This is a practical session for the people who run fuel sites, haul and sell the product, service the equipment, and handle the paperwork behind all of it. Two goals.
Goal 01
Use it as a thought partner, not a glorified Google search.
Most people type a question, take the first answer, and decide the thing is mediocre. That is a search box with extra steps. You are going to watch what changes when you make it ask you questions before it answers.
Goal 02
Understand what tools are and what they actually do.
On its own it can only talk. Tools are what let it open your file, pull four sources into one pile, and draft into a system you actually use. That is where the hours come off your week, and it is also where you decide what it is allowed to touch.
Aunt Chilada's · 11:30 to 1:00. We are using fake sample data. Nothing real of yours hits a chatbot today.
I'm Brennan Gerle. I help organizations look past the hype of artificial intelligence and build practical systems that pay off in real work. I founded POLR AI in Chandler because businesses knew they needed AI and nobody was showing them how to put it in without losing the human touch they had spent years building.
My work is not really about automation. It is about how the business actually runs. I sit with leadership first to find where the operation is stalling and where a system would actually free capacity, then we build, then I do the training so the people doing the job feel equipped rather than replaced. Most of my clients are in construction, the trades, and manufacturing. Companies that run technicians, trucks, and a back office that never quite catches up. Change the equipment and that is a lot of this room.
Prioritize
Leadership sessions and workflow discovery come first. We rank the pain by dollars and hours, so the first thing we build is the one that pays for the rest.
Build
Working software inside the systems you already have. Internal tools, workflow automation, role-specific assistants, dashboards. Built with your people so they own it after I leave.
Ninety days
Month one we find it and rank it. Month two we build weekly with your team in the room. Month three we test, roll out, train, and measure against the baseline we took in month one.
Section 01 / 06 · Room scan
Where is the room?
I sent a short survey before today. Seven people filled it out. That is not this room, and it is not the industry. Raise a hand for the level you are at and we will count it live.
The room, as it stands
Total: 0 people
No counts yet. Click + on each level as the room shows hands.
Never usedYou have not tried a chatbot yet.
Some daysYou have asked it to draft an email or clean up a sheet.
Most daysIt is already part of the work.
Power userYou have a paid plan, saved workflows, and maybe already talk to other software.
From the form · where the seven are
0% said they had never touched it. 14% have tried it once or twice. 43% use it some days. 43% use it most days. Seven replies is not an industry, but nobody in that seven said never.
From the form · what is eating the week
86% named a specific repetitive task. Keying daily sales from remote stores. Reconciling delivery tickets to invoices. Price book updates. Driver and fleet scores. Watching the market. Different piles, same shape.
From the form · the hours
71% put a number on it, and those numbers add up to roughly seventy hours a week. The largest single answer was twenty-five, on delivery tickets and price book updates. The one that named QuickBooks by name was ten.
Why the form still sets the talk
One reply asked for peer stories rather than a product demo. One said AI is not really relevant to their week. Both are fair. So we run one loop end to end on a fake multi-store sales dump, then look at where else that same loop lands. Hands above are the count, not the form.
Seven replies. Those percentages are percentages of seven, not of this industry. Count hands.
Section 02 / 06 · What it is, what it isn't
It is not a search engine and it is not an expert. It predicts what comes next.
That is the whole mechanism. It read an enormous amount and got very good at guessing which word, number, or line should follow. None of what it read is about your operation. That is why it can sound completely certain and be completely wrong, and it is why a person stays at the end of every one of these.
Click any card to flip →
Thirty seconds on safe use
Do not paste store-level sales, cardholder data, or payroll into a free chatbot. Paid plans have better privacy defaults. Today we use a labeled sample, not your numbers.
Keep it secure
If you are going to put real numbers in, use the paid plan on the platform you already live in.
Twenty dollars a month gets you better privacy defaults and longer documents. The free version is fine to learn on. The paid version is what you should be on once store-level numbers are in the chat. Click a platform for the "turn off training" walkthrough.
How to turn off training, by platform · click to expand
Getting good at this has five layers. You are going to watch the first three.
Almost everyone stops at layer one, decides the tool is mediocre, and quits. That is the single most common reason somebody tells me AI did not work for them. The rest of this hour is layers two and three.
The three you will watch today
01
Prompt engineering
Writing the ask so one good answer comes back. Who you are, who it should be, what you want, and when to stop.
You will see this next, as CRIT
02
Context engineering
Deciding what it gets to look at before it answers. A perfect question against the wrong pile is still a wrong answer. Most bad results are this, not the wording.
The four sources, in the plug section
03
Harness engineering
The setup around it. What it may read, what it may change, and where it has to stop and wait for a person. This is the layer that keeps you out of trouble.
Connectors and the two switches
Past today, so you know the words exist
04
Loop engineering
One assistant running the same cycle on its own: do the thing, check the result, fix it, go again. The hard part is not the work. It is teaching it how to know when it is actually done, instead of stopping when it runs out of room.
Runs without you watching
05
Graph engineering
Several assistants with defined jobs handing work to each other, the way a crew does. Who owns what, who checks whom, and what gets called off when the facts change mid-job.
A org chart, not a chatbot
Section 03 / 06 · How you talk to it
Most people ask AI for an answer. CRIT asks it to think with you first.
CRIT is a framework from Geoff Woods, who wrote The AI-Driven Leader. Four parts, in order. It is the difference between a tool that hands you something generic and a thought partner that admits what it does not know before it opens its mouth.
C · Context
The background it does not have. What you run, what landed on your desk, what is at stake, what you already tried. Skip this and it answers a question nobody asked.
R · Role
Who you need it to be. A controller who has closed multi-site books. A service coordinator who knows where billing leaks. The role decides which questions it thinks to ask.
I · Interview
The part everyone skips, and the reason this works. You tell it to ask you questions before it answers. Same sentence every time: interview me one question at a time, up to three questions, to gain more context.
T · Task
What it produces once it has interviewed you. Format, deliverable, and the stop. Draft it, do not post it. The stop is part of the task, not an afterthought.
The part that stays yours.
A good prompt is not a way to think less. Every one of those four letters is a place where you put something in that nothing else has.
The context is yours
It has read almost everything ever written and none of it is about your operation. It does not know your stores, your accounts, or which manager rounds. Every answer worth having starts with something only you could have told it.
The judgment is yours
It drafts. You decide what posts, what gets sent, and what gets held until somebody counts the drawer. That is not a safety rule bolted on the end. It is the whole design.
The name on it is yours
When that note reaches a customer or that entry hits the books, nobody is going to ask what the model thought. Twenty years of somebody trusting you is not in the training data.
Which is the honest case for the Interview line. It is not a trick to get better writing out of the thing. It is the step that makes it come get what is in your head before it guesses at it.
What the Interview line actually changes.
It is not about getting a better answer. Both of these give you a good answer. The difference is how many decisions the thing made for you before you saw it.
A tool that answers
you ask→it guesses→you catch the misses
You end up the reviewer. The work is checking someone else's assumptions about your own operation, and you only find the bad one if you look hard.
A tool that asks first
you ask→it asks back→you decide
You stay the one steering. It cannot guess wrong about the thing it stopped to ask you, and nothing gets posted, sent, or filed without you.
Same messy pile. Three different asks.
Toggle Lazy, then Better, then CRIT on the sample below. This is four minutes. Then we get to goal two, the part a prompt cannot do by itself: getting the numbers out of the chat window and into something that counts.
Live buildSample data only, four fake Arizona C-stores. Every answer below is a real output, run ahead of today, not a mockup. Click through Lazy, Better, CRIT. What changes is not the quality. It is how many decisions the model makes for you without asking.
You
AI
What changed:
Now build one of your own.
Type into each box. The fields turn green as they get specific enough. Copy the prompt. Paste it into ChatGPT or Claude. This still does not post to QuickBooks. That's the next section.
CContextEmpty
Who you are, what landed, what's at stake. Store count, system, the mess.
Type your situation. Names of real stores stay out of the chat.
RRoleEmpty
Who you want the AI to be. Years, specialty, the lens to bring.
Tell the AI exactly who to be at the desk.
IInterview
Locked
This is the secret ingredient. Locked in so you don't forget.
This line tells the AI to ask up to three questions, one at a time, before answering. We've locked it so you don't skip it.
TTaskEmpty
What it should produce after the interview. Format. Deliverables. And the stop.
Tell it exactly what to make. Include the stop.
0% · Start with Context
Your CRIT prompt
Fill in the fields above. Your assembled CRIT prompt will appear here, ready to paste into ChatGPT, Claude, or your platform of choice.
Live response
Apply it · then we plug it in
CRIT gets you a draft. It does not put anything into your books.
You can paste the sample dump into chat and get a clean table. That is useful. The numbers are still only in the chat window. They are not in any system that counts. For that you need a connector.
Step 01
Paste the sample
Copy the fake dump from the black box in the next section. Do not use your POS or a card batch.
Step 02
Run CRIT
Context, Role, Interview, Task. Watch it ask what's missing before it drafts.
Step 03
Read the flags
Missing cash. Estimated gallons. Tenders that don't tie. That's the point of the sample.
Step 04
Then the plug
A draft in chat is not a posted entry. Connectors are how a draft becomes something your system of record can hold. Still with a human stop.
Section 04 / 06 · The plug
Chat can draft a file. It cannot put that file into the system you run on.
For some of you that is QuickBooks. For others it is Sage, Viewpoint, NetSuite, your dispatch software, or a state compliance portal. The rule does not change: moving a number into the record takes a connection you control, and you are the one who hits save. I am going to use QuickBooks for the next ten minutes because it is the one somebody named on the survey. Swap in yours as I go.
Read a file
Open the PDF, the spreadsheet, the scanned close report. It reads them directly instead of you retyping them into a chat box.
Go look something up
Search, fetch a page, check a source and tell you where it came from. This is the one people think is the whole product. It is not.
Actually do the math
Run the arithmetic instead of predicting what the total probably looks like. Different thing entirely, and it is why the sums tie.
Reach a real system
Pull a report out of your books, draft an entry back into them. This is the connector, and it is the only one on this row that you have to grant.
The first three come switched on in most paid plans and cost you nothing but attention. The fourth is the one with the two switches, and it is the rest of this section.
The chat
It can clean a dump, ask what is missing, and draft a sales receipt. It cannot see your books unless you let it.
the plug
Your system of record
QuickBooks here. Sage, Viewpoint, your dispatch board, or the state portal at your shop. Wherever the entry has to land. Letting it look costs you nothing. A draft is useful. Posting is a decision, and the assistant waits for you to make it.
A connector is how you let the model touch a real system: your spreadsheet, your email, your service software, or your accounting system. Without a connector, you copy and paste forever, which is that ten-hour week from the survey.
Three steps: get the data, move it into a draft, then decide whether it posts.
Click a step. We will stay on the remote-sales pile from the survey.
→
→
Level one · look
Letting it look is the safe half.
It can open yesterday's dump or pull a report out of QuickBooks and tell you what it sees. Looking changes nothing in your books. Turn this one on.
Level two · change
Letting it change things is the half that waits.
Same connection, second switch. This is the one that could write an entry or hit post. A draft is as far as it should go, and you hit save.
You hold both switches
These are two permissions, not one.
Granting the connection does not grant posting. Approve looking now. Leave changing off until you have watched it work for a few weeks.
Fake sample · four Arizona C-stores
Paste this live. Not a member's data.
SAMPLE ONLY. Invented for today. Not a member's data. No card numbers.
store_id date cash credit fuel_gal fuel_$ instore_$ notes
PHX-07 Camelback 8/19 1842.50 6210.15 4280 13263.72 4102.10 credit batch late?
PHX-12 DreamyDraw 8/19 2620.02 9856.28 ? 9588.30 2888.00 mgr text: gallons "around 3100"
TEMPE-03 Rural Aug 19 3,031.01 11,402.39 3650 gal 11420.85 3012.55 lottery not in this dump
SCT-01 Scottsdale 8/19/26 --- 12782.25 4102 12630.06 3550.00 cash drawer still open, will send AM
PHX-07 Camelback 8/18 3677.39 13834.00 4410 13622.49 3888.90
TEMPE-03 Rural 8/18 2480.15 10715.00 3400 10638.60 missing instore blank in email
No cardholder data. No real store IDs. If a tender doesn't tie to fuel plus in-store, that's on purpose. The flags are the lesson.
Section 05 / 06 · Where else this shows up
A few places operators already use this on admin work.
Same loop, different pile. Two of these came straight from the form and two are guesses. Tell me if yours is different.
Guess
Invoice coding
You get the PDF, then move it into a coded draft for fuel, c-store, lottery, or store supplies. You look at the GL, and you are the one who posts.
From the form
Vendor recs
You get the statement or the delivery tickets and put them next to what you think you owe. One survey reply spends twenty-five hours a week on this. You decide: is there a missing invoice, a duplicate, or a fuel surcharge you did not approve?
Guess
Field tickets
The tech writes one ticket on site. The office turns it into an invoice and a compliance record that have to agree. You get both as drafts, plus a list of what is missing, what is expiring, and what should not be billed yet. You still sign.
From the form
Market watch
One reply watches the market and customer needs, and said AI is not really relevant to that work. Coming from someone who does that for a living, that is closer to right than wrong. Nothing here sets a price or predicts where the market goes. What it does is read the pile that lands before you get to it, and tell you which of your customers a change actually affects. The call is still the job.
Your turn. Where does a messy pile land on your desk every week? Call it out. We'll map Get, Move, Decide on it if we have time.
None of this replaces your judgment. It gets you to a working draft fast. You still review, edit, and own the post.
Section 06 / 06 · Monday
Do not roll this out across every store on Monday.
Take tomorrow's remote-sales pile, strip names if you need to, run CRIT, and read the flags. Do not paste card data. Do not auto-post.
Tomorrow morning
Copy yesterday's remote-sales dump into a scratch file
Strip card numbers if any landed in the email. They shouldn't.
Run the daily sales CRIT prompt. Let it interview you.
What "done" looks like
A store-by-store table you could paste into a sheet
A QBO-ready draft, still unposted
A flag list you actually look at before anyone hits save
What you skip
No auto-post into QuickBooks
No store-level sales into a free chatbot on the company Wi-Fi "just to try"
No payroll, no cardholder data, no "the model can just log in"
If you want to keep going
Bring tomorrow's pile, whatever yours is
Strip the card data first
Scan the code at the end
A quick word on safe use.
AI sounds confident even when it's wrong. Don't paste store-level sales, cardholder data, or payroll into a free chatbot. Don't let a connector post. Human review stays your job.
Common questions
You're probably wondering one of these.
Click a question. These are the ones that come up every time.
Click a question to see the answer.
Open the floor
This is the part that matters. Your dump. Your books. Your "we already tried ChatGPT and it was meh." Let's go.
Next step
If this is useful and you want to keep going, let's talk.
Bring the actual pile, with the card data already stripped.
Scan this
Everything from today, plus the sample files and the prompts, on one page built for this group.
