AI has arrived, and the anchor point of data assets has changed.

On January 1, 2024, the Provisional Provisions on Accounting Treatment Related to Enterprise Data Resources came into effect, officially opening the gate for data assets to be recorded on financial statements. Back then, almost every enterprise I contacted was asking the same question: can our data be recorded on the balance sheet, and how many hundred million yuan can it be valued at?

Two years on, the tide has shifted.

What is happening on the other side is that Microsoft pays Reddit 60 million USD annually only for the license to use forum posts for AI training; News Corp has signed a content licensing agreement with OpenAI reportedly worth more than 50 million USD per year. Note that what these payments buy is not a copy of the database, but the training right — the qualification to feed data to the model.

The same batch of data has two pricing systems. One is the valuation in the appraisal report, the other is the actual procurement order paid in hard cash by large model companies. The gap between them is not caused by appraisal methods, but by the fact that the anchor has changed: the value of data assets is shifting from "can it be recorded on the balance sheet and can it be transacted" to "can it fully feed the model".

In this article, I will clarify five new trends of data assets under the AI context, and finally provide you with a checklist that can be directly copied into your annual work plan.

Trend 1: The pricing power has moved — from the exchange to GPU server rooms

In the past, when we talked about data asset pricing, there were three reference points: the listing price on data exchanges, the intangible asset appraisal standards of appraisal institutions, and the accounting caliber for financial statement entry. The underlying assumption of this system is that data, just like land and real estate, has a "fair value" waiting to be discovered.

AI has overturned this assumption.

Now the most generous data buyers in the world are not brokers in the exchanges, but large model companies. Their logic for buying data is completely different: they do not care how many TB of data you have, but how much it can make the model stronger. The same set of customer service conversation records may only be depreciated to hundreds of thousands of yuan according to the "cost method" in the appraisal report, but for a company that wants to build an intelligent customer service Agent, it is 100,000 response samples with real human feedback — priceless raw materials.

The mechanism behind this is: AI has changed the marginal value curve of data. In the past, the consumers of data were humans, and there was an upper limit to one's reading volume; now the consumers of data are models, and their appetite for training corpora is measured in trillions of tokens. The entire demand curve shifts to the right, and the price of the scarce part of data — high-quality, labeled, semantically clear — naturally rises steadily.

The inference we can draw from this is: how much your data is worth depends less and less on how beautiful your appraisal report is written, but on "how many models need it".

Golden quote: Recording on the balance sheet is bookkeeping, and fully feeding the model is the real pricing.

Trend 2: The "first reader" of data assets has changed from humans to machines

Let's ask a question: who are the meticulously built indicator system, dashboard, and data catalog designed for?

The answer was very clear in the past: for humans. For business teams, for management, for auditors. Therefore, many designs of data governance essentially cater to human reading habits — visualization needs to be beautiful, calibers need to be reasonably explained, and reports need to be layered.

After the emergence of Agents, a new type of reader has been added. Moreover, the reading volume of this type of reader will soon exceed that of humans: the enterprise's procurement Agent directly reads the material master data to place orders, the financial Agent directly reads the voucher fields to reconcile accounts, and the operation Agent directly reads the interface to fetch data. They do not look at your exquisitely polished dashboard.

The preferences of machine readers are completely different from those of human readers:

• Humans can fill in ambiguities through tacit understanding, while machines will execute ambiguities to the end;

• Humans know that "bearing" and "bearing (old code)" refer to the same thing, but machines will treat them as two different materials, and then make the procurement place an extra order;

• Humans can tolerate slightly arbitrary field naming, but every join operation of the machine is a bet that the semantics of your field is unique.

Therefore, a new standard has emerged for the availability of data assets: machine readability. Whether the data structure is unambiguous, whether the metadata is complete enough to support automatic invocation, and whether interfaces are more "first-class citizens" than reports — these capabilities that were previously considered "nice-to-have" are now part of the composition of assets.

I judge that in the next two years, a number of enterprises will start to build "data catalogs for Agents": not web pages for human browsing, but API catalogs that can be directly retrieved and consumed by machines. The one who takes the lead will make their Agents run first.

Golden quote: Data designed for machines may not look fancy, but machines pay for it.

Trend 3: The acceptance criteria have changed — from "compliant and auditable" to "AI-usable"

In the past, the acceptance of a data governance project was judged by several metrics: standard coverage rate, data accuracy rate, standard implementation rate, and completeness of audit traces. These metrics answer the question: Is this set of data compliant? Can it withstand audit?

In the AI era, we need to add one more question: Dare we feed this set of data directly to the model?

The answers to these two questions are often inconsistent. One client told me that their master data "has an accuracy rate of 98% and passed the audit", but the effect was terrible after being used for RAG — because the same material had three different calibers in three systems, which people used relying on tacit understanding, while the model has no such tacit understanding.

I wrote an article about data standardization before, and there is one sentence in it that I want to repeat here: When humans encounter ambiguity, they will stop and ask, but when an Agent encounters ambiguity, it will guess an answer and keep running. In the past, conflicting calibers would cause misaligned reports, which could be resolved by aligning them in a meeting; now, conflicting calibers will make Agents directly execute wrong decisions, and the execution is large-scale, 24 hours non-stop.

As a result, some new metrics have begun to emerge in the industry. Although there is no unified name yet, the direction is clear:

• Corpus conversion rate: the proportion of structured data that can enter training or retrieval augmentation without loss, without secondary manual cleaning;

• Semantic density: for the same table, whether the annotations and business meanings are complete enough for the machine to understand independently;

• Ambiguity inventory: how many calibers in the business are "only understood by senior employees through tacit understanding" — this is the accident source list in the Agent era.

The nature of governance debt has also changed. In the past, governance debt only affected the appearance of reports and could be delayed; today, governance debt directly determines the lower limit of model output — the model will never be more reasonable than the data you feed it.

Golden quote: Past governance determines how good your reports look, today's governance determines whether your AI talks nonsense.

Trend 4: Three new accounts have emerged for right confirmation and valuation

AI has raised three new questions for data assets that did not exist before.

The first account: external licensing account. When your data is crawled by crawlers for training, should you charge for it, and who should you charge? The practice in the international market is clearly priced — Reddit, News Corp and other companies have already received their checks. China's data intellectual property registration is also advancing. For most enterprises, at least they need to start inventorying: which of my data has the value of external licensing? Are there clauses restricting the use for AI training in the licensing contract?

The second account: reflux data account. After the model is launched, its outputs, user corrections, and manual reviews will precipitate into new data. This batch of data is usually of high quality — it comes back with business feedback. The question is: whose asset is this? Will it be included in the ledger? Who is responsible for governing it? Many enterprises conduct data asset inventory once a year, but in AI scenarios, data proliferates once every time it turns in the flywheel, and static ledgers cannot manage it at all.

The third account: synthetic data account. Is synthetic data generated by models counted as assets? How to confirm its reliability? It has already supported half of the training corpus in some scenarios, but most enterprises' asset management systems do not even have the subject of "synthetic data".

Behind these three accounts is the same mechanism change: The boundary of data assets has become dynamic for the first time. In the past, we defaulted that assets are stock, and inventory is a snapshot; data assets in the AI era are flow and feedback closed loops, and snapshot-style management will miss the fastest growing part.

Golden quote: Static ledgers cannot manage dynamic flywheels.

Trend 5: The truly valuable form has emerged — the data flywheel, not the data pool

Tesla's shadow mode has been talked about many times, but from the perspective of data assets, it is worth reviewing again: millions of cars are running on the road, automatically uploading driving data, the data trains the model, the stronger model helps sell more cars, and more cars generate more data. Note that there is no link called "data asset appraisal" in this closed loop — but no one doubts the value of Tesla's data assets.

In contrast, many enterprises have stored PB-level data lakes that have not been flowed for ten years. There is no labeling, no business semantics, and conflicting calibers. When AI comes and they want to do something, they find that all the data they dig out is silt.

Where is the difference? The data lake is stagnant water, the flywheel is running water. Data that stays still will only depreciate — the business is changing, the calibers are drifting, and yesterday's data cannot match today's problems; data flowing in the closed loop completes a cycle of "generation → use → feedback → correction" every time it turns, and its quality is forced out by use.

Therefore, when evaluating an enterprise's data assets in the AI era, I have a new three-factor formula:

Data Asset Value = Data Stock × Flow Speed × Number of Feedback Closed Loops

Stock is only the base, and the multiplier is what widens the gap. For two enterprises with similar data stock, one's data is lying in the lake, while the other's data is consumed by models and Agents every week, with feedback reflux after consumption. Five years later, the former's data assets are close to depreciation, while the latter's are compounding.

Golden quote: Data pools are sold by the ton, flywheels compound by the loop.

Action Checklist for Peers in Data Governance

The above are judgments. Judgments cannot feed you, and the following are seven items that can be directly copied into next year's plan, sorted by priority:

1. Redo the data asset inventory, add a column of "AI Usability Rating". Score each core data asset item by item: degree of structurization, semantic clarity, ambiguity inventory, licensing status. If two of the four items are not up to standard, they cannot enter the candidate pool of "feeding the model".

2. Push back governance investment from high-value scenarios. First ask which three business scenarios will be launched with AI next year, then decide which data domains are worthy of focused governance. Give up the fantasy of full-volume governance — in the AI era, governance resources should follow the model, not the list.

3. Write corpus licensing into contracts and systems. Externally: clarify the boundary of training use in the contract with data providers; internally: write a whitelist of which data is allowed to enter model training and RAG, do not make up for it after an accident happens.

4. Build a machine-readable version of the data catalog. Prioritize the catalog for Agents over the catalog for humans: prioritize APIs over reports, prioritize field-level semantics over page aesthetics. The judgment standard is very simple — can your Agent find and correctly use this data independently without manual explanation?

5. Launch the special project of "Ambiguity Inventory". Let the senior employees in each business domain dictate "which calibers are only understood through tacit understanding": the meaning of old codes, historically inherited naming, and oral agreement priorities. This inventory used to be tacit knowledge passed on by word of mouth, but now it is the accident source for Agents, which must be made explicit.

6. Establish the ownership and posting rules for reflux data. Clarify the responsible department, quality standard and posting rhythm for model outputs, user corrections and manual labeling. Even if the system is a little rough, it is more important to get it running first than to wait for a perfect system.

7. Change the reporting wording. When reporting data asset work to your boss, replace "how many hundred million yuan have been recorded on the balance sheet" with "how many scenario models have been fully fed, how many person-days have been saved for the business, and how many decisions have been executed by Agents every month". The former means that assets are generating cash flow, while the latter only means that assets are lying on the report.

Conclusion

The concept of data assets has been popular for almost ten years. For ten years, it has been missing a hard anchor — recording on the balance sheet wanted to be this anchor, but it turned out that after bookkeeping, the market did not recognize it.

AI has filled this anchor. For the first time, the value of data can be measured by "how many models it makes smarter, and how much more data these models generate". This is not an advancement of appraisal technology, but a change of value logic.

For data governance practitioners, this is a once-in-a-decade opportunity: in the past, we were the people who kept accounts for data, and from now on, we are the people who supply ammunition for AI.

When the anchor changes, the ruler should be changed. Those who change early can make their story convincing at the budget meeting next year.

This article is from the WeChat Official Account "Data Drives Intelligence" (ID: Data_0101), written by Wang Jianfeng, and published with authorization from 36Kr.