A lot of what gets sold as AI never pays for itself, so I start from the work: one job that gets done over and over, looked at properly, and a written answer on whether software should take it over. You do not need any technical knowledge for that conversation.
One process, one conversation, and my view of whether AI would take that work off someone. It is the smallest AI job you can order here.
How to start
You do not have to prepare anything. Two sentences about the work that takes the most time in your week are enough. If it turns out there is nothing here worth building, you will have lost an afternoon and learned where not to spend money.
Worth reading first
How the audit works, and when a chatbot on a company website pays off, are two separate articles on the blog.
The same four steps every time. Choosing a tool comes after the process is on paper.
I pick a single job where the stake is clear: time, how often it repeats, what an error costs, or what a customer feels. Only one, because a pilot that touches everything is hard to judge afterwards.
I write down the data it uses, the tools it passes through, the decisions inside it and the exceptions that break it. Most AI pilots that fizzle out skipped this step.
Input, context, the action it takes, the point where a person steps in, and a way to measure whether it did the job. It is small, and it handles real cases from day one.
I run it on cases that have already been handled and compare what it did against what a person did. It gets a second job only once the first one holds.
Example flow
A supplier invoice reaches the inbox. The system reads the fields, checks them against the order, updates the CRM and routes mismatches to a person for approval.
Somebody opens an email, reads four numbers off a PDF, compares them with what was ordered, types them into another system and files the PDF. A few minutes each time, dozens of times a month, and nobody notices the work until something goes wrong.
Each part of that flow does a different job. A model reads the document, because it copes with a badly scanned PDF better than ordinary code does. Ordinary code does the comparing and the writing, because a comparison has to give the same answer every time. The last part matters most: when the invoice and the order disagree, the flow stops and shows the case to a person, and those are the only cases that need anyone's attention.
First I run the flow over invoices that have already been processed and compare its results with what was done by hand. It handles live documents only once the two match.
Six things. Each one is a job somebody is doing by hand today.
Where this already runs
Asystent PV is my own chat assistant for solar installers, live at asystentpv.pl. It answers customers under the company's name, says when it does not know and passes the question to the office. You can ask it something right now. The same build goes onto another company's offer, prices and questions.
View the Asystent PV projectFive commitments you can hold me to.
Money leaving, a contract, an answer to a customer that cannot be taken back: the flow prepares it and a person releases it. That step is part of the flow's code, so the model cannot talk its way past it.
Before a flow gets more work, I measure on your own examples how many cases it handled without a person and how many it got wrong.
I build inside the systems you already have: your CRM, your mailbox, your server.
Every automated action is written to a log a person can read afterwards. When something looks off, you and I start from the same record.
If the audit finds AI would not take real work off anyone, the note says so and suggests what would help more. You read it before any build starts.
The tools I reach for most often.
Where AI is not worth it
Plenty of steps do not need a model at all. If a rule with three conditions gives the same answer every time, it beats a model that is right most of the time, and it never surprises anyone on a Monday morning. Where the job is small enough that custom code would be the wrong answer, I wire it up in Zapier or Make and write nothing. And where the honest fix is a better form, a tidier spreadsheet or one fewer step, I will tell you that.
Four projects that run every day. Each write-up shows what was built and how it works.
AI implementation quote
Say what AI should do at your company. Within one working day I reply with a price and a date, or with the questions that have to be answered first.
