# AI agent development

> An AI agent is software that, given a goal, decides which steps to take and carries them out in your systems through tools: it looks up records, reads documents, drafts replies and updates data. Corrplus builds agents with narrow permissions, human approval on every consequential step and a log of everything they do, so you can put them into production with confidence.

URL: https://corrplus.net/services/ai-agent-development/
Updated: 2026-10-22
Corrplus Team

## Who it is for

- Operations teams where closing a request means opening several systems and making a judgement call (returns, supplier emails, order exceptions)
- Companies that started with n8n or RPA and got stuck on steps that cannot be written as rules
- Software and IT teams that want to expose internal systems to assistants such as Claude or ChatGPT through MCP
- Organisations that tried an agent pilot but could not take it live because of permission, cost or error control

## When it is not a fit

- Work whose steps are the same every time. Classic workflow automation is cheaper and more predictable.
- Irreversible decisions where human approval is not acceptable, such as payment release or credit decisions. There the agent only prepares the case.

## What you get

- **Task and permission design**: A written definition of the agent's goal, which tools it can use with which rights, and what it must never do.
- **Tool layer and MCP servers**: Defined read and write tools for your ERP, CRM, email and file systems, exposed as Model Context Protocol (MCP) servers where useful.
- **Human approval steps**: An approval request to the right person before any step that changes a record, reaches a customer or moves money.
- **Evaluation set**: Test cases built from real examples, so the agent's behaviour is re-measured whenever the model or instructions change.
- **Monitoring and logs**: For every run: which tool was called with which input, what it cost and what the outcome was.

## How we work

- **Discover**: We study the work you want to hand over through real examples and separate the decisions made by rule from those made by judgement.
- **Design**: We write down the tool list, permissions, approval points and which personal data reaches which model, under GDPR and Türkiye's data protection law (KVKK). The quote is based on this document.
- **Build**: We start with a version that only reads and proposes. Write permissions are opened one by one as it passes the evaluation set.
- **Run**: We monitor the logs, add wrong decisions to the test set and manage model and cost changes.

## Tools we use

Anthropic Claude and OpenAI tool use, Model Context Protocol (MCP), Google Gemini, LangGraph, n8n, Supabase and PostgreSQL, Langfuse, Microsoft 365 and Google Workspace APIs

**Timing:** The first read-and-propose version usually ships within 4–6 weeks

**Pricing:** Fixed-scope project quoted after the discovery call; model usage billed to your own account; optional monthly operation.

## What can an AI agent do, and what can't it?

Take a return request. Someone reads the email, checks the order system, looks up the shipment, applies the return policy and writes to the customer. Each step is simple, but the order and the decision change with every request. Classic automation breaks on that variation. An AI agent takes the goal, decides which tool to call in which order, and brings you the result.

An agent is not magic. It can only use the tools you give it, and only as far as you allow. Most of our work is drawing those lines: which tool, which permission, and where a person signs off.

## How is an agent different from a chatbot or an LLM integration?

These three services are often confused. The difference is who the software talks to and whether it acts inside your systems.

| Aspect | AI agent | [Chatbot](/services/ai-chatbot-development/) | [LLM integration](/services/llm-integration/) |
|---|---|---|---|
| What it does | Plans several steps toward a goal and executes them with tools | Talks to visitors and answers questions | Adds one AI feature to your existing software |
| Who uses it | Mostly internal teams, in the background | Website and WhatsApp visitors | Existing users of your software |
| Writes to systems | Yes, with permissions and approval | Usually reads only, opens a ticket | In one defined field |
| Main risk | A wrong action | A wrong answer | Cost and consistency |

A chatbot is about the quality of the conversation; an agent is about getting the work done. An LLM integration is a single, predictable feature such as "summarise this service note". All three can work together: a WhatsApp chatbot takes the request, an agent prepares the transaction in the order system, and a person approves it.

## How do we make agents safe to run?

- **Read first, write later.** The first version only gathers information and proposes. Write permissions open one at a time as results hold up on the evaluation set.
- **Human approval.** Messages to customers, record changes and anything with financial impact wait for approval in Microsoft Teams, Slack or email.
- **Standard tools through MCP.** We expose your internal systems as Model Context Protocol servers, so the same tools work across models and permissions live in one place.
- **Logs and cost.** Every run records its steps, tool calls and model cost, so any decision can be traced later.
- **Data protection.** We settle which personal data reaches which model, and in which country, before anything is built.

## Where should you start?

The best first agent handles work where someone currently jumps between several systems, volume is high and each individual decision is low risk. In the first call we identify two or three such candidates. If the rules are already clear, [workflow automation](/services/workflow-automation/) is the better starting point. For a broader roadmap we begin with the two-week [AI discovery sprint](/ai-discovery-sprint/).

## Frequently asked questions

### What is the difference between an AI agent and a chatbot?

A chatbot talks and answers questions. An agent uses tools to reach a goal: it looks up records, creates documents, updates systems. An agent can sit behind a chatbot; the real difference is whether the software acts inside your systems.

### What happens if the agent does something wrong?

That is why agents are built with scoped permissions. Steps that change records or send anything outside the company go through human approval, and every call is logged. The first version only reads and proposes; write access is opened based on test results.

### What is MCP and why would we need it?

The Model Context Protocol is an open standard for connecting AI applications to tools and data. Once your internal systems are exposed as MCP servers, different models and clients that support MCP can use the same tools under the same permissions.

### Which model do you use?

It depends on the task. We start with models strong at tool use (Anthropic Claude, OpenAI, Google Gemini) and move cost- or speed-sensitive steps to smaller models. Because the evaluation set stays fixed, behaviour stays measurable when the model changes.

### How do agents fit with GDPR, KVKK and the EU AI Act?

We decide during design which personal data the agent can reach and which model and country it goes to, masking data or running open models on your servers where needed. Risk classification and legal assessment are done together with your legal counsel.

### Will our existing n8n workflows go to waste?

No. In most projects the agent calls existing n8n workflows as tools. Rule-based steps stay in the workflow; only the judgement call moves to the agent.