2026 is the year in which every business software vendor says \u201cwe have AI\u201d. But between an impressive demo and reliable operation over live accounting data lies a chasm. This article is a map: three levels of how AI actually works over an ERP, which of them you can deploy right away \u2014 and what questions to ask anyone selling you an \u201cAI agent\u201d.<\/p>\n
AI over an ERP reliably handles the first two levels today: answering questions over live data<\/strong> and preparing actions for a human to approve<\/strong>. The third level \u2014 an autonomous agent acting on its own \u2014 so far only makes sense for low-risk tasks under supervision.<\/p>\n<\/div>\n This is the safest bet today \u2014 and for most companies the first real contact with AI over their own data. An assistant (Claude, ChatGPT\u2026) connects to the ERP and answers the questions people used to wait until the next meeting for:<\/p>\n Technically this is enabled by the MCP standard, which connects AI assistants to business systems \u2014 we explain it here<\/a>. We have connection guides for both ChatGPT<\/a> and Claude<\/a>.<\/p>\n The second level adds writing \u2014 but with a human as the final authority. The AI drafts reminder texts based on actual receivables, puts together the figures for a VAT return, proposes a purchase order based on stock levels. A person reviews and confirms; only then is the action carried out, via the ERP’s official interface with all its validations. This principle even has a name \u2014 human in the loop<\/em> \u2014 and it is the most sensible way today to let AI write into your accounts.<\/p>\n A fully autonomous agent already makes sense where the task is well-bounded and a mistake is cheap: watchdogs and alerts, sorting and matching by clear rules, preparing reports. Unsupervised accounting entries, on the other hand, are still a bad idea \u2014 not because it can’t be done technically, but because of three things: hallucinations<\/strong> (a model can be confidently wrong), accountability<\/strong> (who signs off a faulty invoice?) and audit<\/strong> (you must be able to prove who did what and why). Progress is fast; the sensible strategy is to grow level by level, not to jump straight to the third.<\/p>\n 1. Does the AI see our live data, or is it a chatbot trained on documentation?<\/strong> The difference between \u201ctells you where the button is\u201d and \u201ctells you who owes us what\u201d. 2. Is there an audit log of every query and action?<\/strong> 3. What happens on an error \u2014 who sees it and how is it fixed?<\/strong> 4. What data leaves the company, where does it flow, and are models trained on it?<\/strong> A serious vendor answers without dodging; a demo with no answers puts you at the level of marketing, not operations.<\/p>\n<\/div>\n The good news to finish with: you don’t have to change systems because of AI.<\/strong> Levels 1 and 2 can be built on top of what you already have \u2014 POHODA, Money S4\/S5, Helios, KARAT and others. The AI layer connects to your existing ERP’s data via an MCP server<\/a>, starts in read-only mode, and you decide which agendas the AI can see. How it fits into the wider puzzle of business systems is covered in our guide to ERP integrations<\/a>.<\/p>\n No \u2014 it moves the routine. Chasing numbers, retyping and preparing figures will shrink; review, decision-making and responsibility stay with people. In practice, AI helps most those accountants who are swamped with questions from sales and management.<\/p>\n<\/details>\n Usually not. An AI layer can be built on top of your existing system via MCP \u2014 no migration, no ERP replacement. Swapping systems because of AI is currently the most expensive possible route to a result you can have for a fraction of the price.<\/p>\n<\/details>\n At levels 1 and 2, answers rest on figures pulled from the ERP database, not on the model’s memory. Even so, insist that answers can be substantiated \u2014 with references to specific documents or records. Where they can’t, treat the output as a draft to be checked.<\/p>\n<\/details>\n With a level-one pilot: connect the AI read-only to a few agendas and let the team ask questions for two weeks. Our MCP servers for ERP start at 2 490 K\u010d per month \u2014 details at POHODA MCP<\/a> and for other systems<\/a>.<\/p>\n<\/details>\n","protected":false},"excerpt":{"rendered":" Every ERP \u201chas AI\u201d these days. Three levels of how artificial intelligence actually works over business data \u2014 what to deploy now, what needs human sign-off, and where to tread carefully.<\/p>\n","protected":false},"author":0,"featured_media":67182,"template":"","class_list":["post-71883","erpedie","type-erpedie","status-publish","has-post-thumbnail","hentry"],"acf":[],"_links":{"self":[{"href":"https:\/\/symmy.com\/en\/wp-json\/wp\/v2\/erpedie\/71883","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/symmy.com\/en\/wp-json\/wp\/v2\/erpedie"}],"about":[{"href":"https:\/\/symmy.com\/en\/wp-json\/wp\/v2\/types\/erpedie"}],"version-history":[{"count":1,"href":"https:\/\/symmy.com\/en\/wp-json\/wp\/v2\/erpedie\/71883\/revisions"}],"predecessor-version":[{"id":71946,"href":"https:\/\/symmy.com\/en\/wp-json\/wp\/v2\/erpedie\/71883\/revisions\/71946"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/symmy.com\/en\/wp-json\/wp\/v2\/media\/67182"}],"wp:attachment":[{"href":"https:\/\/symmy.com\/en\/wp-json\/wp\/v2\/media?parent=71883"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}\n
Level 1: AI that knows your numbers<\/h2>\n
\n
Level 2: AI that prepares the work<\/h2>\n
Level 3: the autonomous agent<\/h2>\n
Telling marketing from reality: four questions for the vendor<\/h2>\n
What this means for your ERP<\/h2>\n
Frequently asked questions<\/h2>\n
Will an AI agent replace accountants?<\/summary>\n
Do we need a new \u201cERP with AI\u201d?<\/summary>\n
What about hallucinations \u2014 won’t the AI make numbers up?<\/summary>\n
How do we start?<\/summary>\n