The old days
Two years ago, if a mid-sized business wanted to add AI capabilities - a customer support chatbot, document processing, or intelligent search - the conversation started at €10,000 and quickly climbed to six figures.
The reason was simple: the infrastructure didn't exist yet. You needed custom model training, dedicated ML engineers, expensive cloud GPU instances, and months of development time. The tools were powerful but raw. Building anything production-ready meant assembling a team of specialists and hoping they could deliver something that actually worked.
For most businesses, the math didn't add up. AI was exciting in theory and prohibitively expensive in practice.
The old days
€10K-€100K+
to add AI capabilities
- Custom model training
- Dedicated ML engineers
- GPU instances
- Months of development
Today
€3K-€15K
for the same outcomes
- Foundation models you build on
- RAG for your own knowledge
- Pay-per-token APIs, no GPUs
- Days to weeks
What changed
Three things happened almost simultaneously, and together they rewrote the economics of AI adoption.
1. Foundation models got good enough
GPT-4, Claude, and their successors crossed a critical threshold: they became reliably useful for real business tasks. Not perfect - but good enough that you could build products on top of them without training your own models. The shift from "build your own AI" to "use existing AI intelligently" was enormous.
2. RAG made custom knowledge practical
Retrieval-Augmented Generation (RAG) solved the biggest objection businesses had: "But the AI doesn't know anything about my company." With RAG, you take your existing documentation, product catalogs, FAQ pages, and internal knowledge, convert them into vector embeddings, and give the AI access to search through them at query time.
The result is an AI that answers questions using your specific information, not generic internet knowledge. And the cost of setting this up dropped from months of work to days.
3. APIs commoditized the compute layer
You no longer need GPU clusters. You make an API call, you get a response, you pay per token. OpenAI, Anthropic, and others handle the infrastructure. The marginal cost of an AI interaction dropped to fractions of a cent.
What this means for businesses
The practical upshot is that AI capabilities that used to require a €100K+ budget are now achievable for €3K-€15K depending on complexity. Here's what's realistically within reach:
Customer support agents - an AI chatbot trained on your knowledge base that handles routine inquiries 24/7. Not the frustrating rule-based chatbots of the past, but conversational AI that understands context, maintains conversation history, and knows when to escalate to a human.
Document analysis - feed the AI your contracts, proposals, or regulatory documents and ask questions in natural language. "What are our payment terms with supplier X?" "Which clauses mention liability?" This used to require custom NLP pipelines. Now it's a RAG system with a good chunking strategy.
Workflow automation - AI that processes incoming emails, categorizes support tickets, extracts data from invoices, or generates first drafts of reports. The tasks that eat hours of human time but follow patterns that AI handles well.
Intelligent search - replace the basic keyword search on your website or intranet with semantic search that understands what people mean, not just what they type.
The catch (there's always a catch)
Accessible doesn't mean trivial. The technology is ready, but implementing it well still requires judgment:
Data quality matters more than model quality. The best AI in the world can't give good answers from bad data. Before building any AI system, you need to audit and clean the knowledge it will draw from. Outdated FAQ pages, contradictory documentation, and missing information will all surface as AI hallucinations.
Integration is the real work. The AI itself is the easy part. Making it work within your existing systems - your CRM, your ticketing tool, your website, your internal processes - is where the engineering effort lives.
Monitoring isn't optional. AI systems need ongoing attention. Responses should be reviewed, knowledge bases updated, and edge cases addressed. "Set it and forget it" leads to an AI that confidently gives wrong answers, which is worse than no AI at all.
Not everything needs AI. This might be counterintuitive coming from someone who builds AI solutions, but it's true. If a simple form, a well-written FAQ page, or a basic automation script solves the problem, use that. AI adds value when the task requires understanding natural language, handling ambiguity, or processing unstructured data. If the task is structured and predictable, simpler tools are better.
How to think about AI adoption
The businesses that get the most value from AI share a pattern: they start with a specific problem, not a technology.
"Our support team spends 60% of their time answering the same 20 questions" - that's a great starting point for an AI support agent.
"We want to use AI" - that's a budget waiting to be wasted.
If you're considering AI for your business, start by identifying the repetitive, language-heavy tasks that consume disproportionate time. Those are your highest-ROI opportunities.
Where we come in
At Evorbi, we build AI solutions for businesses that have identified a real problem. We're a small studio that handles the technical implementation - RAG systems, chatbot development, workflow automation - so you don't need to hire an ML team or learn prompt engineering.
The typical engagement is scoped, time-bound, and focused on one clear outcome. No six-month research phases. No vague "AI transformation" roadmaps. Just a working system that solves the problem you described.
If that sounds like what you need, let's talk.
Frequently asked questions
How much does custom AI actually cost now?
Capabilities that used to require a €100K+ budget are now realistic for roughly €3K-€15K, depending on complexity. The drop comes from three shifts: foundation models you can build on instead of training your own, RAG for custom knowledge, and pay-per-token APIs that remove the need for GPU infrastructure.
Do I need to hire an ML team?
No. With foundation models and APIs handling the heavy lifting, most business AI - support agents, document analysis, workflow automation - can be built on top of existing models. You don't need to train models or run GPU clusters.
Will the AI actually know about my business?
Yes, through RAG (Retrieval-Augmented Generation). Your documentation, product catalogs, and internal knowledge become searchable embeddings the AI draws from at query time, so it answers using your specific information rather than generic internet knowledge.
What's the catch?
Accessible doesn't mean trivial. Data quality matters more than model quality, integrating with your existing systems is the real work, and AI needs ongoing monitoring. And not everything needs AI - if a simple form or script solves the problem, that's the better tool.
How do I know if AI is right for my business?
Start with a specific, repetitive, language-heavy task that eats disproportionate time, not a vague "we want to use AI." Those high-frequency, pattern-following problems are the highest-ROI starting points.
