On the morning of November 5, SAP published a rather encouraging article: Danish wholesaler Lemvigh-Müller had three agents collaboratively handle supplier order confirmations—100,000 a year, with more than 90% untouched. That afternoon, ERP LEADERS published a short commentary, and the line in its title, "Takeda's numbers are forecasts," made me laugh.
Not to be argumentative. That short commentary went on to count four unanswered questions: how much the agent costs, who handles the remaining less than 10%, how many people this work originally required, and when it went live. None of them were answered. In the same SAP article, Takeda Pharmaceutical's 10% productivity improvement and 25% reduction in stockout losses were described as expected results.
I looked at these two articles side by side for a long time. It wasn’t that I thought the 90% was inflated, but rather that I realized one thing: AI in ERP is already running fast enough—so fast that we haven’t even had time to agree on how to issue it an acceptance form or how to settle payment with it.
DualEntry client Graphite's monthly closing cycle was used as a sample for "billing by workload"
1. The Bill Has a New Yardstick
On the same day, another agentic ERP vendor called DualEntry, through BusinessWire, put forward a term: Agents as a Service—charging for the work completed by agents, rather than for how many people log in to the system.
I wrote down a line from co-founder Benedict Dohmen: SaaS has always only competed for software budgets, while agents compete for headcount budgets. He had a more vivid one too: SaaS gives every accountant an account, AaaS gives every accountant a team. The company's examples are Slash, where two people manage the finances of $600 million in annual recurring revenue; and Graphite, where the monthly close was cut from 20 days to 1 day.
This is not an isolated voice. On November 1, Yonyou was already talking about “moving from selling software to taking on business outcomes” and exploring outcome-based charging. The two companies, four days apart and half a world away, are saying the same thing: the pricing anchor for software is shifting from “access rights” to “output.”
Seat system
What you are buying is “how many people can use it”; costs grow linearly with the number of users, and whether it is used well has nothing to do with the vendor's revenue.
Workload system
What is being bought is “how many items are completed”; if no results are produced, payment should not be made, but how to count “one item” has become a new problem.
That sounds favorable to the client, but I’d like to pour a little cold water on it. Seat-based pricing has at least one advantage: the bill is easy to understand—50 users times the unit price, and finance can glance at it and know whether it’s right. Switch to workload-based pricing, and the first thing to argue over is the definition—does one agent run count as one unit, or does it count only after a person approves and posts it? Does identifying a discrepancy count as completion, or must the discrepancy be resolved? Should the three retries after a failure be billed?
If the terms aren't written into the contract, the new billing model will be harder to understand than the old one. Especially in finance scenarios, "processing" and "posting" are two separate actions, with approval, period locking, and audit trails in between. DualEntry itself admits that for agents to autonomously complete accounting work, they must obtain the complete financial context behind transactions, and these controls must remain effective regardless of whether the action is initiated by a human or an agent.
In other words, when the billing unit changes, the acceptance criteria must change accordingly. Previously, software acceptance was about “whether the feature exists”; in the future, it must be about “whether this thing was done correctly, and who can prove it was done correctly.”
II. 90% is a good number, but it is not an acceptance form
Back to the SAP article. Its technical formulation actually carries a lot of weight: agents directly execute production orders and issue warehouse tasks within SAP Cloud ERP, with no middleware between decision-making and execution. ERP LEADERS' reminder to architects is also precise—since there is no middleware, the access rules of the core system are the agents' rules.
This sentence is worth posting on the wall of the project room. In many companies, permissions and segregation of duties are inherently fuzzy, held together by veteran employees' tacit knowledge. Once agents are introduced, the fuzzy areas will not disappear; they will only be amplified at machine speed.
| Numbers / Wording | Nature | What else is missing for making investment decisions? |
|---|---|---|
| 100,000 confirmations / year, >90% contactless | Published actual measurements | Original headcount, launch time, cost savings |
| Takeda productivity +10%, stockout losses -25% | expected results (expected value) | Measured window and control baseline |
| agent executes directly without middleware | Architecture Facts | Permission model, segregation of duties, and audit trail |
| Gartner's Inaugural Supply Chain AI Agent Quadrant | Category Assessment (Report 10-28) | The comparability of our own business processes within the quadrant |
Gartner's first formal evaluation of this category (Market Shapers is one of the tiers) is itself a signal: supply chain agents are no longer a concept from launch events; they have become a market that can be compared side by side. In the same 2026 supply chain technology trends, there is also a prediction that by 2030, half of cross-functional supply chain solutions will let agents autonomously execute decisions, rather than just give recommendations.
But when it comes to your own company, do not copy the word "autonomy" verbatim. A relatively stable approach in the industry now is to grant permissions in three tiers: let the agent perform routine and reversible actions, with a record left after completion; for time-sensitive but medium-risk matters, let the agent act first, and a human can intervene within a set window; for high-stakes or never-before-encountered situations, the agent is only responsible for preparing analysis and alternatives, and humans still make the final decision.
I've seen many projects fail, not because the model isn't smart enough, but because "can demo" was directly treated as "can hand over." A demo only needs to run successfully once; acceptance requires it to run correctly a hundred times, and you also have to clearly explain who found the one time it went wrong.
III. Odoo 20's Answer: It's Only Done When It Can Be Proven
From November 3 to 5, several Odoo implementation partners coincidentally dug out details about the implementation side of Odoo 20 AI. This part is most useful for those who actually need to go live.
First, the good news. Odoo 20 Enterprise's agent has built-in sources, Restrict to Sources, and inline citations [SOURCE: id]. Facts in answers that come from knowledge sources will have clickable citations, and will only link to sources the current user has permission to access; citations will not become a backdoor for unauthorized reading. When there is not enough information, it will clearly state that there is no answer in the provided sources, rather than casually making one up.
That checklist from Auboros gave five guardrails; let me go through them in my own words:
- Read first, then write: facts can only come from records actually read during this run, not from the model's memory-based impressions of your company.
- Cite item by item: amount, date, name, status; each must point to an invoice number, order number, or work order number.
- If records are missing, abstain from answering: if you can't find it, just say you can't find it, ask questions, don't fill in the gaps.
- Recompute the numbers with plain code: reopen the referenced records, add up the total, and if it doesn't match the number in the draft, hold it.
- Drafts are always left to humans: before approval, nothing is sent, posted, or changed.
Citations only prove that the agent has seen it, not that it read it correctly. Auboros gave a painfully apt example: $4,812.50 was written as $4,182.50, and the numbers are so close that the human eye almost never stops on them. What can really catch it is the fourth recalculation, and that one is not built into Odoo 20, nor is it in MYOB Acumatica 2026.1. In addition, inline citations cover knowledge-type sources; invoice amounts read back through tools do not automatically carry citations, so you have to write the rules yourself in the instructions.
This is not just technical perfectionism. In its guidance for commercial AI products, Australia’s OAIC requires companies to be able to identify “the data and systems used to generate that AI output”; the National AI Centre’s adoption guidance requires continuous testing and monitoring and the retention of human control. In ERP terms, this means drafts, referenced records, recalculated results, and approvers—four things kept on the same document. Without these, that output is, in legal and audit terms, a guess with a company logo.
There are two more charges that are easy to miss. First, cost attribution: Odoo AI uses IAP credits, which are accounted for separately from user licenses. Vauxoo's advice is very practical—hold your own API key; whoever holds the key pays; use two separate keys for testing and production, each with its own monthly cap; otherwise, a misconfigured test agent can consume an entire month's budget. Second, inference location: after checking the IAP policy on November 5, Wavect found that Ask AI and AI Agents use global inference, and from v20 onward automatically select models; prompts, context, and retrieval materials may be sent to the vendor. Having the database hosted in the EU does not mean AI inference happens in the EU. For those with cross-border business, this field on the procurement questionnaire must be answered separately.
By the way, a note on upgrades: Odoo 19's agent uses Topics, 20 uses Skills, and the 20.0 API-key documentation was still being built at the time. So when upgrading from 19 to 20, AI workflows need to go through acceptance all over again like financial processes, keep a release register, and don't assume that "if the interface opens, the process is alive."
IV. Before signing the contract, sort out the three accounts first
First, the measurement ledger.Choose one: per seat, per call, or per completed item, and write “completion” as a decidable condition: which step counts as delivery, who reviews it, and what anomaly rate triggers a retrospective. Don't let the vendor define both the workload and completion.
The second entry: the ledger of evidence.Every external document must be traceable to its source record, its numbers must be recomputable by an independent piece of code, drafts must pass human review, and these audit trails must be kept together with the document. Referencing and recomputation are two separate things; the platform may provide the former, but you'll basically have to write the latter yourself.
Third entry: the boundary account. Who holds the keys, how much each call costs, which region the context is sent to, and whether the hosting region covers inference—these four questions require written answers. Confirm once more: the permission rules of the core system are the permission rules of the agent.
A word of truth for CIOs and finance leaders: what you should do most right now is not choose which model is stronger, but first produce the acceptance checklist. You can pilot just one process—for example, supplier order confirmation or accounts receivable dunning emails—and hard-code the unit of measurement, reviewer, exception fallback, and audit trail location, then run it for a month. If it runs smoothly, then discuss expanding coverage; if it does not, you only lose one month.
Gartner says that by 2030, half of cross-functional supply chain solutions will execute decisions autonomously. I believe that, but there are four years in between—and what stands in between isn't algorithms, but piles of acceptance forms that no one is willing to fill out in advance.
Agents have made "doing things" cheaper, but "proving that it was done right" has instead become a new cost center. Only by first getting a clear accounting of the latter can you truly dare to hand the work over.
Sources: SAP News Center and ERP LEADERS (2026-11-05; Gartner’s “Emerging Market Quadrant for Integrated Supply Chain AI Agents” report dated 2026-10-28), BusinessWire / Morningstar (DualEntry AaaS, Epicor IDC MarketScape, 2026-11-05), Auboros “AI Hallucination Guardrails for ERP Agents” (verified 2026-11-03), Vauxoo / Odiware / Wavect Odoo 20 AI implementation analysis (2026-11-05), Gartner 2026 supply chain technology trends (as cited by Supply Chain Management Review), and public analyses related to Yonyou’s 2026 interim report (2026-11-01).
This article was compiled and written by Guangzhou Zhongdi Information Technology for reference by enterprise digitalization and finance leaders, and does not constitute procurement advice. The accompanying images were generated by AI.
