The AI Decision Framework

A thinking tool for commercial contractors deciding where, and where not, to put AI to work.

Most AI advice tells you what to build. This framework tells you how to decide.

Every question below pulls an answer out of your own business. The solution that follows tells you what that answer means and what to do about it. The point is not to hand you a shopping list of tools. It is to give you a way of thinking that holds up on a live job, not just in a meeting.

The framework runs in three tracks, ten questions each, because the right question depends on where you sit. You are not meant to answer all three. Find your seat and work your ten.

• The Owner looks at other people's work: the management layer and the crews.
• The Operations Manager looks at their own desk, the hub everything runs through.
• The Individual Contributor looks at their own work as a craftsperson.

The same backbone runs through all three. Find the repeat work. Gate it on who owns the result and who can catch a mistake. Check that the data is clean and the freed time actually matters. Then make sure it can be maintained, and that it opens the door to the next thing.

If you take only one idea from this, take this one. The line between AI that works on a jobsite and AI that just sounds good in a meeting is always the same three questions: what does a mistake cost, whose name is on the output, and can that person spot a bad one. Those three questions show up in every track below. Watch for them.

---

Track 1: The Owner

The owner has two lenses. One on the management layer, meaning estimators, PMs, controllers, and office managers. One on the boots-on-the-ground crews. The questions split that way.

Looking at the management layer

1. Which of my managers spends real hours every week rebuilding the same report, schedule, or spreadsheet?

If you can name the person and the cadence, you have found a build candidate. High-frequency repeat work pays back fastest. If you cannot name both, it is not ready yet. Target the recurring weekly task first, and aim the recovered hours back at bidding and running jobs.

2. What is my office consistently doing late or doing badly because nobody has the hours?

These are margin leaks, not time problems. Quote follow-ups that slip, billing that lags, job costing that is always a week behind. AI wins here by holding a floor nobody currently has time to hold. Prioritize the leak costing the most money, not the task that annoys you most.

3. When one of these office tasks goes wrong, what does it cost me, and who catches it today?

This sorts your candidates by risk. Cheap, self-correcting errors can run loose. Expensive errors, a wrong bid number or a missed submittal, need a named human checkpoint before output leaves the building. If no one catches it today, that is the first thing to fix, with or without AI.

4. For each manager task, is this a judgment call or the same steps every time?

Same-steps-every-time work is what you automate: pulling numbers together, formatting, flagging jobs that are behind. Judgment work stays with the person: pricing a tough bid, dropping a vendor, reading a client. Sort every candidate into one bucket. Build the consistency tasks. Leave the judgment tasks for AI to assist, not decide.

5. If AI does a manager's work, whose name is still on the result?

Every workflow needs one owner who can look at the output and know whether it is right. If the answer to a bad result is "the computer did it," you have an ownership gap, and that gap has to close before you automate. No owner, no go.

6. Can that manager actually tell when the AI got it wrong?

This is the discernment test. If the person cannot spot a bad number, automating that task is risky, not helpful. When the answer is no, the move is to sharpen the person on the task first, then automate. The human has to stay smarter than the tool.

Looking at the boots-on-the-ground crews

7. Where do my field crews lose time to paperwork or chasing information that should reach them automatically?

Daily logs, photo documentation, material requests, RFIs. Look for information that should flow to the field without someone keying it twice. Target the highest-frequency field task that currently pulls a person off productive work.

8. What do my crews currently have to call the office to find out?

Each of those calls is a delay and a person interrupted. If the answer could be served up without pulling someone off another job, that is a candidate. Build the information path before you build anything fancy. Faster answers usually beat smarter ones in the field.

9. When the field and the office are not talking, what does it cost me in rework, delays, or wrong material on site?

This tells you where better information actually changes the job outcome versus where it is just nice to have. Put effort where a faster, cleaner handoff prevents a real cost. Skip the places where the disconnect is annoying but harmless.

10. If I free up office time and field time, what do I want that capacity going toward?

Recovered hours with no destination are slack, not savings. Name the target before you start: more jobs, faster closeouts, fewer weekend hours. If you cannot name where the time goes, the savings will quietly evaporate and you will wonder what you paid for.

---

Track 2: The Operations Manager

This person runs scheduling, vendors, job flow, and the daily fires. These questions are about their own desk, not their team's.

1. What do I rebuild from scratch every week that I dread?

The schedule, the status report, the same email to the same vendors. The task you dread is usually the repeat task, and the repeat task is your first build candidate. Start there, not with the flashiest idea.

2. What am I always behind on, not because it is hard but because I run out of day?

That backlog is where AI buys you the most, because the work is not difficult, there is just never enough time. Point AI at the time problem, not the skill problem. The skill problems are where you still earn your keep.

3. When I rush these tasks, what slips?

A wrong date, a missed order, a vendor left out of the loop. Identify which of those actually costs the company money, and put a checkpoint there. The cheap slips can run on their own. The expensive ones get a human read before they go out.

4. Which parts of my job are pulling information together versus making the call?

AI does the gathering. You make the call. Map each task into one or the other, then offload the gathering so your hours go to the deciding. If a task is mostly judgment, leave it alone. If it is mostly assembly, it is a strong candidate.

5. If I hand a task to AI, can I still look at what it produced and know in five seconds whether it is right?

If yes, you are ready to hand it over and stay in the loop as the reviewer. If no, you are not ready, because you cannot catch a bad output you cannot recognize. Close that gap first by getting clear on what "right" looks like.

6. Is the data I would be running this on actually clean?

Your job list, your QuickBooks, your vendor pricing. Messy inputs produce confident, wrong answers, and you are the one who looks bad. Before anything goes live, confirm the source data is right and know who keeps it right. If it is messy, fixing the data is step one, not an afterthought.

7. When I free up a few hours a week, where do I want them going?

Getting ahead instead of staying even. Catching problems while they are small instead of fighting them as fires. Decide the destination up front so the time does not just refill with more of the same.

8. When our pricing changes or we switch a tool, can I adjust this workflow myself?

If you can adjust it, you own it. If you are stuck waiting on someone else every time the business shifts, you have bought a crutch, not a tool. Build or accept only workflows you can maintain, or get trained to maintain them as part of the deal.

9. What is the one task that, if I automated it, would make three other things on my plate easier?

That is where you start, because it compounds. The flashy standalone win helps once. The upstream task that unclogs everything downstream pays you every week. Rank candidates by what they unlock, not just what they save.

10. Once I have one working, what does it teach me to spot next?

The goal is not one shortcut, it is getting good at recognizing these opportunities yourself. Treat the first workflow as training for your own eye. When you can spot the next one without asking permission, the framework is working.

---

Track 3: The Individual Contributor

The estimator, the PM, the admin, the field lead. One person thinking about how they get their own work done well.

1. What part of my day is the same thing over and over?

The repetitive piece is usually where you can get help without losing control of the work. Point AI at the repetition first. The varied, thinking parts of your job are not the target.

2. What takes me longer than it should, and what makes it slow?

Hunting for information, retyping the same things, formatting. Those are the parts a tool can take off your plate. Identify the specific slow step, then offload that step rather than trying to hand over the whole task at once.

3. Where do mistakes sneak into my work when I am moving fast?

A tool that catches those before they go out makes you look sharper, not lazier. Use AI as a second set of eyes on the error-prone steps. Keep yourself as the final read, since the mistake is still yours if it slips through.

4. On any task, which part is real skill and which part is busywork?

Keep the skill, offload the busywork. The judgment is why they hired you. The retyping is not. Sort each task into the two, and only hand over the busywork half.

5. If I use AI to draft or build something, can I look at it and know whether it is good?

If you can judge it, the tool helps you. If you cannot, you are not ready to lean on it yet, because your name is still on the result. Build your own judgment on the task before you let AI carry it.

6. Am I actually checking the output, or starting to trust it blindly?

The day you stop reading what it gives you is the day it bites you. Keep a habit of reviewing every output, even when it has been right a hundred times. You stay responsible for what leaves your hands, always.

7. What do I want to do with the time I get back?

Better work on the parts that matter, or just leave earlier. Both are fine, but decide which, so the time does not vanish into busywork you did not mean to keep. Name the destination before you start.

8. Is what I am feeding it accurate?

Garbage in, garbage out. If your inputs are sloppy, the output will be too, and that is on you. Check your own inputs first. Clean inputs are the cheapest quality improvement you can make.

9. When my work changes, can I adjust how I am using the tool, or did I just memorize one trick?

Understanding beats memorizing. If you only know one trick, you are stuck the moment the work shifts. Learn the tool well enough to bend it to a new task, not just repeat the one thing you were shown.

10. What did this teach me that I can use on the next task?

The goal is not one shortcut, it is getting better at spotting where this helps across everything you do. After each win, ask what it taught you. That habit is what turns a single tool into a real edge.

---

The Through-Line

Across all three tracks, the questions do the same four jobs in the same order.

Diagnosis, what should we even target? Questions 1 through 4 in each track.

The human-in-the-loop gate, who owns it and can they catch a mistake? The ownership and discernment questions.

Reality check, is the data clean and does the saved time matter? The data and recovered-time questions.

Durability and compounding, who maintains it and what comes next? The maintenance and next-step questions.

If you only remember three questions, remember these. What does a mistake cost. Whose name is on the output. Can that person spot a bad one. Answer those honestly and the rest of the framework falls into place. Skip them, and you end up with AI that looks good in a demo and fails on a live job.

---

Want to put this to work?

If a few of these questions landed on a task you already know is costing you, that is the place to start. The right next step is a short conversation about that one workflow. No pitch. We will tell you honestly whether AI fits it, and whether we are the right people to build it.

Schedule a call

Prefer email? Reach John directly at john@clearoakconsulting.com and tell us the one task you would most like off your plate.