AI in Your Company: Five Mistakes Most SMBs Make

Zsigmond Máriás, founder of Loginet, joined Robesz Pintér on the Simple Trends Podcast to talk about what goes wrong when businesses try to adopt AI. Máriás founded Loginet in 2008, straight out of university. The company has since become a major player in Hungarian web and mobile development: enterprise custom builds, in-house products, outsourced product development for startups. In recent years, more clients come to them with AI implementation questions. These are SMBs who know they should be doing something with AI but have no idea what or how. The conversation drew on Loginet's 25-question AI implementation guide. Five points stood out.

Start With a Problem That Hurts

Most AI adoption attempts begin with a grand vision. Strategy decks, six-month projects, rounds of workshops. Zsigmond thinks that order is backwards.

"Pick a problem that drives you crazy, and solve that one."

Say you run B2B e-commerce and orders arrive through Viber, email, and photographed faxes.

Staff enter them by hand. You don't build full automation on day one. You build the piece that turns ten typical orders into a spreadsheet. If that works and your team accepts it, you move to the next step: plugging into the CRM or the webshop.

One small win inside the organization makes the second and third project far easier to greenlight than a six-month abstract strategy program ever could.

Messy Processes Plus AI Equals Faster Mess

If a process has missing steps and unreliable data, AI will reproduce those errors at higher speed.

"Feed it a bad prompt, ask for the wrong thing, and it'll give you the wrong answer fast. Scaling that up helps nobody."

Zsigmond recommends fixing one process at a time. Trace where data originates, find the gaps, close the broken links, then layer AI on top. Your input quality caps your output quality, and that is an organizational problem before it becomes a technical one.

The other common mistake: inventing a new problem for AI to solve instead of tackling the long-standing ones. Zsigmond sees AI's biggest payoff in tasks employees dislike, do poorly, and still spend hours on.

People Stay. Busywork Goes.

Companies that lay off developers and cite AI are, in most cases, doing what Zsigmond calls "AI washing." The numbers needed to improve, and AI gave them a convenient excuse.

"Companies that run their existing teams well can redirect people from pointless tasks to meaningful ones."

Two examples from Loginet. First: landing pages used to require a developer, a copywriter, and a coordinator, so they rarely got built. Now one person ships one end-to-end in a morning. Second: post-workshop memos that nobody used to write because the effort was enormous. Now AI processes the recording, notes, and a good template into a solid summary.

Both cases follow the same pattern: a task that used to get skipped now gets done.

Fix Your Knowledge Base Before You Deploy AI

If your internal data is disorganized, AI will return the internet's lowest common denominator: generic, useless answers. You need clean data, current documentation, and digitally accessible company knowledge for anything better.

"You need your own data, your own knowledge base, and most companies don't have one."

Zsigmond says every client claims on first ask that their documentation is excellent and their specs are thorough. In practice, that is never the case. Documentation is stale, knowledge lives in people's heads, and the spec has ballooned to three thousand pages riddled with inconsistencies.
AI can help here too. It makes economically viable what used to be too expensive and slow: documenting a large software system, consolidating knowledge scattered across team members. But you also need workflows that keep that knowledge base current, not a one-time cleanup project.

Keep the AI Engine Swappable

The AI toolkit shifts month to month. The best model today didn't exist six months ago. As Zsigmond and his team joke: "The new industry standard has been the standard since yesterday, or for the last twenty minutes."

That makes vendor independence critical.

"If you build something on an OpenAI solution, build it so you can swap it out."

This is the model-agnostic approach: make the building blocks replaceable. When a better model ships from any provider, you shouldn't have to rebuild the whole system. Vendors want deep integration because that creates lock-in. The more complex a bundled solution you buy, the harder individual parts become to replace, and the more dependent you get.

Loginet follows the same philosophy in its own products: they build tools where swapping components stays simple.

Your Next Step

Zsigmond advice: find one area that hurts. Bring in expertise, internal or external, to decide whether AI can address it and at what cost. Build it small. If it works, show it off inside the company and start the next one.

"Once you can point to one success story inside the company, selling the second and third project gets much easier."

Find more articles on enterprise AI applications on the Loginet blog, including the 25-question AI implementation guide that sparked this conversation.