A CFO at a manufacturing company with annual revenue of 3 billion asked me last week: The automated invoice recognition system we implemented last year—why did our bill suddenly rise by 40% this quarter?
He isn't using Odoo; he's using a cloud ERP from an international vendor, with AI capabilities hooked into Power Platform. The answer is simple—after November 1, in that system's newly renewed licenses, the AI credits that originally came “free with seat purchase” are gone, replaced by separately purchased Copilot Credits billed by usage. Every invoice extraction and every document recognition now starts being priced separately.
This is not an isolated case. If you put together the recent developments in enterprise ERP and AI integration over the past few days, you will find a fact ignored by most people: the explicit benefits of AI entering ERP (saving labor, faster reconciliation) have been discussed for a whole year, but its hidden cost line has only just surfaced. Today I want to talk about three accounts.
First account: the licensing line, where AI goes from “giveaway” to “pay-per-use”
This change by Microsoft, effective November 1, looks to outsiders like a minor adjustment to licensing terms, but to insiders it is a shift in the billing paradigm.
In the past, eligible Power Platform and Dynamics 365 licenses included "seed" AI Builder credits, so when enterprises used AI actions such as invoice extraction and text recognition, they generally did not pay extra. Now, new subscriptions or renewals no longer include these credits, and the existing AI Builder capacity add-on is also counting down to retirement. After migration, AI actions in Power Apps / Power Automate will consume Copilot Credits first, while agent flows in Copilot Studio will be billed directly in Copilot Credits—and these two types of credits cannot be automatically converted between each other.
What does it mean for the CFO? AI capabilities have shifted from "CapEx-style bundling into seats" to "OpEx-style per-use metering." Previously, when finance looked at AI investment, it looked at annual subscription fees; going forward, it will have to look at "how many cents each invoice recognition costs, and how much is charged for each run of an agent flow." It will force companies, for the first time, to account for the per-use cost of every automated process the way they read a water or electricity meter.
This is also why I have always recommended that growth-stage companies take the open-source route seriously. Take Odoo 20 as an example: its AI runs on IAP credits, with publicly listed unit prices and transparent usage, so every cent you spend can be mapped to a specific action and won't be swallowed up by the gray area between "included in the seat" and "purchased separately by volume." An open-source foundation means you can see right through the cost structure yourself.
Second account: compliance line, agents must pass review before going online
The second thing happened at SAP TechEd (October 30). SAP released AI Governance Assistant, which can assess and flag AI agents that "appear high-risk but have not yet been classified" under the EU AI Act, and every output is explicitly labeled: this is a draft "for legal and compliance review."
This is not just a nice-to-have feature demo. Article 99 of the EU AI Act stipulates that penalties for violating high-risk obligations can reach €15 million or 3% of global turnover, whichever is higher. When AI agents shift from "assistants that answer questions" to "executors that can initiate transactions and modify databases," regulatory attention naturally follows to this layer. SAP turning it into a pre-launch self-check switch shows one thing: for companies with business in the EU, or that are themselves cross-border groups, conducting a compliance assessment before connecting AI agents to ERP has shifted from "optional" to "mandatory."
There is a key point that is often overlooked: compliance responsibility does not shift to the vendor just because you use a major vendor's system. What decisions the agent makes on your data, which record it changes, and which set of rules it relies on—the final signatory is still the controller on your side. So whether you use SAP, Oracle, or Odoo, the governance logic is the same—authenticated identity, least privilege, full traceability of actions, and rollback of anomalies. The open-source foundation actually has an advantage in this respect: you write the rules yourself, keep the logs yourself, and own 100% of the data, without relying on the vendor to do "black-box compliance" for you.
The Third Ledger: The Acquisition Bill and the ROI Question in AI M&A
The third account is the most heart-wrenching, and also the most easily overshadowed by the grand narrative of financial reports.
SAP's third-quarter financial report disclosed a figure: the AI startup Prior Labs, which it acquired for about €518 million, generated less than €1 million in sales revenue for SAP in the previous quarter. Meanwhile, SAP paid a total of about €2.1 billion during the year for three data / AI companies: Reltio, Dremio, and Prior Labs.
From the buyer's perspective, this raises a very practical question: when "building in-house or acquiring AI capabilities" becomes an arms race among ERP vendors, as an enterprise customer, how do you distinguish which are capabilities that can already be put into production and have a clearly calculable ROI, and which are merely forward-looking stories on the vendor's balance sheet? The former deserve to enter your processes; the latter should only enter the vendor's PPT.
The way to judge is actually simple: see whether it can, on your own data, compress a specific process (such as monthly closing reconciliation or supplier risk assessment) from N days to N hours, and whether you can clearly explain where the money saved or earned comes from. If you can't explain it clearly, treat it as a "future capability" and don't write it into this year's budget return assumptions in advance.
Three accounts, one main thread
Connecting these three developments makes the main thread clear: as AI enters ERP, the labor saved is only the tip of the iceberg above the water. Beneath the surface lie three hidden cost lines: the reconfiguration of license billing, the front-loading of compliance responsibility, and the uncertain returns on vendors' AI acquisitions.
For CIOs and CFOs who are currently selecting solutions or preparing to scale up, I have three honest points:
First, change AI's unit of measurement from "seats" to "actions". The contract must clearly specify: the per-use cost cap for each high-frequency automated action, and the alert threshold for usage overage. Don't wait until the bill arrives to find out.
Second, write governance policies before use. First determine who configures, who reviews, and who is responsible for data integrity, then grant the agent execution permissions. Never set high-stakes actions (changing accounts, making payments, issuing orders) to "fully automatic".
Third, keep sovereignty over the foundation in your own hands. Whether you choose open source or closed source, ensure that data can be exported, logs can be retained, and costs can be verified. The ledger can be handed over to an agent to operate, but responsibility and control cannot be outsourced to a black box that charges by usage.
Ultimately, as enterprise digitalization reaches the AI stage, the competition is not about who plugs agents in first, but about who first gets these three accounts straight. Get them straight, and AI is leverage; fail to, and it is just a new bill you cannot make sense of.
