AI in business: 7 lessons from real-world projects

In today’s market, not using AI means falling behind. Many companies rush in, only to end up with costly projects that fail to deliver. 

We’ve been there too—hitting dead ends, learning what doesn’t work, and discovering what truly matters. 

The lesson? Success with AI isn’t just about picking the right model, but about mastering the whole process, from training to integration.

Let me explain with an example: imagine an AI tool as a new employee. 

• First, they must go through general training, where they acquire basic skills. This is essentially pre-training. 

• Then comes a specialized training for their role, during which the employee receives feedback on their performance. This is what we call reinforcement learning. 

• Finally, the employee is reviewed by senior colleagues, who help refine the employee’s performance. This is known as human evaluation.

Taking this path isn’t easy. You’ll face many obstacles that you’ll need to overcome. So, what should you watch out for during training?

Tip #1: Solve specific problems

With the growing popularity of artificial intelligence, many companies try to label their products as “AI-powered,” even when the technology doesn’t add real value. That’s why we set a simple rule in our company: the key is not to “have AI,” but to solve a specific problem with it.

Real value only comes when artificial intelligence delivers measurable results—saving time, producing more accurate outcomes, or improving the user experience.

“The key is not to have AI, but to solve a specific problem with it.”

Tip #2: Use relevant data

There’s a saying: “Garbage in, garbage out.” If you don’t have quality sources, you can’t expect quality results. With AI, this is doubly true. 

We encountered the problem of irrelevant data while developing a specialized assistant for legal analysis. The model had been trained on general texts and didn’t understand legal language at all.

We also faced a question: How to gather enough quality data while still protecting personal information? Our solution was to create synthetic data that mimicked real cases but contained no sensitive information. This also helped us train the model more precisely for the specific areas we needed.

Another challenge was cost. Training and running advanced AI models is far from cheap. Often, you need powerful GPU clusters, which can cost millions. That’s why we focused on two categories of models:

• The first are infrastructure models. These run in data centers, are the most powerful, and operate in clusters via the internet.

• The second are small optimized models that run locally, capable of simpler tasks without sending data online. Surprisingly, we found that local models aren’t always faster—specialized compute units (TPU/NPU) in clusters are often so powerful that they outperform local deployment.

“If you don’t have quality sources, you can’t expect quality results. With AI, this is doubly true.”

On top of that, AI models are often hard to interpret. This means that while a model may generate convincing outputs, it’s often unclear why it made a particular decision.

In some fields, where you need transparency about how a model reached its conclusion, this can be a serious problem. That’s why we started using methods known as explainable AI (XAI). These make systems more transparent and help us understand how the model makes decisions.

Tip #3: Don’t forget legislation and your technology

AI systems rarely function on their own. They usually need to communicate with other applications, databases, or APIs. 

When we were developing an AI tutor for internal purposes, we had to ensure it could easily connect with all our existing project management tools and our CRM system. At that point, we hit another challenge: integration was technically demanding because the systems used different standards and protocols. But without it, the product would not have been usable.

Expect other hurdles too—especially legal ones. With new regulations like the EU AI Act, it’s increasingly important to ensure that AI systems don’t violate users’ rights, protect personal data, and avoid discriminatory behavior. 

In practice, this means that already during the design phase, you need to ask questions like: “Does the model respect user privacy? Is it fair? Can it be audited?” If the model falls short in these aspects, you’ll have no choice but to rework it.

“Instead of launching an ambitious project with an uncertain outcome, we move forward step by step.”

Tip #4: Stay grounded

Time and again, we’ve benefited from following the rule “think big, start small.” Instead of launching an ambitious project with an uncertain outcome, we move forward gradually. 

• First, we research the market to see whether customers actually want the product. 

• Then comes a proof of concept, where we test whether our plan is technically feasible. 

• Finally, we create an MVP—a simple but functional version. 

This significantly reduces the risk of spending huge amounts of money on something doomed to fail, or on a product we wouldn’t be able to improve over time.

If you’re using large language models such as GPT, I’d recommend this approach: first, adapt the model on your own data. This ensures high accuracy for your specific purpose. Then, by crafting well-formulated prompts, you can shape the model into the form you’re satisfied with. 

Which strategy you choose depends on the task the model should solve, the availability of data, and your budget. For some applications, investing in quality prompts may be more effective, while in others it pays off to customize the model for a specific domain.

Tip #5: Don’t stop at implementation

Implementing a model is not the end of your work. AI systems must be continuously monitored, tested, and improved. Users change, their needs evolve, and your technology must adapt. New, more capable models are constantly emerging as well.

“AI systems must be continuously monitored, tested, and improved.”

Tip #6: Track your return on investment

Developing AI systems isn’t cheap. Costs for experts, infrastructure, and operations can quickly add up. That’s why you should set clear milestones at the start and track your progress toward them.

Decide what matters most to you—is it revenue growth, user satisfaction, or time savings? A clear plan from the start will guide future development.

Tip #7: Start now, but wisely

If AI hasn’t yet found its way into your company, now’s the time. Start with small, well-defined projects that deliver quick returns. You’ll gain experience you can later apply to more ambitious solutions. All you need at the beginning is a clear vision of what you want to achieve, realistic expectations, and a willingness to learn from mistakes. 

Remember: AI isn’t a silver bullet. But if you use it rationally, it can become a powerful ally for your business.