email:

"Per our records your invoice #4021 is past due. Remit payment immediately."

Ask the LLM to rewrite it as firm but human, from a small local company that wants to keep the relationship. You'll get something that still asks for the money but doesn't read like a parking ticket. In our experience, a warmer tone tends to get paid faster.

We'll go deeper on the exact framework in a follow-up post. For now, the takeaway is simple: role, context, constraint. Stop typing one-liners.

What it's good at, and where it'll burn you

LLMs are strong at language work. Rewriting something rough into something clean. Summarizing a long email thread. Drafting a first version of anything. Brainstorming names, subject lines, ways to phrase an awkward conversation. Reformatting messy notes into something tidy. If the job is "make these words better," it's good.

Where it falls apart is anything that needs to be factually correct.

Don't assume it got the math right. Ask it to total an estimate and it might add wrong, confidently. Newer tools often run a calculator behind the scenes and handle the arithmetic fine, but not every time, so check every total yourself. It also doesn't know current facts: today's permit fees, this week's material costs, whether a part is in stock. And it should never be your final word on anything legal or medical. An LLM will write you a confident, professional, completely incorrect paragraph about lien deadlines or a patient's symptoms without blinking. It doesn't know it's wrong. It can't.

The rule of thumb: trust the structure, verify the specifics. Let it build the estimate, the email, the summary. Then you check every number, every date, every claim before it reaches a customer. The LLM writes the draft. You're still the one who signs off.

That's the whole game. It's a fast, tireless writer that knows nothing about your business until you tell it, and gets details wrong when you don't check. Used that way, it'll save you real time every single day. Treated like a magic answer box, it'll embarrass you.

Want help finding where this fits?

We do a free AI workflow audit. We look at how your business actually runs and find one specific spot where AI saves you real time, whether that's estimates, follow-up emails, scheduling, or something else entirely. No pitch, no obligation, just a straight look at where it'd help. Find us at marshland.dev.