An Exact Online AI agent is an autonomous AI system that executes multiple steps in your Exact Online administration without you approving each action. While an AI assistant waits for your next question, an agent plans the sequence of actions itself — reading an email, looking up a customer, creating a sales order, sending a confirmation — and runs them in sequence. This article: how it works technically, what you need to start today, and which pitfalls to avoid.
What is an Exact Online AI agent?
An Exact Online AI agent is software that translates natural language into concrete actions in your administration. The agent uses a large language model (LLM) like ChatGPT, Claude or Copilot to make decisions, and an MCP server like Ledger Botje to convert those decisions into API calls to Exact Online.
The difference with conventional integrations is autonomy: an agent receives a goal ("process all outstanding order requests") and decides itself which steps are needed to reach that goal. No predefined workflow, no if-this-then-that. The agent improvises within the boundaries you set.
AI assistant vs AI agent: the difference in one sentence
An AI assistant responds to questions, an AI agent executes tasks. More concretely:
- Assistant: "How many outstanding invoices do I have?" → the AI fetches and shows it.
- Agent: "Send reminders for all invoices overdue by more than 30 days and log the action in CRM." → the AI fetches the list, drafts the emails, sends them, and logs the action — without per-step approval.
Read also our article on AI agents for Exact Online automation for a broader overview of applications.
How does an Exact Online AI agent work technically?
An AI agent for Exact Online consists of three components:
- An LLM that plans and reasons (Claude, GPT-4, Gemini, etc.).
- An MCP server that manages the connection between LLM and Exact Online. Ledger Botje exposes 219+ tools (read and write) via the Model Context Protocol.
- An agent runtime that runs the agentic loop: goal → plan → tool call → evaluation → next step.
The agentic loop is the heart of every AI agent. At each iteration, the LLM receives the goal, the current state, and the list of available tools. It picks a tool, runs it via MCP, reads the result, and decides what to do next. The loop stops when the goal is reached — or when a preset limit (number of steps, time, cost) is exceeded.
7 tasks an Exact Online AI agent takes over
1. Processing incoming orders
The agent reads an order request email, recognises the customer (or creates a new one), checks stock, creates the sales order in Exact Online, and sends a confirmation. Time per order: ~30 seconds instead of 4–8 minutes manually.
2. Matching purchase invoices
The agent links each purchase invoice to the corresponding purchase order and goods receipt. Quantities and prices match? Book it. Discrepancy? Flag for review. A typical SMB saves 1–3 hours per day here.
3. Proactive stock reordering
The agent monitors stock levels, compares them with minimum stock and outstanding sales orders, and places purchase orders with preferred suppliers when stock drops below the reorder point. No stockouts, no overstock.
4. Automated debtor management
The agent runs a daily check on overdue invoices, segments by ageing category, drafts reminder emails with the right tone per category, and flags critical debtors for human follow-up.
5. Monitoring production orders
For manufacturing, the agent tracks production order progress, signals material shortages or delays on the shop floor, and automatically books material issues and shop order completions based on scan input.
6. Planning warehouse transfers
The agent calculates which items need to be transferred between warehouses based on outstanding sales orders, and creates the transfer orders including route and quantities.
7. Generating periodic reports
Every Monday, quarter or year-end, the agent gathers the relevant data, builds a report (revenue, margins, debtors, stock turnover) and emails it to the right people. No more manual Excel exercises.
What do you need to connect an AI agent?
For a working Exact Online AI agent you need three things:
- An Exact Online subscription with API access — applies to virtually all business plans.
- An MCP server with write access — Ledger Botje Advanced (€59/month) gives access to all 219+ tools, including write actions.
- An AI tool that supports agentic workflows — not every LLM can run agent mode, but most mainstream tools can (see below).
Setup time: under 5 minutes. No developers, no API documentation, no OAuth implementation. Follow the MCP setup guide for your AI tool of choice.
Which AI tools work as agents with Exact Online?
Almost every modern AI platform with MCP support can function as an agent for Exact Online:
- Claude — strong in autonomous workflows, native MCP support since late 2024.
- ChatGPT — Custom GPTs and agent mode (2025).
- Microsoft Copilot — agent functionality within Microsoft 365.
- Google Gemini — agent capabilities in Workspace.
- Cursor — for developers building or integrating agents.
- OpenClaw — open-source with full flexibility for custom agents.
The choice depends on your work environment: already using Microsoft 365? Copilot is the obvious pick. Working in Google Workspace? Gemini. Want maximum autonomy? Claude or OpenClaw.
Working safely with autonomy: what to set up
An agent that executes actions independently sounds scary — and that’s healthy. A good setup starts with boundaries:
- Per-tool write access: Ledger Botje lets you decide per tool whether the agent can write. Limit to what is strictly necessary.
- Audit log: every action is recorded with timestamp, user and payload. Available via the admin dashboard.
- Amount limits: set maximums per transaction (e.g. purchase order < €5000 without approval).
- Human-in-the-loop for critical paths: let the agent generate proposals, you approve at high amounts or new suppliers.
- Read-only sandbox: test first with a Basic plan (read-only) before enabling write access.
More detail on the security architecture in MCP security for Exact Online.
Step by step: your first agent in 5 minutes
- Create a free account on the Ledger Botje pricing page (14-day free trial, no credit card).
- Connect Exact Online via OAuth — takes 30 seconds.
- Install the MCP connection in your AI tool of choice (see setup instructions).
- Give your first agent task, for example: "Create a report of all outstanding invoices over 30 days old, grouped by customer, and draft reminder emails."
- Review the result, refine your prompt, and gradually expand to autonomous workflows.
Frequently asked questions about Exact Online AI agents
What does an Exact Online AI agent cost?
Ledger Botje Advanced (€59/month) contains everything you need for an agent with write access. On top of that you pay for your AI tool (ChatGPT Plus from €20/month, Claude Pro from €18/month, Copilot from €22/month). Total: ~€80–90 per month. Read more about the costs of an Exact Online AI integration.
Is an AI agent safe enough for my accounting?
Yes, provided you set boundaries. Ledger Botje uses OAuth 2.1, encrypted token storage and logs every action. Combine this with amount limits and human-in-the-loop for critical actions. For 95% of routine operations (creating orders, querying stock, matching invoices), full autonomy is safe.
Does an AI agent also work with the Exact Online API directly?
Technically yes, but not practical. The Exact Online REST API requires OAuth implementation, rate-limit handling and code per tool. MCP via Ledger Botje does this for you and exposes the API as 219+ ready-made tools that any AI can use directly. See MCP vs Exact Online API.
Can I build an AI agent myself?
Yes, with agent frameworks like LangChain, AutoGen or the Anthropic SDK. But you still need an MCP server to talk to Exact Online. Ledger Botje serves as that layer — build your own logic, use our tools for the Exact Online side.
How does an AI agent differ from RPA (Robotic Process Automation)?
RPA follows scripts: steps 1, 2, 3, 4, in fixed order. An AI agent improvises: it picks which steps are needed based on context. RPA breaks at every UI change. An AI agent adapts. For modern ERP automation, an agent is more flexible and cheaper to maintain.
Which LLM is best for Exact Online agents?
For stable, autonomous workflows: Claude (strong tool use, long context). For broad availability: ChatGPT. For Microsoft integration: Copilot. In practice, all major models work fine — pick the tool that best fits your work environment.
Ready to set up your own Exact Online AI agent?
An Exact Online AI agent is no longer a future technology — it is workable today for any SMB with an Exact Online subscription. Start small (a report, an order intake), see the time savings, and expand to autonomous workflows once you have confidence.
See the complete Exact Online AI integration for the full overview of what AI can do in your administration, or start directly via the pricing page. Find inspiration for your first prompts in the prompt library.