AI agents that see a task through from start to finish.
An agent checks email, CRM and ERP on its own and sees the task through. When a decision costs money, it waits for a person to approve it.
What is an AI agent.
It is a program built on a language model that gets a goal and chooses the steps to reach it. It uses tools: reads email, checks data in the CRM and ERP, drafts documents. After each step it looks at the result and decides what to do next.
The same client message, handled three ways
Chatbot
Answers questions
Sends a link to the customer portal and asks the customer to write to sales about extra units.
Automation
Runs fixed steps
If the email has an order number, it sends the status from the ERP. It can’t handle the request for extra units, because that isn’t in the script.
AI agent
Chooses the steps itself
Finds the order by company name, checks its status and the price list, replies to the customer and prepares the new order for a rep to approve.
When the steps are always the same, you don’t need an agent. Use process automation.
Examples of AI agents in a company.
Each agent has one job, access only to the systems it needs and clear points where a person decides.
Customer service agent
Answers questions about orders, payments and complaints using the ERP and the client’s history in the CRM.
a person decidesreturns, discounts, exceptions
Quote agent
Turns a request into a quote: picks products, checks the price list, stock and the client’s terms.
a person decidessending the quote, off‑list prices
Meeting prep agent
Before a call it gathers information about the client’s company, past contacts and open issues. After the call it updates the CRM.
a person decideswhat goes to the client
Company knowledge agent
Answers staff questions about procedures, contracts and products from your documents, with a link to the source.
a person decideschanges nothing, only reads
Reporting agent
Every week it pulls data from several systems, compares it with the previous week and writes what changed and why.
a person decideswhat to do with the findings
An agent gets access like a new hire.
We connect the agent to your systems through an MCP server: one entry point through which the agent sees only what you allow.
- Reading and changing are separate. The agent can check data in the ERP but places an order only after approval.
- Every action is logged. You can see what the agent did, when and based on which data.
- The server stays with you. The next agent, and your team working in Claude, can use the same entry point.
- reads, replies about order status
- CRM
- reads, adds notes
- ERP
- reads, new order after sales rep approval
- Price list
- reads
- Accounting
- no access
10:14:06 crm.find_client “Client Ltd”
10:14:07 erp.status SO/4471 → shipped
10:14:09 erp.new_order awaiting approval: Mark
Why agent projects fail.
40%+
of agentic AI projects will be cancelled by the end of 2027 due to rising costs, unclear value or weak risk controls.
Gartner, June 2025An agent where something simpler would do.
We start with the simplest solution. We build an agent only where the next steps depend on the situation.
Value nobody calculated.
Before building, we measure how long the task takes today and what the agent’s work will cost.
Full autonomy from day one.
The agent starts with proposals your team approves. Its autonomy grows with its results.
No control over what the agent does.
Access, limits and the action log are agreed before launch, not after the first mistake.
The agent’s autonomy grows with its results.
- 01
Discovery Sprint
In one week we pick the task where an agent pays back fastest and calculate the cost of its work.
€1,900 · deducted from the build
- 02
The agent proposes
The agent drafts replies and actions and your team approves them. We see where it gets things wrong and fix it.
- 03
The agent acts
The agent handles simple cases on its own, while exceptions and money decisions still go to people.
- 04
Handover
Code, MCP server, documentation and a trained team. The agent stays with you, not with us.
This is for you if
Companies come to us when the decision on AI is made and the plan isn't there yet.
- Your company has up to a hundred people and is growing.
- You’re hiring an AI person, or it is already part of someone’s role.
- You don't have an IT department, or your tech team is a few people.
- Your team has tried new tools, but the way you work hasn’t changed.
Let's talk.
Tell us what isn’t working today. We’ll tell you straight whether we can help and which model to start with.
- 30 minutes with a founder, not a salesperson
- A concrete answer on where to start
- No commitment
Typical reply within 24h.
Prefer to pick a time right away?
Questions about AI agents.
The questions we hear most before a rollout. If yours isn’t here, ask us on a call.
A chatbot answers questions. An agent does tasks: it checks data in your systems, drafts documents and closes the case, asking a person for approval where needed.
Automation always runs the same steps. An agent chooses the steps to fit the situation. If your process looks the same every time, automation is enough and costs less.
Only where you allow it. At first the agent prepares actions and your team approves them. We widen its autonomy once the results show it is safe.
The agent works within its access and limits, and a person approves anything that costs money or goes to a client. Every action is logged, so you can see what the agent did and why.
Mostly Anthropic’s Claude models, and OpenAI models when they suit the task better. We pick the model for the task and the cost of running it.
MCP (Model Context Protocol) is an open standard that lets models use a company’s data and tools through one controlled entry point. You decide in it what the agent can read and what it can change.
Before the rollout we agree what data the agent processes, in which tools and who has access to it. Personal data is handled in line with GDPR, and we write the rules down together with you.
We start with a Discovery Sprint: one week for €1,900, deducted from the build. After it you know the cost of building the agent and the monthly cost of running it.