Last week a client asked me to look at their IT budget spreadsheet. The spreadsheet was very detailed—servers, bandwidth, software subscriptions, outsourced labor, listed column by column down to two decimal places. I asked one question: What about the AI part.
He paused for a moment and said AI isn't under my purview. Sales is using it on their own, supply chain has its own setup, HR is also trying it, and the money comes from their respective budgets. On my side, I only manage the underlying platform.
This answer has become increasingly common this year. The way AI spends money is different from the procurement logic of enterprise software over the past two decades: it is not like ERP, with a project initiation, a single tender, and one contract; rather, each business line tries it out, buys it, and renews it on its own. The advantage of spending in a scattered way is speed; the disadvantage is that no one knows the total amount, and therefore no one knows whether it is worth it.
1. Those two IDC predictions are about the same thing
On November 18, IDC held the FutureScape 2027 China ICT Market Forecast Forum in Beijing, where it released two predictions that appear to point in opposite directions.
One is: by 2027, 60% of AI funding will fall outside a single IT budget. Another is: by 2027, CFOs and CIOs at China's top 2,000 enterprises will adjust 40% of new AI investments, shifting from "securing investment based solely on pilot results" to "projects that define cost baselines and expected business outcomes before launch".
One says money is being disbursed in scattered ways; the other says the threshold is tightening. Seen together, they are actually two sides of the same thing: money can be spent in a dispersed manner, but it must be aggregatable, allocatable, and comparable. This is exactly the job of management accounting, not the job of the procurement department.
Why these two things happen at the same time isn't hard to deduce logically. In the pilot stage, what is spent is small money, buying a story; a story doesn't need to be recorded under any account category—as long as it's told movingly. By the rollout stage, what is spent is big money, and it must be renewed every year thereafter; at this point the question becomes: which department's budget will this money be deducted from, and which financial statement will the saved money go into. A story can go unaccounted for, but results must be accounted for.
In her opening remarks, Kitty Fok, President of IDC China, offered a summary: AI is not addition, but "rewrite." At the strategic level, this sounds like direction; at the financial level, it sounds like a reminder—rewrite means the original account structure may no longer be sufficient.
II. Two-thirds of companies saw net costs rise after AI went live
On the same day, TTEC released “CX Trends 2027.” There is one figure in the report worth pausing to look at: 66% of customer experience leaders report that net costs increased after adopting AI. Meanwhile, only 1% believe their operating model is “highly adaptable and built for continuous change.”
When this conclusion first came out, many people's first reaction was "AI is bragging again." But putting the circumstantial evidence together makes the picture clearer. Gartner surveyed 199 service and support leaders from April to May 2026 and found that AI spending in this field grew by 38%, while the total budget for the field grew by only 2%. The difference must be moved from somewhere else; according to Gartner analysts, it is "shifted from labor and overhead costs."
The problem is that the workforce hasn't really gone away. Another Gartner survey covering 321 service leaders shows: 85% are adding new tasks to the frontline, 75% are moving the frontline into entirely new roles, only 31% have implemented or plan to implement layoffs due to AI before the first quarter of 2027, and 63% choose to gradually absorb it through natural attrition.
So a rather awkward structure has appeared on the books: platform subscriptions are rising, integration projects are rising, data governance is rising, and the headcount for oversight and review has not come down—the original people are still there, only the work they do has changed. Four newly added sums of money, plus one sum that was never cut, are all pressing down on the same income statement.
The cost hasn't disappeared—it has just changed accounts.
This sentence is the key to understanding 66%. AI does reduce some work hours, but it simultaneously introduces four new types of costs:
- Platform and subscription fees—per seat, per call volume, per credits, prepaid monthly or annually
- Integration and customization costs—interfaces, permissions, data cleansing, mostly a one-time investment plus ongoing maintenance
- Data governance costs——Master data cleansing, knowledge base maintenance, and metric definition alignment, the hardest to get approved as a project yet the most expensive
- Supervision and review hours — AI-generated results need someone to review them, and the reviewers are still the same people as before
These four payments have one essential difference from labor costs: labor costs are incurred monthly, have payroll slips, and can be allocated by department; these four are paid annually, come according to bills, and have no corresponding documents. Finance cannot see them, not because it is not diligent, but because they are not in the voucher flow.
A specific accounting pitfall: Odoo's AI features are consumed through IAP prepaid credits. At the moment of top-up, it is a prepayment, not an expense. If recorded directly as a current-period expense, monthly gross margin will jump with the top-up cadence rather than change with actual usage. The correct approach is to first record it as a prepayment, then recognize it based on actual consumption—otherwise you can't even answer "how much did the agent spend this month."
III. After having an identity, the next question is how much it cost
IDC also had another prediction at the same forum: by 2028, 70% of production-grade Agents at China's Top 500 enterprises will have independent non-human identities, obtain temporary least privileges on a per-task basis, and fully record authorization, actions, and lifecycle.
The matter of "issuing employee IDs to agents" has been discussed before; here I will only discuss the accounting consequences it brings: once an agent has an identity, it meets the condition for becoming a cost object. Only with a cost object can unit costs be calculated, departmental allocation be done, and output be compared. Without this layer, all AI spending can only be squeezed into one big pool of "IT expenses"; at year-end a total is aggregated, and no one can say which scenario is the most worthwhile.
I once saw a very typical scene at a company. They spent over 800,000 a year on AI, and when the boss asked which scenario was most worth continuing to invest in, the technical lead couldn't answer. The reason was simple: all calls went through the same API key, the bill was a monthly total, and it wasn't split by scenario. They could say "how much was spent in total," but they couldn't say "how much the quoting assistant spent, how much customer service summaries spent, how much quality inspection reports spent."
This is not a technical problem. The logs actually contain call records—it's just that no one has matched them against costs. And the direct consequence of failing to match them is: when budgets are cut at year-end, the only option is a blanket cut, cutting away both what should be kept most and what should be cut most.
IV. In Odoo, when opening an account for an agent, which areas are affected?
If you want to implement this in an existing ERP, it is simpler than you might think, but there are more pitfalls than you might imagine. Taking Odoo 17 to 19 as an example (fields in version 20 may be slightly adjusted, so verify before implementation), it roughly comes down to these points:
| Things that are about to fall | Locations in Odoo | What happens if it's not done? |
|---|---|---|
| Cost object | Analytic account (account.analytic.account), open by scenario, not by vendor or model | All expenses are mixed into one pool, making it impossible to compare scenarios |
| Pre-recharge processing | IAP credits are recorded as prepaid and transferred to expenses based on actual consumption | Gross profit fluctuates with the recharge cadence, and costs are distorted |
| Allocation rules | Include it in the allocation based on usage volume or headcount, with a methodology note attached. | Allocation becomes arbitrary guesswork, and departments refuse to accept it |
| Budget Control | Analyze account budgets or Enterprise budgets, and set alert thresholds | Overspending is only discovered at year-end, too late to put on the brakes |
| Output comparison | Custom metric fields: document count, error rate, processing cycle | Only the claim "saves work hours" remains, and it cannot be verified. |
| Identity and Logs | Dedicated service accounts, recording rules, write operation audit trails | No revocation upon departure, no alert on privilege escalation, broken audit trail |
The rows most easily skipped in this table are the second and fifth rows.
First, let's talk about prepaid recharge
Odoo's IAP mechanism is to buy credits first and then consume them. Many companies, to save effort, expense the top-up directly in the month of purchase. This does not show a problem in months with stable business volume; however, once they top up a year's worth in a single month, that month's expenses suddenly swell, making gross margin look bad; the next month it is empty again, making it look as if efficiency has improved. When finance uses this table for analysis, the conclusions they draw are all noise.
The accounting treatment is not complicated: record top-ups as prepaid expenses; at month-end, transfer costs based on actual consumption, and leave the balance on the asset side. One extra journal entry buys you a credible number for "how much was spent on AI each month."
More on Output Comparison
The benefits of most AI projects are written as "how many work hours were saved." This metric sounds intuitive, but the problem is that it is hard to verify—did the saved time turn into more output, or into idle time? If the saved work hours do not correspond to an increase in output or document volume, then it does not exist on the report.
The more verifiable metrics fall into three categories: throughput (number of documents, work orders, and inquiries), error rate (rework rate, review return rate), and cycle time (duration from receipt to completion). All three have native record points in Odoo, requiring no additional development; they only need to be defined as acceptance indicators before go-live.
A recurring pitfall: Analysis accounts only carry revenue, not costs. Many companies open analysis accounts to see how much a certain business line earns, while costs remain in the general ledger and are allocated by department. In this case, even if you open an analysis account for an AI scenario, what is attached to it is only revenue or simply empty, leaving no way to allocate costs. When opening an account, you need to think clearly: which types of costs should this account bear, and from which account subject should they be transferred.
V. If You Only Do Three Things
1. Lay out the bills from the past three months and identify four types of costs.Platform subscriptions, API calls, integration labor, and oversight hours. See whether you can add them up to a single number. If you can't, it means your AI spending is currently in an "invisible" state; at this point discussing ROI is meaningless—you don't even have the denominator yet.
2. Open an analysis account for the single most expensive scenario, attach the four sums of money to it, and run it for three months.Do not open accounts for all scenarios at once; that will end up unfinished. Pick the one with the largest amount and the clearest boundaries—for example, quotation assistance, invoice matching, or customer service summaries. After three months, you will get a table showing how much this scenario actually cost each month. The questions this table can answer are more than any presentation.
III. Set a go-live rule: before any new agent goes live, it must clearly state both the "cost baseline" and the "verification metrics." The cost baseline is the estimated monthly spend and which department bears it; the verification metrics are the numbers to be used three months later to judge whether it has been useful. Both must be findable in the system, not merely written in a PPT. The "40% of new investment will be adjusted" that IDC mentions refers to adjusting projects that lack these two things.
Return to the IT lead from the beginning. Later he did something not particularly big: he had each business line cc a copy of AI-related bills to finance every month, and finance set up analysis accounts by scenario. The numbers aggregated in the first month surprised him somewhat—the total was 40% higher than his estimate, and nearly a third of it was three pilots that no one was using anymore still renewing their subscriptions.
Accounting for AI costs is not about saving money. It is about making it comparable. An investment that cannot be compared will never make it to the front in a budget meeting; nor can it ever be clear whether it should continue. And when 60% of the money has already been dispersed to business lines, while 40% of the investment must have a cost baseline before launch, the very question of 'whether it can be calculated clearly' itself becomes a threshold.
Data sources in this article: forecasts released by IDC China on 2026-11-18 at the Beijing FutureScape 2027 China ICT Market Forecast Forum (60% of AI funding exceeds a single IT budget, 40% of new AI investment is adjusted, by 2028 42% of leading enterprises will restructure agent and application assets, 40% of production-grade agents have self-evolution capabilities, 70% have independent non-human identities); TTEC “CX Trends 2027” released on 2026-11-18 (66% net cost increase, 1% highly adaptive operating models); two Gartner surveys in April–May 2026 (199 service and support leaders: AI spending +38% vs. total budget +2%; 321 respondents: 85% increased frontline tasks, 75% reassigned, 31% plan layoffs, 63% natural attrition). Odoo models and fields are exemplified by versions 17 to 19; version 20 may change, so please verify the official documentation before implementation.
Zhongdi Information · Observations on Enterprise Management Digitalization · The enterprise scenarios in this article are drawn from implementation interviews and have been anonymized.
