AI agents in your ERP: five controls to demand before opening up your data

Oct 6

4 min read

On 1 October, the software vendor Acumatica unveiled its 2026 R2 release: AI agents that act inside the ERP, and an open door for Claude, ChatGPT or Gemini through the MCP protocol.

For a Swiss SME, the question is no longer whether AI will enter the ERP, but on what terms. Here are the five controls we recommend demanding before you connect anything.

What Acumatica just announced, and why it goes beyond one vendor

Acumatica is a cloud ERP widely used by distribution, services and manufacturing SMEs. Its 2026 R2 release, announced on 1 October 2026, puts AI at the heart of the product, according to the report published by ERP Today.

Three new features caught our attention, because they describe what most ERPs will soon offer:

AI agents configured in natural language, triggered by a business event or an existing workflow, and able to hand tasks over to one another;
MCP servers that give external assistants read-only access to published ERP queries;
logs that record every access and all AI consumption, with alerts when thresholds are reached.

The examples the vendor gives are telling: price adjustments, lead qualification, credit risk assessment. These are no longer writing gadgets, they are management decisions.

Acumatica is not an isolated case. Most major ERP and CRM vendors have announced agents in recent months. If your ERP does not offer them yet, it probably will with its next major update.

The real shift: AI moves from writing to acting

Until now, AI in SMEs was mostly used to write. An employee pasted text into an assistant, reviewed the answer, then made the decision themselves. The risk stayed contained, because a human stood between the machine and your data.

According to the UBS study published in May 2026, based on around 2,500 Swiss companies, 60% of businesses already use AI. Automation is the leading use case, cited by 41% of them.

Yet only 5% place AI at the core of their business model.

An agent inside the ERP changes that equation. It reads customer, inventory and invoice data, then proposes or executes an action: change a price, block an order, chase a debtor.

The potential gain is real, because the ERP is where the company's numerical truth lives. So is the risk: a mistake is no longer a clumsy sentence, it is a wrong accounting entry or an annoyed customer.

That is why the question to ask your vendor or integrator is not "what can the agent do?", but "how is it controlled?".

The five controls to demand before switching on an agent

Here is the checklist we use with our clients. It boils down to five questions, and each one needs a written answer before anything goes into production.

1. Does the agent run under a defined role?

An agent should never have global access. It inherits the permissions of a specific role, ideally created for it, and limited to the screens and entities it actually needs.

2. Which data is sent to the language model?

Ask for the list of fields sent and the ability to mask some of them: salaries, bank details, negotiated pricing terms. The Swiss Federal Act on Data Protection also applies to these flows.

3. Is access for external assistants read-only?

Connecting Claude or ChatGPT to the ERP through MCP is useful for querying your figures in plain language. Start read-only, on published and approved queries, never on the full set of tables.

4. Is every action logged and reversible?

You must be able to answer the question "who changed this price, when, and on what basis?". If the agent acts, the log must say so, and the action must be reversible.

5. Is the cost capped?

Agents consume AI units billed on usage. Set a monthly budget and threshold alerts, as Acumatica now allows, to avoid a nasty surprise on the invoice.

If your vendor cannot answer one of these questions, that is not necessarily a deal-breaker. It is a signal to wait, or to frame the use with a clear internal procedure.

An example: debt collection at a distributor in Vaud

Take a case representative of what we see: a technical equipment distributor with 45 employees, based in the canton of Vaud, issuing around 9,000 invoices a year from its ERP.

Chasing late payers takes an accountant two days a month. She checks due dates, reviews each customer's history, writes the emails and decides who deserves a phone call rather than an automatic reminder.

An agent can prepare most of this work. Every morning, it spots overdue invoices, ranks customers by risk level and drafts a reminder suited to the relationship.

Applied with the checklist, the setup looks like this:

a "collections agent" role that reads invoices and contacts, but can neither issue credit notes nor change a credit limit;
pricing terms and bank details masked before anything is sent to the model;
reminders sent only after the accountant has approved them during the first two months;
a monthly consumption cap and a log reviewed at the end of each month.

The target outcome: the accountant moves from data entry to decision-making. The two days a month become half a day of review, and sensitive cases finally get a real phone call.

The benefit does not come from the agent alone. It comes from the company first clarifying who decides what, then delegating the repetitive part.

What we take away

AI agents arrive in the ERP through an update, not through a project. The main risk for an SME is therefore not missing the wave, but switching it on without having framed it.

A well-controlled agent is one more team member; a poorly controlled agent is a faceless administrator account.

Our advice comes down to three moves:

list now the AI features planned in your ERP's roadmap;
pick a first process that is repetitive, measurable and low-risk, such as payment reminders or invoice matching;
apply the five controls before activation, and document them.

Would you like to assess where your company stands before getting started? The 39 Advisory AI Barometer gives you a first reading in a few minutes, and we are happy to discuss it with you.