How we count AI adoption
We read company reports for evidence that a business is using AI, either in its own work or in a product used by customers. Working trials count too. Plans alone stay out.
1,464 ÷ 2,727 = 53.7%. Companies reporting use or trials, divided by companies with a recent report. Each company counts once.
Which companies do we cover?
The list contains 2,826 companies. We cover today’s top 2,000 mainland-listed companies by market cap, plus 826 Hong Kong-listed companies with mainland operations. A company listed in both places counts once. Separately listed parents and subsidiaries count separately.
The headline percentage uses the 2,727 companies with a readable annual, interim or quarterly report for fiscal 2025 or later. That is 1,904 mainland listings and 823 Hong Kong additions. Companies without an eligible report are left out of the calculation.
Hong Kong selection and report availability
We began with HKEX’s mainland-enterprise category and selected 1,000 candidates by reported Hong Kong market value. Of these, 172 were already in the mainland list. We checked the other 828 against company filings: 826 had a mainland operating headquarters or predominantly mainland operations. Export-oriented mainland businesses can qualify. Ownership, incorporation or a mainland subsidiary address alone is insufficient.
We excluded Boyaa, which divested its mainland game business, and China Qinfa, which disposed of its mainland mines and continues its coal business in Indonesia. We did not replace them. This is a selected group, not every Hong Kong-listed business with mainland operations.
Current reports must contain at least 2,000 extracted characters. The 3 Hong Kong companies with only prospectuses or older reports remain searchable but are outside the percentage. The mainland total includes 1902 FY2025 annual reports and 2 companies with eligible interim reports. The historical annual-report series uses a lower text threshold of 500 characters.
Market caps are summed for the companies in each sector from listing snapshots. Mainland values are in RMB; Hong Kong values are in HKD and cover the Hong Kong listing, rather than the combined value of A and H shares. We show the currencies separately without an exchange-rate conversion. Some sector lists overlap, so their market caps should not be added.
What do we mean by AI?
We are interested in generative AI, including language, image, video and audio applications. Reports often leave the model unnamed, so we use a broad definition. For company AI or machine-learning projects reported in FY2025 onward, we include them as generative AI unless the report identifies a conventional method for that system.
Meeting the technology definition does not establish that the company is using it. We check that separately.
The full inclusion rules and exceptions
- Named generative models and clearly described generative functions count.
- Agents, AI tools, multimodal AI, large models, AI toys and glasses count under our definition. These can qualify even when used for prediction or recognition.
- AI assistants, secretaries and butlers, and AI audio applications also count. Natural-language and AI-vision applications described in 2025 or later are included as generative AI. A report’s year does not prove when the project began.
- We treat any 生成 (“generate”) wording in a classified passage as grounds for inclusion. This deliberately broad rule can include ordinary software output generation or automation. It does not prove use of a generative model.
- 智能 (“smart” or “intelligent”) prompts a review of the surrounding paragraph; the word alone is insufficient. 智能体 and 智能體 mean agents and follow the agent rule.
- A specifically identified conventional method on the same system normally stays out. The generation-word rule and recorded editorial decisions can override that exclusion. Ordinary digitisation without AI evidence does not automatically count.
These choices mean some conventional systems may be included. The number is a broad reading of company reports, rather than a verified count of LLM installations. We do not borrow a model name or evidence of use from an unrelated project on the same page.
When does it count as use?
A report must describe a system doing real work or undergoing a working trial. Internal applications and products used by customers both count. Selling chips, servers or computing capacity for AI does not, by itself, show that the company uses an AI application.
| Report Description | Included in the Use Count? |
|---|---|
| A system is being used, or is in a trial | Yes |
| A project is announced or planned | No |
| AI training, awards, product features or market trends | Not enough by themselves |
For example, “seven AI pilot scenarios went live” counts as a trial. A future AI project can fit our technology definition while remaining a plan. A supplier saying AI demand increased sales does not establish its own use.
Each company counts once if any collected report supplies qualifying evidence. The headline therefore covers use reported across different periods, not activity measured on one date. We have checked the source passages; we have not independently tested the systems or verified their results.
Status and user labels
“In use” means operational work. “In use at scale” also has reported figures for usage or results. “Trial” means a working test. “In development,” “Announced” and “Planned” remain separate. General commentary is excluded. Completed research, product descriptions and awards alone do not establish use.
“Inside the company” covers its employees or processes. “For customers” covers delivered AI capabilities used by customers. “Company and customers” requires both for the same system. Vendor research is not automatically internal adoption; use by a parent group or customer is not automatically use by the issuer. A plan can have an intended user without being in use.
One combined use count
All 1,464 companies with generative AI use or trials identified under these rules count together. We apply the same use requirement whether a report names a model, describes an application or qualifies through the broader inclusion rules.
The remaining companies have generative-AI projects in development or planned (332), other company AI projects underway (60), or no qualifying project found (871). Each company goes into its highest qualifying group, in that order. Other AI projects must be in use, in trials or being developed; future ambitions alone do not count. A company may use AI without disclosing it.
Why does the historical chart differ from the headline?
Both include mainland and Hong Kong companies. The headline combines evidence across collected filings; the historical chart uses each fiscal year’s annual reports only.
We follow the same 2,000 mainland companies and 826 verified Hong Kong additions backward, rather than selecting the largest companies afresh each year. Each percentage divides companies reporting use or trials by companies with an available annual report for that year. Missing and pre-listing reports stay unavailable.
Our AI-project, language and vision assumptions are broader from FY2025 onward. The increase therefore reflects changes in what we count and which reports are available, as well as companies reporting more AI activity. It cannot be read as a clean measure of adoption growth. Selecting today’s largest companies also leaves out businesses that have since shrunk or disappeared.
The company page shows two annual series using the same available reports. One counts company generative-AI projects, including development and concrete plans; the other requires use or trials. General AI discussion, market commentary and supplier demand alone are excluded from both.
How should we read the Ramp comparison?
Ramp counts US business customers with a recorded payment for an AI product or service in a month. We count Chinese listed companies reporting use or trials across collected filings. Our 2,727-company sample is significantly smaller than Ramp’s overall panel of more than 85,000 businesses. Neither represents every business in its country.
The table offers sector context. It cannot establish a national adoption gap because payments and reported use capture different things, over different periods, in different company populations. Our reports do not provide a consistent measure of AI spending.
Technology coverage and sector matching
Ramp covers language models, agents, AI search, image and video generators. Its spending categories also include GPU cloud and model serving. It may miss free tools, employee-paid accounts and internally hosted AI. Our measure includes internal and customer-facing applications and can include conventional systems under the stated assumptions. Technology coverage is therefore approximate. Selling an AI product alone does not qualify a business for Ramp’s payment measure.
Ramp uses NAICS sectors. Food and beverage production and vehicle manufacturing belong to manufacturing; farming and transport services do not. We remove mainland farms, distributors, services and uncertain mixed mining/processing businesses from our producer groups. Hong Kong companies are mapped from their principal business in company filings.
The six matched rows remain imperfect: Information includes some hardware, and Construction includes some building-material producers. Retail, healthcare and utilities are omitted because our groups are too mixed. Some companies appear in more than one row; they count once within each row. Ramp does not publish sector sample sizes in its public table.
Ramp’s method · Covered AI categories (§3.1) · NAICS definitions · Exact sector mappings
What do the reporting scores measure?
These describe how much detail companies give about AI and how their reporting changed between FY2024 and FY2025. They include plans. They do not measure the share using AI, system quality or financial returns.
Scores retain the original company groups; current use percentages include the Hong Kong additions. Broader classification rules in 2025 also affect the change score. Some companies belong to more than one sector, so sector totals should not be added.
Formulas and comparable figures
The stored names are “state” for reporting detail and “velocity” for reporting change. Company generative-AI project passages at all stages enter the scores; market commentary and supplier exposure alone do not. Metrics must appear in the project quotation. Coverage and density use metric names from FY2022–25. Repeated names alone do not establish comparable measurements. Other / Conglomerates and the Hong Kong group are unscored.
growth = (claims_fy25 + 5) / (claims_fy24 + 5) state = min(10, round(4*coverage + min(3, recurring_series/max(members,1)*2) + min(3,density_median/4),1)) velocity = min(10, round(min(4,max(0,growth-1)*1.3) + 3*rising_share + 3*new_entrant_share,1)) composite = round((2*state + velocity)/3,1)
Rising share uses companies with statements in both FY2024 and FY2025. New entrants first report an annual generative-AI statement in FY2025. Density is the median metric-name count in company-years with evidence; absent years are not assigned zero. Python rounds each component to one decimal, with ties to even, before calculating the composite. Download the exact inputs and company lists.
Charts of reported figures require the same company, named measure and system, compatible units and consistent cumulative status. Conflicting same-year values, intervals, compound changes and extreme jumps are excluded. Approximate values and thresholds remain approximate. Bank and insurance charts have separately reviewed measures. A nearby figure is not automatically attributable to the AI system in a quote.
How much does the definition change the number?
Our headline uses the broad rules above. Requiring explicit generative-AI wording, a large model or a named model gives a smaller count. Both calculations still require use or a working trial, and use the same company list.
Broad definition: 1,464 companies (53.7%). This is the headline.
Explicit generative or model references: 731 companies (26.8%). This narrower reading leaves out 733 companies supported only by the broader rules.
The narrower count can miss real generative AI when a company does not name the technology. The broader count can include conventional systems. These are two readings of the disclosures, not a confidence interval.
What we checked for accuracy
In November 2026, we re-reviewed the 190 companies added in the previous release, checked source passages flagged by the smart-word and activity rules, and selected fixed random samples of 100 previously counted companies and 100 excluded companies. For the excluded sample, we also searched available report pages for missed projects.
In the counted sample, 9 of 100 no longer qualified for the headline after review. In the excluded sample, 2 of 100 gained qualifying evidence. These are disagreements found by the audit, not estimates of the tracker’s accuracy.
The checks use AI-assisted source review, deterministic evidence checks and analyst spot checks. They are not independent human validation. The saved audit lists the sample, source coverage, changed companies and reasons.
Rui’s earlier editorial examples are preserved as technology-scope regression cases. A reporting year or an earlier label cannot turn “smart” into AI. Use, trial and company ownership still need separate evidence.
Download the accuracy audit and definition sensitivity · Saved editorial examples and regression cases · Analyst source corrections
Sources, checks and downloadable data
We use AI to extract and classify passages, check quotations against the original reports, and review proposed additions against the full source page. Editorial decisions can correct those classifications. These checks reduce errors; they do not make the reports complete or independently verify company claims.
One company can have many passages about the same project. Hong Kong reports publish selected short excerpts, prioritizing use and trials; mainland entries are more extensive. Statement counts therefore cannot rank AI usage across the two markets. Unlisted institutions and companies outside the headline list remain searchable but do not enter the percentage.
Versioning, source checks and audit trail
The public explorer shows company-project passages. The broader historical source archive contains 32,440 classified statements and is retained for audit, including entries excluded from the project counts. The source archive retains 101,300 quote-checked extraction records, including repeated versions. Public files call statements “claims.” A record is grouped by document, PDF page, normalized quote and business function. Among completed, quote-checked runs without a sweep flag, the latest run wins across all technology labels. Ties use run ID and mention ID. A later exclusion or status correction replaces an older positive label.
Recorded editorial decisions take precedence and are checked against the source page. Original text and earlier classification fields are retained for audit. These fields do not create separate public categories of generative AI. Hong Kong candidates receive a separate full-page check of use and company role. Altered or unverifiable quotations cannot establish use; recovered evidence must match the source and receive a new check.
The release manifest records source and output hashes. Counts, exports, company groups and scores are generated together without new model calls or database changes. The November 2026 review fixed PDF line-break and keyword-case failures, expanded the assumptions, and recovered missed passages. Before-and-after counts measure these corrections, not new market adoption during the review.
Annual and interim reports are company sources. Institutional supplements may use official publications, regulators or labelled media reports. Six older mainland company keys outside the top 2,000 are matched to the same Hong Kong issuers, rather than counted as additional companies.
- Company-project passages (CSV) · Historical source archive (CSV) · Search the reports
- Project categories and review checks · Project evidence and report links · Classification rules and counts · Earlier scope-review archive
- Reviewed sector corrections and sources
- Annual counts · Annual report history and sources · Sector company lists
- Hong Kong company checks · Report inventory · Hong Kong classifications and sources
- Editorial decisions · Replaced classifications · Release manifest