
Model Context Protocol (MCP): connecting AI to company systems safely
What MCP is, when it pays off more than a conventional integration, what risks it brings and what to check before an MCP server gets access to your company’s data.
Analysis
This article is for owners, COOs, CFOs and IT managers in trading and manufacturing companies considering connecting an AI agent to their ERP, CRM or warehouse management system. An agent that only answers questions and an agent that can change an order’s status are two different projects. We show how they differ, where they deliver results, why it is usually the data rather than the model that fails, and what determines the cost.

A conversation about an AI agent in the ERP becomes concrete once someone asks exactly what it is allowed to do with the data. Read, propose and write are three different projects under one name. The table below compares their risks and the safeguards they need.
With read access, the agent answers in sentences instead of tables: what the status of an order is, how much is left in stock, when the last invoice was issued to a customer. The main risk is showing data to someone who should not see it. This is the cheapest option to start with.
With proposals, the agent prepares a document but does not approve it: a draft order, a draft correction, a preliminary match of payments to invoices. A person checks the prepared entry and approves it with one click. This level usually gives the best balance of benefit and risk.
With write access, the agent changes data in the system itself, e.g. it creates a customer record or changes an order’s status. This requires limits, data checks before saving, a log of operations and the ability to roll back changes. The projects that cause the most trouble usually jumped straight from read to write.
| Level | What the agent does | Main risk | Safeguards needed |
|---|---|---|---|
| Read | Answers questions about operational data | Showing data beyond the user’s permissions | The same permissions as in the ERP, showing only data available to the person asking, logging the questions asked |
| Propose | Prepares a document for approval | Approving on autopilot, without reading | Highlighting fields filled in by the agent, flagging items the agent is unsure of, spot checks |
| Write | Changes data in the system itself | A series of errors made faster than anyone can react | Limits on the number and value of operations, checks before saving, a log, rollback, tests after every system change |
The processes that work best are those where people switch between systems or retype data from documents in varying formats. The list below shows five such cases.
A simple estimate helps gauge the scale. Suppose 10 people in customer service each spend 30 minutes a day checking statuses in two systems. That is 300 minutes, or 5 hours a day and about 100 hours a month over 20 working days.

The most dangerous fault of an agent in the ERP does not look like a fault. The agent does not refuse to answer or invent a customer; it gives a figure calculated correctly, but according to a different definition from the one the person asking had in mind. It also gives no sign that it had to choose at all.
The question “how many active customers do we have?” has three answers if the system allows them to be counted in three ways. It might be a flag in the customer record, the absence of an end date for the relationship, or a purchase in the last 12 months.
The second form of the problem is duplicate item codes. If the same product has three codes in the item master because different people entered it over the years, the agent may create orders for the wrong one. A person makes such a mistake once; an agent repeats it across hundreds of documents.
The third is free-text fields: order remarks, call notes, comments on complaints. Commercial arrangements are often recorded only there, with no expiry date. The agent may draw a conclusion from them that was true two years ago.
A data review sounds like a separate consulting project. In a mid-sized company, it usually takes a few days of work by someone who knows the data, and the four questions in the list below, asked for every area the agent will touch.
The last question brings the most surprises. In many ERP systems, permissions are assigned to the program’s screens, not to the data. An agent that reads data directly from the database then bypasses them completely, so the permissions have to be recreated in the agent’s settings.

If the system has three definitions of an “active customer”, the agent will pick one of them – and do so faster than anyone notices it picked the wrong one.
An agent with write access needs safeguards that a traditional integration did not. An integration does exactly what it was programmed to do, whereas an agent makes decisions based on content. That is why the three mechanisms in the list below are a precondition, not an extra.
An example of a limit: the agent may create 20 documents an hour, not 2,000, and will not approve an operation above a set amount on its own. A limit does not mean distrust of the model. Every system makes errors in series, and a limit turns a series into a single problem to fix.
The price depends above all on the number of systems the agent connects to, and whether it is allowed to write to them. The AI model and the number of users usually matter less. The table below lists the main cost components and ways to limit them.
Maintenance matters more here than in an ordinary IT project. An ERP update, a new field in a master record or a change to permissions all require the connection to be adjusted and retested. An agent with no maintenance budget usually stops handling some cases quietly within a year.
In a typical implementation, the cost of the AI model itself is a smaller item than maintaining the connections. It grows quickly with poor decisions, though: overly long instructions, retrieving many documents for every question or using the most expensive model for simple document-type recognition.
| Component | What it depends on | How to limit it |
|---|---|---|
| Connections to systems | Number of systems, availability of a data exchange interface, write access | Start with one system and read access only |
| Reviewing and cleaning data | Number of inconsistent definitions, duplicates and empty fields | Limit the first stage to one process and a few concepts |
| Safeguards | Level of access, audit requirements, number of thresholds | Stay at the proposal level until accuracy is stable |
| Licences and model usage | Number of questions, length of instructions, choice of model; in Copilot Studio – credits for the whole company | A cheaper model for simple tasks, short instructions, retrieving only the data needed |
| Maintenance | How often the ERP and processes change | Tests after every ERP update, a responsible person on the company side |
A data exchange interface (API) is an orderly way for one program to fetch data from another and save data to it, much like a counter in a public office with a set procedure. Modern ERP systems have one; older ones often do not. An agent will still work, but differently and at a higher cost.
That leaves three routes. Reading directly from the database is fast, but the permissions must be recreated outside the ERP. An intermediate layer – a separate program that exposes selected data to the agent – is safer, but it has to be built and maintained. Automating clicks on the program’s screens (RPA) is the most fragile, because any change to a screen’s layout breaks it.
Such connections are increasingly built using the MCP standard, which lets you connect the same system to different AI tools without writing a separate connection for each. It does not solve the data or permissions problem, though – that still has to be designed.
Start with one process at the read level. Choose the process where people most often switch between systems to answer a single question – usually order statuses or product availability. An agent that only reads quickly shows the real state of the data before you decide on write access.
An agent will not pay off if the data is in such a state that people themselves do not trust the ERP reports. Nor does it make sense if the ERP vendor offers a built-in assistant that covers your needs. First check what your licence already includes, and only then consider your own agent for the gaps.
Related topics in the knowledge base
FAQ
Questions that management boards and IT departments of mid-sized companies usually ask after their first conversation about an agent connected to the ERP.
About this page
About the author
Kacper Włodarczyk
CEO of ALGORCOMP
Kacper specializes in implementing Microsoft 365 Copilot, Copilot Studio, Power Platform (Power Automate, Power Apps, SharePoint) and AI agents for mid-sized B2B companies in Poland. He leads dozens of projects covering AI strategy, Power Platform governance, document workflow automation and sales processes. In his writing he focuses on the practical side of AI adoption in organizations — from the first proof of concept to company-wide rollout — with particular attention to data security, compliance (GDPR, NIS2, the AI Act) and return on investment.
Meet the teamWe will go through the five most important concepts in your ERP with you and assess whether an agent can work on them safely – and if not, what needs sorting out first.
Featured

What MCP is, when it pays off more than a conventional integration, what risks it brings and what to check before an MCP server gets access to your company’s data.

How to build a golden set of reference cases from your company’s history and which measures suit each type of agent. Also: when to trust automated scoring and how to catch a silent decline after a model change.

AI agents that click through programs like an employee: how they work, how they differ from RPA robots and integrations, what they cost in Copilot Studio and how to run a first trial safely.