Last month I reconciled accounts with a client in industrial vision. Their AI service budget for this year is 800,000, and by the end of October they had spent 610,000. Finance asked one question: What did this 610,000 buy, what is the unit price, and how much quantity remains?
Procurement pulled out the contract, which stated: "Annual agent service package, including model invocation and operations and maintenance support."
There is no unit of measurement, no unit price, no usage cap. What was bought for 800,000 is a single sentence.
This is not an isolated case. Over the past two years, when companies bought AI services, most were priced by annual package, by seat, or by project, because vendors could not quote by usage either—each provider has a different basis for token pricing, and prices change as soon as the model changes. All companies can do is control the total amount; they cannot control unit cost.
This document from Guangdong dated November 11 affects this layer.
Guangdong has turned tokens into something tradable
On the afternoon of November 11, at a press conference in Guangzhou, the Guangdong Provincial Government Services and Data Management Bureau, together with 11 departments, officially released the Guangdong Province Action Plan for Promoting High-Quality Development of the Token Economy (2026–2027), featuring six major actions and 20 key tasks: building a foundation with data elements, cultivating token factories, developing the circulation system, incubating token scenarios, discovering star token enterprises, and optimizing the token environment.
Launched on the same day was the “Wanshu · Yuanqi” platform, backed by the Guangdong Token Trading and Service Center, which was led and established by the Guangzhou Data Exchange. Among the five capabilities officially provided, two are directly related to corporate ledgers:
First, trusted metering—standardized accounting is carried out using tokens as the unified unit, with tamper-proof data evidence throughout the entire process.
Second, enterprise-benefiting policies — uniformly consolidate AI support policies at all levels, and rely on trusted transaction credentials to achieve intelligent policy matching and precise fulfillment.
Putting these two sentences together, the meaning is clear: in the future, when enterprises obtain AI-related policy funding, the supporting documents will no longer be just contracts and invoices, but will also include a verifiable usage record.
The policy foundation did not first emerge on this day. In June 2026, the Provincial Government Services and Data Administration had already issued the "Several Measures for Guangdong Province to Unleash the Value of Data Elements and Promote the High-Quality Development of the Token Economy," which was called China's first provincial-level special policy for the token economy, with the goal of making Guangdong a national highland for token supply, a benchmark for intelligent applications, and a hub for token going global by the end of 2027. The event on November 11 was about translating the system into platforms and transactions—among the five categories of cooperation signed on site, the first was token procurement orders.
When the unit of measurement changes, contracts and invoices must change accordingly
Getting back to that 610,000.
If the contract states "X yuan per thousand tokens, an annual planned usage of Y thousand tokens, and monthly settlement," finance should at least be able to answer three questions: what the unit cost is, how much has been used, and how much remains. The fact that they cannot answer now is not because finance is incapable; it is because there is nothing calculable in the contract.
When we revise this type of contract for clients, we always add four lines:
- Unit of measurement: tokens or thousand tokens; clearly specify whether pricing is separate for input and output or a flat rate.
- Unit price: tax-inclusive/tax-exclusive, whether it is adjusted with model versions, advance notice period for adjustment
- Settlement cycle: monthly or quarterly, and from which date does the payment term start counting?
- Reconciliation method: whose measurement data prevails; what discrepancy threshold triggers review
The third and fourth lines are the easiest to miss, and also the easiest to argue over. It is normal for the usage shown in the supplier's backend to disagree with the enterprise's own records; a difference within 5% is usually accepted, but a difference of 20% means you have to go through the logs. Writing in the contract upfront that “the supplier's backend shall prevail, and the enterprise has the right to inspect the original records” is less trouble than arguing afterward.
The steps on the system side are not complicated, but they must be done before procurement, not supplemented afterward.
In Odoo, create a service-type product for the AI service, set the unit of measure to "thousand tokens", and enter the planned usage and unit price on the purchase order line. Each month the supplier sends a usage statement, which is attached to the vendor bill as an attachment, and the bill lines are analytically distributed by project. This way, at month-end you can produce two reports: one showing token consumption by project, and one showing unit cost changes by supplier.
There is one detail worth mentioning. One of our customers has an agent project that serves three downstream customers at the same time and shares a single set of model calls. Previously, all costs were recorded under a single account, "R&D Expenses—Cloud Services," and year-end allocation relied on manual estimates. After switching to splitting by analytic account, the bill for the same month is allocated to the three projects based on call proportions, and only then can each project's gross margin be calculated accurately. This has nothing to do with policy, but without a contract that measures usage, it cannot be done.
Foreign trade enterprises have one more layer. "Token Going Global" is one of the platform's five capabilities, offering differentiated service packages to small and medium-sized enterprises, Agent platform providers, and large enterprises going global. For those doing cross-border business, if they purchase invocation services from overseas Agent platforms, this involves currencies, exchange rates, and cross-border payment vouchers, and the accounting treatment is more complex than for domestic business. There are currently no detailed rules in this area; my suggestion is to first follow the existing standards for foreign exchange and cross-border trade in services, and not invent accounting subjects yourself.
We must be honest here: as of November 13, we have not obtained the specific format and issuance rules for the “trusted transaction certificate.” The platform has just launched, and the fulfillment of enterprise-benefit policies has only completed the signing of the “first batch of cooperation intentions”; no enterprise has actually received money yet. So what we can do now is keep good records of usage, not promise how much can be redeemed.
What the list of 211 companies solves is not a money problem
Another matter was also announced the same day: Guangdong Province's first batch of digital economy innovative enterprises included in the list—211 in total, 83 seed-type, 82 potential gazelle-type, and 46 potential unicorn-type.
The basis is two documents—Several Measures on Strengthening the Cultivation of Innovative Digital Economy Enterprises (Fa Gai Shu Ju [2025] No. 1154) and the Work Plan for Cultivating and Upgrading Innovative Digital Economy Enterprises in Guangdong Province (Yue Zheng Shu Han [2026] No. 339). The latter proposes the “Star Exploration Program,” which is advanced through a dynamically managed three-tier gradient cultivation database, a set of tiered evaluation standards, a service system, and a set of operational monitoring mechanisms.
What enterprises included in the pool receive is policy element matching, financing channels, scenario opening, professional coaching, and support for going public—not direct funding. This point must be made clear, otherwise some enterprises will again assume that being on the list means there is money.
For service providers, the value of this list lies at the other end: if a client is on it, then in all subsequent application materials, operating revenue, R&D expenses, and the proportion of data business revenue will be checked by reviewers as established facts. The figures filled in at the time of application and the figures now produced in the system must match. During due diligence, we have seen cases where they did not match; the reason was that at application the figures were filled in on a "consolidated basis," while the system contained only single-company data, resulting in a discrepancy of more than 30%, which is very troublesome to explain.
So when helping clients calculate these kinds of metrics, first settle three things: which legal entity, which period, and whether tax is included. If these three are not fixed, the numbers will definitely not match.
Three things you can do this week
First, go through all AI service contracts for this year, looking at only four lines: billing unit, unit price, settlement cycle, and reconciliation method. If two or more of the four lines are missing, add them when renewing; if the amount exceeds 500,000, sign a supplementary agreement this year. The judgment criterion is simple—whether finance can calculate from the contract “how much per thousand tokens”; if it cannot, then it is missing.
Second, create an AI service product and corresponding unit of measure in the system, and start recording by usage from the December bill. If historical contracts cannot be backfilled, make opening entries and clearly state “measured by usage from month X.” This costs no money, just two hours.
Third, if the client or you yourself are preparing applications for things like digital innovation enterprises or specialized, refined, distinctive, and innovative enterprises, first pull the three figures—operating revenue, R&D expenses, and the proportion of data business revenue—from the system once, and check them against last year's application form. For any discrepancies, identify the reasons before the end of November; don't wait until the December application window opens to start digging back through the accounts. The tiered cultivation measures in Guangdong Industry and Information Technology Regulatory Document [2026] No. 5 will take effect on December 1, and the criteria will only become more detailed.
