Integrating AI in your business: use cases that actually work

Most AI projects that fail begin with "we need to do something with AI" instead of with a concrete problem. AI is not a goal, it is a tool that delivers enormous gains on specific tasks and burns money on others. After more than 30 products built, we know fairly precisely where it does and does not pay off.

Use cases that do work

AI delivers most on tasks that are repetitive, language-based and well-defined. These are the applications we build most often and that pay for themselves.

• Answering questions from your own documents. A chatbot that gives answers based on your manuals, contracts or knowledge base. This is called RAG and is the most valuable application for most companies.
• Classification and routing. Automatically labeling incoming emails or tickets and sending them to the right department.
• Summarizing and extracting. Summarizing long documents or pulling specific data from invoices and reports.
• Draft content. First versions of text, replies or reports that a human then checks.

Lexi AI, our collective-labor-agreement assistant, and ClaimHandler use AI exactly this way: well-defined, with human control where it counts.

What RAG is and why it is crucial

RAG stands for Retrieval-Augmented Generation. Instead of letting an AI model guess based on its general training, you first give it your own relevant documents and let it answer from those.

Without RAG an AI model invents plausible-sounding nonsense about your company. With RAG it answers based on your real documents, with source attribution.

The difference is enormous. A chatbot without RAG that answers questions about your product gives errors that damage your brand. A RAG system pulls the answer from your documentation and can even point to the source. That is the difference between a toy and a working business tool.

What it costs

An AI integration costs us between €8,000 and €35,000, depending on complexity. The build cost is one-time, but AI also has ongoing costs you need to understand up front.

• Model cost per use. Every question to a model like Claude or GPT costs a fraction of a cent to a few cents. At high traffic that adds up.
• Hosting and infrastructure. A RAG system needs storage for your documents and a search index.
• Maintenance. Models change, your data changes, and you want to keep measuring quality.

We work with Claude, GPT and Mistral, and pick the model with the best price-quality ratio per use case. Not every problem needs the most expensive model.

The pitfalls we keep seeing

Three mistakes cost companies the most money:

• AI without a problem. Starting with the technology instead of with a measurable task. First ask: which hours or which errors am I going to save with this?
• No own data. Relying on the bare model without RAG yields general, often wrong answers about your specific situation.
• No control layer. AI makes mistakes. Anything that matters needs a human or a control mechanism in the flow, plus logging so you can trace errors.

Start small and measurable

The best AI projects start with one use case where you can measure the result. How much time it saves, how many errors it prevents, how many customer questions it handles on its own. Only once that is proven do you expand. Building a broad "AI platform" before one application proves itself is the same mistake as fully building out a SaaS for the first customer.

Privacy and data

With AI on business data, privacy is no side issue. Where does your data sit, does it go to an external model, and what happens to it. We build integrations so that sensitive data stays protected and you comply with what your sector requires. For some clients that means local or European processing.

Want to know whether your process lends itself to AI and what it realistically returns? Look at our AI development approach. We start with the question of which problem you solve, not which model we use.

How to recognize the right use case

Not every process deserves AI. The best candidates share a few traits. Run your process through this checklist before you invest:

• It is repetitive. A task that recurs a hundred times a week delivers a hundredfold gain when automated.
• It is language-based or unstructured. AI excels at text, documents and conversations, exactly where classic software gets stuck.
• An error is recoverable or checkable. Tasks where one error has big consequences always need human control.
• You can measure the result. Hours saved, fewer errors or more questions handled. Without a metric you never know whether it pays off.

If a process scores on all four, it is a strong candidate. If it misses two or more, AI is probably the wrong solution.

Build or use an existing tool

Sometimes custom work is not needed. For general tasks there are ready-made AI tools that work right away. Custom work pays off when you need to connect AI to your own data and systems, when privacy must be strictly arranged, or when the AI must support its own workflow that no standard tool covers. The trade-off is the same as with any software: off-the-shelf for the generic, custom for what sets you apart. We regularly advise starting with an existing tool and only building custom once it hits its limits.