GenAI Integration for Business: LLM in Your CRM and ERP

What GenAI Integration Is and Who It's For

GenAI integration means connecting a language model to a specific company system — CRM, ERP, an internal knowledge base, or documentation — so it answers questions and performs tasks against real, current data instead of guessing from what it learned during training. The practical difference: the model knows the current order status, the company's actual procedures, or the wording of an internal policy, instead of producing a plausible-sounding generic answer.

This kind of integration makes sense for companies where a large share of daily work is searching for information and answering repetitive questions: HR fielding leave and payroll questions, customer service checking order status in a CRM, sales teams digging through contact history. I build this for small and mid-sized companies in Poland and EU clients that already run a CRM or ERP and want the language model to use it, rather than operate as a separate, disconnected tool.

How It Works Under the Hood

At the core is routing across more than 35 language models. A simple question goes to a cheaper, faster model; analyzing a long legal document goes to a stronger one. Routing cuts cost and latency, because not every request needs the most expensive model on the market. On top of that sits tool use and the MCP protocol, which let the model do more than generate text — it calls functions in the client's CRM or ERP: creating a record, updating a status, sending a document.

When a company's knowledge base runs into hundreds or thousands of documents, simply stuffing them into the model's context window stops working. That's where RAG comes in — retrieving the document fragments relevant to a question and feeding them to the model as context, often combining semantic search with classic keyword-based BM25. The model answers from a specific retrieved passage rather than from memory, which cuts down on fabricated answers.

Clients can bring their own API keys (BYOK), which keeps full control over billing and lets them use terms already negotiated with a model provider. I also set cost limits per model and per user so the integration never produces a runaway bill.

Real Scenarios: From CRM Records to HR Documents

A concrete example of GenAI tied to a knowledge base is Janina, an HR assistant that has been running in production since 2026. It answers HR questions using hybrid RAG — semantic search combined with BM25 — over 40 pieces of legislation and more than 900 documents, including Polish Supreme Court rulings. Every answer cites the specific article of the relevant law, so HR staff can verify it instead of just trusting it.

Janina tracks social security and tax filing deadlines, generates HR documents, integrates with Poland's KSeF e-invoicing system, and pseudonymizes employee data. It runs on WhatsApp, where employees already message anyway. It illustrates a pattern I reuse elsewhere: a model connected to a large, specific document base, citing its source, integrated with a system the company already uses — not necessarily HR, just as often a sales CRM or a service ticketing system.

Integrations and Technical Stack

On the stack side, I work mainly with Anthropic and OpenAI models, the MCP protocol for communication between the model and tools, tool use for calling functions in client systems, and vector databases for semantic search over documents. The specific model and vector database depend on what systems the client already runs and how many documents need to be searchable.

Hosting for agents and models can run on a dedicated server per agent, in an EU cloud or on-premise, with automatic failover across 35-plus LLMs — the same setup I run at ClawLabs, where I host production AI agents across 13 communication channels with our own apps published on the Microsoft Store and Google Play. That infrastructure carries over directly into a GenAI integration built for a specific company.

Cost and Timeline

The process starts with a one-hour call: the client explains what eats their team's time, I explain what can actually be automated and which systems need connecting. Then I build a prototype, usually within two weeks, running on the client's real data rather than sample data. Only after the prototype do we agree on production rollout — client cloud, EU cloud, or on-premise — with logging, cost controls and GDPR compliance built in from day one.

Across the Polish and EU market, ballpark pricing for a GenAI integration project in 2026 typically runs from the low thousands to tens of thousands of euros for the first phase, depending on how many systems get connected and how large the knowledge base is — this is a general market indication, not a fixed price list for this specific service. The actual cost and timeline get set after the initial call, once I've seen the systems and data involved.

When GenAI in Production Makes Sense — and When It Doesn't

A GenAI integration makes sense when a company has a repeatable process built around documents or system data, and the volume of questions is high enough that automating it genuinely frees up the team. It also makes sense when the company's data is sensitive and needs EU-based or on-premise processing — something I settle on day one, along with pseudonymization where it's needed.

It is not worth building when a process only covers rare, one-off cases where handling them manually is faster than building and maintaining an integration, or when the company's documentation is too incomplete or inconsistent for the model to draw reliable answers from. In those cases I say so during the initial call instead of selling an integration that will not pay for itself.