AI agents for companies: agents that work inside your processes.
I implement AI agents in companies. An agent reads emails and documents, checks data in your systems, prepares a reply, a report or a decision, and hands the case to a person wherever a person should decide. I take every agent to production and answer for how it behaves.
- 12 h → 2 minassistant for an advisory firm: a research agent over more than half a million documents
- 200+workflows at an e-commerce client, including agentic product research
- 190,000+users of n8n-mcp, the tool AI agents use to build workflows
An AI agent is not a chatbot. A chatbot answers a question. An agent gets a task, pulls the data it needs through tools (CRM, inbox, document store, APIs), works through the steps and ends with a result someone can check.
Many agents never get past the demo, because a demo does not have to handle exceptions, missing data or the moment the model gets it wrong. That is where I start: what the agent may decide alone, what it only proposes, when it hands over to a person, and how we will know it made a mistake.
Agents I implement
Document and knowledge agents
The agent searches the company's documents, brings the sources together and drafts an analysis with links to the passages it relied on. That is how the assistant for an advisory firm works, across more than half a million documents.
Conversation analysis agents
Call and interview transcripts turned into structured findings: the agent answers every question on the agenda, checks each answer against the transcript and flags what nobody asked about.
Data-checking agents
Every day they compare what the order, contract or plan says with what is actually happening in the system, and flag discrepancies to the person responsible before they turn into cost.
Enquiry and sales agents
Lead qualification, ticket triage, drafting a quote or a reply from CRM data. The agent prepares; a person approves and sends.
Research agents
They gather information from the web and from your systems, put it into one format and prepare it for a decision: product research, competitor analysis, enriching data on customers and suppliers.
Shared memory for a company's AI agents
Every employee works with their own AI assistant, and every one of them starts from zero. I connect their assistants (ChatGPT, Claude, Codex) through one MCP connector to a shared company memory: each person keeps private notes for their own assistant, while decisions, who is working on what and how the systems work go into a shared space that colleagues' agents search. Sign-in with the company account, data on your own server. The first implementation is in pilot at an e-commerce client.
System maintenance agents
When an automation fails, the agent pulls the error context, the node configuration and the last 30 runs by itself. It states the cause with a confidence level, assesses the impact and proposes the smallest fix; the same error is diagnosed once a day and repeats are only logged. An agent like this watches my own n8n instance, and a person approves every fix. The same pattern extends from errors to anomalies: a drop in volume, unusual values in the data, runs that never happened.
Who this is for
- Companies where several people read documents, emails or reports every day to make the same kind of decision.
- Teams that already have an agent demo in ChatGPT or Copilot Studio and do not know how to get it into production.
- Companies whose data is spread across several systems that nobody has time to reconcile.
- Organisations whose data cannot leave their own infrastructure.
How I implement an agent
- 01Scope of decisions
Before the agent exists, I agree with the process owner what it does alone, what it only proposes and when it hands over to a person. An agent that changes nothing without approval is easier to roll out and easier to defend in front of IT.
- 02Built on real data
I build and test the agent on your data, edge cases included: empty fields, wrong formats, duplicates, a system that does not respond.
- 03Trial run
The agent works alongside a person and we compare the results. It goes into production once the result is repeatable, not the first time it works.
- 04Production and handover
Error handling with alerts, monitoring, documentation anyone can take over, and a trained process owner who knows how to use the agent and whom to call.
Agents in production
Agents running at clients today, and one you can talk to on this site.
Over 200 workflows on one platform — marketplace automation across 6 countries, agentic product research, custom microservices and an internal automation team trained to work independently.
Read the case studyA self-updating base of more than half a million documents, hybrid search, a research agent and analysis drafts. From ~12 hours of work to ~2 minutes.
Read the case studyA due-diligence consultancy's interviews turned into structured findings automatically — AI answers every agenda question, validates each answer against the transcript, and catches the unprompted disclosures that matter in a deal. Runs entirely on the firm's own in-country cloud.
Read the case studyA voice agent on aiadvisors.pl: a voice model, a reasoning model in the background, my own gateway and more than a dozen n8n workflows. Answers from the site's knowledge, qualifies enquiries and books meetings.
Read the case studyAgents for specific channels
An agent that talks to customers or connects systems has requirements of its own. I describe those services separately.
A voice agent answers the phone and talks to visitors on your website, in several languages.
A text assistant answers from your company's documents and handles customer enquiries.
n8n automation in which the agent is one step of the process.
The technical layer: MCP servers, RAG and model fine-tuning when off-the-shelf tools fall short.
Guides
How to build shared memory for a company's AI agents, how the AI Agent in n8n works, the tool I often build agents on, and how an AI agent builds workflows by itself.
How do you build a company brain for your team's AI agents?
A company brain is memory your team's AI assistants write and read together: each person keeps private notes for their own assistant, while decisions, work in progress and how systems work go into a shared space that everyone's agents search before they start. It connects to ChatGPT, Claude and Codex through one MCP connector and runs on the company's own server. Unlike a knowledge base, it is written by agents during the work, not by people after the fact.
Read the guideWhat is the n8n AI Agent and how does it work?
The n8n AI Agent is a node in which a language model receives instructions (a system message) and a set of tools, then decides on its own which tools to call to complete a task. Without tools it is just a chat. With tools, most often sub-workflows, the agent can read a document, check data in a spreadsheet and make a judgement call no plain rule can handle. Since August 2026 n8n also has a separate Agents category: agents with their own memory, channels and skills that live outside a single workflow.
Read the guideHow do you connect Claude Code to n8n through MCP, and what can the agent do then?
Claude Code connects to n8n through the n8n-mcp MCP server: one terminal command adds the server, and the agent gets documentation for every n8n node, configuration validation and workflow operations on your instance. From then on it can build, fix and test a workflow from a description in plain language. More than 190,000 people use n8n-mcp, and its telemetry shows agents spend four times more calls reading and checking than writing.
Read the guideCommon questions
- How is an AI agent different from a chatbot?
- A chatbot answers questions. An agent carries out a task: it pulls data through tools such as your CRM, inbox or document store, works through the steps and ends with a result a person can check and approve.
- How much does an AI agent for a business cost?
- Mostly on how much infrastructure you already have. If you run n8n or have a working pilot, configuring and testing an agent starts at around EUR 1,000 net. Building everything from scratch, with several channels, integrations and agents, is usually up to around EUR 10,000 net. I give the exact figure after a short brief, before any work starts. You pay AI model and infrastructure costs directly to the providers, with no margin from me.
- How long does it take to implement an AI agent?
- A first agent usually takes a few weeks from the moment data access is in place. Waiting for that access is usually the longest part, not the build.
- Which models do you build agents on?
- I pick the model for the task: OpenAI and Anthropic models, also inside the client's own cloud such as Azure OpenAI, and, where it makes sense, models fine-tuned for one specific job. Your data does not have to leave your infrastructure.
- Can the agent break something in our systems?
- Only if we let it. By default the agent reads data and prepares a proposal, and a person approves any change in the system. The agent gets write access only once the trial-run results justify it.
- Can the agent fix workflows by itself?
- Today my maintenance agent diagnoses the error and proposes a fix, and a person applies it. Over time the agent can be allowed to fix selected, low-risk classes of errors on its own, but only once its proposals have proved themselves in practice.
- We already have an agent built in Copilot Studio or ChatGPT. Can you help?
- Yes. I start with a review: what the agent does, on what data and where it goes wrong. Sometimes fixing it is enough; sometimes it has to move somewhere it can be tested and monitored.