DIY ChatGPT or a custom AI agent: what your company actually needs

The question always arrives in the same shape: "can't I already do all of this with ChatGPT?" Sometimes the honest answer is yes, and I say so even when it means not selling a project. Other times DIY costs far more than the 20 € a month it appears to cost, because the real cost is people's time and the errors nobody catches.

What DIY is genuinely good at

A general-purpose assistant subscription is perfect for work a person does once and reviews: writing and rewriting copy, summarising a document, drafting a difficult email, translating, brainstorming, making sense of a contract. Human supervision is natural, volume is low, and the person supplies the context each time. Building custom software for that is almost always a waste.

Where DIY stops working

The wall appears when you need one of these: access to real company data (current prices, stock, customer history), doing the same thing dozens of times a day without anyone copy-pasting, a consistent answer regardless of who writes the message, an audit trail, or work that continues when nobody is at a computer. A general assistant is not connected to your world: it starts from zero every time and quality depends on who is typing.

The numeric rule for deciding

Multiply how often the task happens per month by the minutes it costs each time. Under 5 hours a month, stay with DIY and train people to use it better — that is the economically correct choice. Between 5 and 20 hours a month, a single automation makes sense, typically from 1,200 €, and pays for itself in a few months. Above 20 hours a month, or when an error has a real commercial cost, you need an agent connected to your systems: the usual range is 2,500-8,000 €.

The middle path I often recommend

You do not have to choose: in most SMBs the best configuration is training the team to use the general tool well for creative work, and building custom software only for the high-volume process — usually customer requests, quotes or data entry. It costs little, produces a measurable result within weeks, and teaches the company to work with AI before spending larger amounts.

Watch the hidden cost of DIY

When ten people each use their own account, nobody knows what data gets pasted where, customer answers are inconsistent, and the work leaves no trace. This is not theoretical: it is why several companies call me after a year of DIY, and the first thing we do is create order, not add technology.

If you want to know which side your case falls on, bring me one process and its numbers in a free 15-minute call. If the answer is "leave it as it is", I will tell you.