AI Agents for Business: Autonomous AI Agent on WhatsApp and Email

What an AI agent is, and who it's for

An AI agent for business differs from a chatbot in one specific way: it has access to tools and uses them on its own. A chatbot answers from a fixed knowledge base. An agent checks documents, calls functions in the client's systems, opens tickets, sends notifications, and can chain several steps together before handing a case to a human.

I build agents for companies where the same type of request repeats every day: HR questions, complaints, quote requests, appointment booking, order status checks. That workload eats up a team's time, and it can be automated without losing answer quality, because the agent always grounds its answers in that specific company's documents and data, not the model's general knowledge.

The target audience is small and mid-sized companies in Poland, and EU clients for the English-language version - HR, customer service, sales, operations, and agriculture teams. I run the project personally and take full responsibility for the outcome; on larger engagements I bring in Infinity Tech Group or ClawLabs.

How an autonomous AI agent actually works

The technical foundation is large language models with tool use - the agent doesn't just generate text, it calls specific tools: a document search, the client's system API, a calendar, a database. I work on the Anthropic and OpenAI stack, and I wire up integrations with client tools through MCP (Model Context Protocol), which standardizes how an agent 'sees' external systems.

To search company documents I use RAG (retrieval-augmented generation) - the agent doesn't guess, it retrieves the document passage that answers the question and cites the source. For larger document sets I combine semantic (vector) search with classic keyword search (BM25) - a hybrid approach, proven in the 'Janina' project, that returns more accurate matches than either method alone.

A single agent with a set of tools is enough for simple requests. When a process spans multiple steps and systems - a complaint that touches the CRM, warehouse, and invoicing, for example - I move to multi-agent systems: orchestration, subagents, evals, and observability, so I know what the agent did and why.

Channels: WhatsApp, email, phone, and apps

AI agent WhatsApp integration is the most common choice for end customers in Poland and across the EU today - the conversation feels like texting a person, with no extra app to install. An agent handling email reads messages arriving at the company inbox, parses attachments, classifies the topic, and either answers directly or routes it to the right person with a ready summary.

For phone calls I connect a voice assistant built on ElevenLabs - it speaks naturally, answers strictly from the client's own knowledge base, books appointments, and hands harder cases to a human. That's a separate service, but the technical core - access to the same documents and tools - is shared with the text-based agent.

All channels can draw on one knowledge base and one decision logic, so a customer gets a consistent answer whether they message on WhatsApp, send an email, or call.

Real deployments

'Janina' is an AI HR assistant that has been in production since 2026. It runs hybrid search across 40 legal acts and over 900 documents, including Polish Supreme Court rulings, cites the exact article number with every answer, tracks social-security and tax filing deadlines, generates HR documents, and integrates with Poland's KSeF e-invoicing system. Personal data is pseudonymized, and the agent runs on WhatsApp.

'AgriClaw' is a digital agronomist, currently in beta since 2026 - it combines 10-metre satellite imagery, cloud-penetrating radar, and soil moisture data. A farmer gets a one-sentence field analysis on WhatsApp, the first analysis lands 90 seconds after registration, and the agent retains memory across two or more growing seasons.

Production hosting for these agents runs on ClawLabs infrastructure: a dedicated server per agent, EU cloud or on-premise, over 35 LLMs with automatic failover, and 13 communication channels. It's the same infrastructure I build my own deployments on, and I can bring it into a client project as well.

Cost and timeline

Across the Polish and EU market in 2026, a single AI agent covering one channel (WhatsApp or email, say) typically runs from a few thousand up to the low tens of thousands of PLN (or the euro equivalent) for the build, plus an ongoing fee for maintenance and model usage. These are rough market ranges, not a fixed price list for this company - actual cost depends on the number of integrations, the volume of documents to search, and hosting requirements (client cloud, EU cloud, on-premise).

I start with a 60-minute conversation: you tell me what eats up your team's time, I tell you what can realistically be automated. Then I build a prototype - a working agent on your real data - in 2 weeks. Only after that prototype is validated do we move to production rollout and ongoing care: monitoring, updates, and cost control.

When an AI agent makes sense, and when it doesn't

An AI agent pays off where requests are frequent, repetitive, and grounded in documents that can be gathered in one place - policies, procedures, price lists, legal texts. The more repeatable that request volume, the faster the implementation pays for itself.

I don't recommend an agent where request volume is very low, company data is scattered and undigitized, or decisions require human judgment calls that can't be written down as a procedure. In those cases it's better to organize the documentation first and revisit automation afterward.

Security is a starting point, not an add-on: data can stay in the EU or move on-premise, deployments account for GDPR from day one, and the agent always escalates difficult or ambiguous cases to a human instead of guessing.