Free guide · Construction edition

How AI works, in plain English.

A short, visual guide for business owners and leadership teams. Part 1 covers what AI is and how businesses use it. Part 2 explains how it actually works, so you can use it well and know when not to trust it.

Start the guide Download the PDF
Examples from construction or real estate
Part 1

The big picture

What AI is, what it's good at, and how a business actually puts it to work.

  1. 01What AI is
  2. 02What it's good and bad at
  3. 03Three ways businesses use it
  4. 04Start with a workflow
  5. 05What it costs
  6. 06A 30-day pilot
  7. 07Protecting your data
Part 1 · The big picture
01

"AI" is a family of tools, not one thing.

Most of what people mean by AI today is a language model: software that has learned the patterns of written language well enough to read, write and reason in it. It sits inside a much older field.

Artificial intelligence Machine learning Generative AI Languagemodels
Part 1 · The big picture
02

It's very good at some work and unreliable at other work.

The pattern is simple. Language models are strong with words and patterns, and weak wherever exactness or facts they weren't given matter.

Strong
Drafting and rewriting
Summarizing long documents
Pulling details out of messy files
Comparing documents side by side
Answering from sources you provide
Check carefully
Exact math across large data
Facts it wasn't given
Recent events
Judgment calls about people
Anything you can't verify
Part 1 · The big picture
03

Businesses use it in three ways, each with more independence.

An assistant answers when you ask. An automation runs the same step every time. An agent works toward a goal and decides its own next steps. More independence means more value and more need for a person to sign off.

Assistant You askAI answersYou decide Automation Something arrivesAI does one stepResult filed Agent A goalPlans, uses tools, checksYou approve
Part 1 · The big picture
04

Start with a workflow, not a tool.

Map one job your team does often, step by step, and mark the steps where AI could take on the reading, sorting or first draft. People keep the decisions. Here's bid leveling on a construction job.

  1. SubcontractorsA dozen proposals arrive, each in its own format
  2. AIChecks each one against your scope sheet
  3. AIFlags the gaps and drafts the call list
  4. Your estimatorMakes the calls and decides what to plug
  5. OutcomeA leveled bid, ready for the estimate
A person decidesAI handles the stepThe outcome
Part 1 · The big picture
05

The software is only part of the cost.

There are three kinds of cost to plan for. The third is the one most budgets leave out.

Setup and your team's timeMapping the work, connecting your systems, training people and checking output
UsageCharged per token when you build with AI directly, so the bill grows with how much you run
SeatsA monthly subscription per person for tools like ChatGPT, Claude or Copilot
Part 1 · The big picture
06

Prove it with a 30-day pilot.

Pick one workflow, measure how it runs today, and give a small group a month with the new version. Training sticks when it's built around the work people already do.

1Week 1Pick the workflow and measure how it runs today
2Week 2Set up the tool and train the group on their own work
3Week 3Run it for real, with a person checking every output
4Week 4Compare against week 1 and decide whether to expand, adjust or stop
Part 1 · The big picture
07

Know where your data goes before you paste it in.

A few habits cover most of the risk, and none of them require a technical team.

Use business plansBusiness and enterprise plans usually keep your data out of model training. Free and personal plans may not.
Decide what never goes inWrite down the data your team shouldn't paste into any AI tool, such as client financials or personal records.
Keep a list of your toolsKnow which AI tools your team uses and what each one can see. A one-page policy is enough to start.
Part 2

How AI actually works

You don't need to be technical to use AI well. You do need a working picture of what's happening inside, because it explains both what AI is good at and where it goes wrong.

  1. 08Tokens
  2. 09Predicting the next word
  3. 10How a model is trained
  4. 11The context window
  5. 12Why it makes things up
  6. 13Answering from your documents
  7. 14Tools and agents
  8. 15Giving good instructions
  9. 16Checking the work
Part 2 · How AI actually works
08

It reads and writes in small pieces called tokens.

A model doesn't see words the way you do. It breaks text into tokens, which can be a whole short word, part of a longer word, or a punctuation mark. Tokens are also how usage is measured and billed.

An illustration. Every model splits text its own way, so real token counts differ.

Part 2 · How AI actually works
09

It writes by predicting the next token, over and over.

At each step the model weighs every possible next token, picks a likely one, adds it, and repeats. Fluent writing comes out of that loop. So does a confident sentence that isn't true.

The estimator checked the

An illustration. The bars show relative likelihood, not real figures.

Part 2 · How AI actually works
10

It learned from a vast amount of text, then was taught to be helpful.

First the model reads an enormous amount of writing and learns how language fits together. Then people show it good answers and rate its attempts, which shapes it into an assistant that follows instructions.

PretrainingThe model reads a huge range of text and learns the patterns of language.
Fine-tuningIt learns from examples of good questions paired with good answers.
Human feedbackPeople rate its answers, and it learns which kinds of answer they prefer.
The assistant you useIts knowledge stops at a cutoff date, unless a tool such as web search gives it newer information.
Part 2 · How AI actually works
11

It only knows what's in front of it right now.

Everything a model can use in a conversation sits in its context window: your messages, any files you add, and its own replies. Today's windows are very large. Anthropic's current models hold up to a million tokens, which it puts at roughly 555,000 words, enough for a whole stack of contracts, specs and subcontractor proposals at once.

Anything outside the window, it doesn't see: last week's chat, unless the product saves it, or the oldest part of a long one once it's dropped or summarized.

Even inside, research has found models use the start and end more reliably than the middle. Position matters, not age: it rereads the whole window each time.

Out of the window
Last week's chat, or the oldest part of a long one. Not seen.
Context window
Your messages, your files and its replies, all at once
Start and end: used most reliablyMiddle: easier to miss
Part 2 · How AI actually works
12

When it doesn't know, it can still sound sure.

The model is built to produce likely text, so a gap in what it knows gets filled with something that sounds right. This is often called hallucination. It's why you give it sources and check anything that matters.

What was the change-order total on the Lone Pine job?
The Lone Pine job had four approved change orders. They totaled $48,200, mostly for added site lighting.Sounds right. Isn't in any document it was given. The largest was approved in March.

An invented example, to show the pattern.

Part 2 · How AI actually works
13

Give it your documents, and it answers from them.

Many business tools first search your files for the passages that matter, then hand those to the model along with your question. This is called retrieval, or RAG. Answers get more accurate, and they can point to where each fact came from.

You ask a question
The tool searches your filesContracts, specs and past bids
It pulls out the passages that answer the question
The model writes an answer using only those passages
You get the answer with links to its sourcesSo you can check each fact yourself
Part 2 · How AI actually works
14

With tools, an agent can do the work itself and report back.

On its own, a model can only tell you what to do. Connect it to tools such as search, a spreadsheet or your email, and it can take a step, look at the result and decide what to do next. That loop is what people mean by an agent. The more it can do on its own, the more a person should approve before anything goes out.

Plan Use a tool Check the result Next step You approve
Part 2 · How AI actually works
15

Better instructions get better work.

Treat it like a capable new hire on day one. Tell it the situation, the job, what good looks like and the shape you want back.

ContextI run a 40-person general contractor. This is a subcontractor's electrical proposal.
TaskList everything it excludes or doesn't mention compared with our scope sheet.
FormatA table: item, included or excluded, page number.
GuardrailIf you aren't sure, say so. Don't guess.
Part 2 · How AI actually works
16

Build the check into the workflow.

AI output is a first draft. Decide in advance who reviews it, what they compare it against, and who signs off. Having a second AI review the first helps, but it doesn't replace a person for anything that matters.

  1. AIWrites a first draft and cites its sources
  2. ReviewerChecks the draft against those sources
  3. ReviewerCorrects it, or sends it back with notes
  4. Sign-offA named person approves it before it goes out
AI fluency on one page

What to remember.

It predicts likely text, so fluent isn't the same as true.
It only knows what's in its context window, so give it the sources.
Start from one workflow, and keep people on the decisions.
Budget for setup and your team's time, not just seats.
Use business plans and decide what data never goes in.
Build the review step in before anything goes out.

Want help applying this to your business?

We'll set up 30 minutes to talk through where AI could help and roughly what it would cost. If it isn't the right time yet, we'll say so.

Start a conversation Download the PDF