Reliable AI agents: best practices
Guardrails, evaluation and human oversight for AI in your business processes.
An AI programmer is not a developer with a ChatGPT subscription — it's someone who knows when AI actually solves a problem and when it just adds complexity. The term gets used everywhere now, from job listings to LinkedIn profiles, without it being clear what the work actually involves. In this article I explain what I do, how to recognise a good AI programmer, and why I don't try to solve every project with AI — even though it's my field.
An AI programmer builds software where an AI model (like GPT-4, Claude or Gemini) performs a task that used to be done manually: understanding text, classifying, summarising, or holding a conversation. That's a specific layer on top of regular software development — you still need to know how to set up a database, build an API and secure a system. The difference from a "regular" developer is in the layer in between: prompt engineering, constraining AI to reliable answers, evaluating whether output is correct, and knowing when a task doesn't actually need AI at all.
What an AI programmer is not: someone who solves everything "with AI" just because they can. The best AI programmers I know (and what I try to be myself) say "you don't need this" at least as often as "I can build this for you".
These three roles overlap but aren't interchangeable. A traditional developer builds software without necessarily involving AI — fine for a webshop or booking system without any smart components. A no-code specialist (think Bubble, Webflow, or automation with Zapier/Make) builds quickly within a platform's limits, but gets stuck once the logic becomes AI-specific or business-specific. An AI programmer combines classic software knowledge with an understanding of how language models work, what they're good at, and where they become unreliable.
In practice you're often not looking for one label but for someone who can switch between these three — building a classic database, adding an AI step where it genuinely adds value, and using an automation tool where that's faster. See also my comparison of AI automation tools for exactly where that line sits.
If your answer to all four is "no", you probably just need a good developer or a no-code specialist — not someone specifically labelled "AI programmer". That's exactly the honest advice I give when a project like that lands on my desk.
Alongside websites and custom software, I work as an AI trainer at Outlier.ai, where I evaluate and improve the quality of ChatGPT, Gemini and Grok responses on a daily basis. That work gives me a constant, hands-on view of where these models are strong and where they fail — knowledge I carry directly into client projects. I know first-hand how an AI model reacts to an unclear instruction, why it sometimes invents facts, and how the right setup largely prevents that.
ultimAItech is registered with the Dutch Chamber of Commerce (KvK 98782800) and based in Wanneperveen. I work on every project myself — no account manager between you and the developer, no team of juniors executing while a senior sells it. At VYBR!S, I built a full AI chatbot platform serving multiple clients at once, and in order flow automation projects for various SMB clients, I combine AI classification with classic workflow logic.
A good project follows roughly these steps:
| Project type | Indicative price | Example |
|---|---|---|
| Add an AI step to an existing process | €500 – €2,000 | Automatically classify and summarise emails |
| Custom chatbot | €1,500 – €5,000 | Assistant trained on your own business data |
| Multi-step AI agent | €3,000 – €10,000+ | Agent that processes, checks and escalates requests |
| Custom platform with an AI core | On request | Like the multi-tenant platform built for VYBR!S |
Red flag when hiring: if someone immediately says "yes, that can be done with AI" without first looking at your data or process, be careful. A good AI programmer wants to know how clean your data is and how consistent your process is first — AI is never better than the input it gets.
Also see our broader custom software and AI services, or read how I keep AI agents specifically reliable in my article on AI agent best practices.
A regular developer builds software without necessarily involving AI. An AI programmer adds knowledge of how language models work, prompt engineering, evaluating AI output, and building guardrails to keep it reliable.
Only if you want to automate a task that needs to understand, generate or judge text — or an agent that carries out steps on its own. For simple workflow automation without an AI component, you often just need a regular developer or a no-code specialist.
That depends heavily on scope: adding a single AI step to an existing process starts around €500, a custom chatbot runs €1,500-€5,000, and a full AI agent or custom platform starts from €3,000. I work with a fixed price per phase, set after a feasibility check.
Often that's actually better — you work directly with the person building it, without an account manager in between. Look for demonstrable experience, references, and whether they're willing to say "no" when AI isn't the right solution.
A consultant often advises without building anything themselves. An AI programmer does both: advises on what's feasible, and actually implements, tests and hands it over.
A good AI programmer is neither a magician nor a hype salesperson — it's someone who understands your process, knows when AI is and isn't the answer, and then builds it solidly. If you're unsure whether your idea actually needs AI, that's exactly the conversation I'm happy to have before anything gets built.
Considering an AI project for your business? Book a free call or send a WhatsApp message. I'll give you an honest answer, even if that means you don't need an AI programmer (yet).
Guardrails, evaluation and human oversight for AI in your business processes.
n8n, Make, Zapier and the OpenAI/Claude API compared.
What we stand for and how a project with us runs.