Models local Transcript not retained Session privileged
Everyone is still deciding what to do about AI.We spent two years shipping it.
A fuel distributor runs its pricing, its dispatch, its tank levels, its store and its back office on what we built, and the people who do that work use agents to do it every day. There are trucks on the road behind every part of that sentence. That is where we learned which work these models can be trusted with.
Forty hours a week of back-office work never became a job that company had to fill. Reading supplier price emails, chasing tank levels by phone, comparing terminals on a spreadsheet, planning routes, matching bills of lading. Nobody was replaced, because nobody was hired. The same three people cover the ground that would have taken four and a half.
The models run on hardware we own, so your data is not sent to a model vendor and none of it becomes training material for the next version of anything.
02 What we build Three examples, all running
No two of these started the same way.
Three jobs clients named when they arrived. Each one runs in a real company today. Treat them as the range rather than the menu, because the fourth one will be yours and we have not built it yet.
A supplier sends a price change as an attachment at four in the afternoon. The parser reads it and the prices are in the system before anyone opens the mail. It records why it picked each number, so a disputed line can be opened and checked.
Running in production
A staffing company's numbers sat in a system that would not report on them, so every branch answered the same question by hand each week. One page now holds the week, built against the data the existing system already had.
Running, used weekly
Books a year behind the bank, with three different businesses filed as one, so nobody could say which jobs made money. We dump everything, rank what is broken by what it costs, and work down that list. The decisions stay with the owner and the accountant.
Running
03 Two years in the trenches Status: running today
The hard part was never the model. It was knowing which job to point it at.
Most of the money going into this technology is being spent on deciding. Strategies, readiness scores, a pilot that proves the thing can be done and then stops. The spend is easy to approve and the return is hard to find, and a committee cannot tell you which of your jobs will pay for itself, because that answer lives in the work.
We went the other way. Two years inside a fuel distributor, where a wrong number is a wrong load and a wrong load is real money. Dispatch, pricing, tank levels, the store and the back office all run on software we built for it, and the people who send the trucks out and run the sales day use agents to do it. We found out what these models are bad at by putting them in front of our own operation first, and paying for it when they were.
The company is Better Day Energy. We own and operate it, which is why the numbers on this page are ours to publish and why we can show you the system that produced them. Different name on this door, same people behind it.
| What it is | A fuel distributor |
|---|---|
| Runs on | Software we built for it |
| Covers | Dispatch, pricing, tank levels, the store, the back office |
| Who uses it | Dispatchers, salespeople, the store floor |
| Prices read per day | 832, out of six suppliers' unstructured email |
| People it replaced | None. It is what they do instead of hiring |
| Status | Running today |
04 Containment Terms on request
A careful company stalls at the same place every time: where does the data go?
A finance team will not paste the general ledger into a chat window, and they are right not to.
So we do not ask them to. The models we use run on machines we own and administer. Your data goes into that room, the findings come out, and the data stays.
Every firm in this market says that, so ask us for it in writing. We sign it before the reading starts. For a buyer with a compliance function, that is the difference between a six-week legal review and having nothing to review.
The older sense of the word applies too. We do not publish findings. We do not put a client's numbers in the next pitch.
Models we run
Does not leave
- Your data
- Your ledger
- Your customer list
- Your payroll
Comes out
- The finding
- The proposal
- The fix
We will sign
- A data-processing agreement
- A named list of every component that touches your data
- No training on anything you give us
- Deletion on request, and on the day the work ends
- A log of every query we run against your data
05 The reading The second way in
When nobody can name the job yet, finding it is the job.
A company that already knows it wants a parser does not need this step. If you have been sitting through meeting after meeting about AI and what it could do for your company, let us go and find you the answer instead.
We start with the systems you already run: what they record, what people actually do in them, where the workarounds are. We sit with the people who do the work and we ask what they think is true. When the interviews and the data disagree, we follow the data.
It is free. Before it starts we write down how long it will take for your company, and that written time box is the control. What you get is an ordered list: a top item, the reason it is first, and a price against each line. In the working session we show the record behind every number rather than a summary of it.
It ends with the person who can decide, and a proposal that prices only what the reading found. On one reading the finding was that the client had just installed a system which already did the job, and the report said so. That cut our own scope, which is the test: a finding that takes work away from us is what makes the other findings worth reading.
The work after the reading is paid. A deeper look at logs, usage, the screens your people live in, and what is at risk is what justifies anything larger.
What we build for you is yours. The code, the credentials and the documentation transfer when the work ends, and nothing we write for you needs us to keep running it.
06 How a change gets weighed Order is fixed
Staying as you are gets weighed first.
A large replacement is expensive, it disrupts the company for a year or more, and it often ends with you on the same system as your competitors.
- 01Stay as you areWeighed first
- 02A thin layer on what you haveWeighed second
- 03Replace itHas to be justified
We recommend the third when the company needs to operate differently from its competitors, or when the standard package is the wrong purchase. We have made that call ourselves. The packaged system every other fuel distributor rents cost more than our business could afford and would have made us identical to the field, so we built instead. The people who dispatch the trucks use what we built every day.
When the answer is replace, the path is a bridge. A thin client, a skin, orchestration, so the old application stays usable while the new one takes the work.
07 Who we take Screened at intake
Shape Owner-operated, and one person can say yes Done this in Fuel distribution, staffing, petroleum services, convenience retail Also A large company in a crisis, when its leaders will listen and act
We take
- Owner-operators who will open the systems and the books
- Direct access to the person who decides
- A large company in a crisis, when its leaders will listen and act
We decline
- A firm that wants an AI strategy and a meeting cadence
- A leader who needs the finding softened for a board or an investor
- Anyone who will not act on what we find
The fear is the disqualifier.
08 The people No handoff
The person who sells you the work is the person who does it.
Matt Mizell
Sells, consults, delivers
Twelve years as chief information officer of Laird Technologies and Laird Performance Materials, through the sale to DuPont. Runs the fuel company on this page.
John Paukovits
Sells, consults, delivers
Ran applications and infrastructure at Laird Technologies and Laird Performance Materials before the DuPont sale, then technology and innovation at Supply Technologies, a ParkOhio company, until 2026.
There is no account team and no bench behind us. Whoever takes your call owns the finding.
09 Open a channel Privileged
Tell us who you are, the company, and the system you already run.
One of us reads it. Nothing you send here is published, quoted, or passed on.
- ReplyOne of us reads it and writes back.
- Time boxWe agree in writing how long the reading takes.
- The readingIt starts, and it ends in a working session and a priced proposal.