Domain AI Readiness

• Domain AI Readiness is defined as the degree to which an industrial domain is technologically integrated with AI, factoring in both external and firm-level capabilities.
• It employs a patent co-occurrence methodology using IPC4 codes to generate readiness scores that reveal the interaction between AI integration and domain evolution.
• The concept guides strategic decisions such as delaying AI investments until readiness thresholds are met, emphasizing that AI benefits rely on external technological integration.

Domain AI readiness denotes “the degree to which an industrial domain is technologically integrated with AI.” In this formulation, readiness is not a synonym for firm-level AI investment, national digital capacity, or dataset quality alone; it is an external technological condition of the domain itself. Zeng, Wang, and Sun operationalize the concept through patent co-occurrence between AI-related and domain-specific IPC4 fields, and show that firm-level AI capabilities produce larger productivity and innovation gains when deployed in domains with higher readiness, while benefits are limited in domains that are technologically unprepared or already obsolete (Zeng et al., 13 Aug 2025).

1. Definition, scope, and theoretical rationale

Domain AI readiness is defined as “the degree to which an industrial domain is technologically integrated with AI.” The underlying theoretical claim is one of complementarity: AI is treated as a General-Purpose Technology whose returns depend not only on internal investments such as data pipelines, talent, and management practices, but also on the external technological environment of the domain in which it is deployed (Zeng et al., 13 Aug 2025).

The construct is explicitly domain-specific. Unlike traditional IT, AI is described as relying heavily on context-specific data, metaknowledge, and iterative feedback loops, all of which vary by industry domain. The relevant complements are therefore not only internal firm assets, but also industry-level data ecosystems, shared standards, and collective knowledge stock. In this sense, readiness is an attribute of the surrounding technological landscape rather than merely a property of the adopting firm.

This framing also distinguishes domain AI readiness from broader readiness indices. Some frameworks assess readiness at the level of governments, institutions, datasets, or operational systems; domain AI readiness instead asks whether a given technical field has already become sufficiently entangled with AI that firms operating in that field can convert AI capability into output gains. A common misconception is that AI capability alone should improve firm performance wherever it is introduced. The reported results reject that view: AI capability by itself can be ineffective, and even harmful, in low-readiness domains (Zeng et al., 13 Aug 2025).

2. Patent-based operationalization

The empirical construct is built from patent data using an IPC4 co-occurrence methodology. For each year $t$, the procedure identifies a set $A_t$ of AI-related IPC4 codes, approximately $20$–$30$ codes. For every non-AI IPC4 code $j$, it then counts

$$f_{j,t} \equiv \text{number of patents in year } t \text{ whose IPC4 list contains } j \text{ and at least one code from } A_t.$$

To mitigate skew, the $f_{j,t}$ values are ranked across all $j$ into deciles. Let $d_{j,t}\in\{1,\dots,10\}$ denote the decile of $j$ in year $A_t$0. The field-level readiness score is then defined as

$A_t$1

so that $A_t$2, with $A_t$3 if $A_t$4. The interpretation is direct: $A_t$5 is the domain AI readiness of technical field $A_t$6 at date $A_t$7 (Zeng et al., 13 Aug 2025).

Firm-level domain readiness is an exposure-weighted aggregation over the firm’s patent portfolio. Let firm $A_t$8 hold $A_t$9 patents in IPC4 field $20$0 in year $20$1, and let $20$2. The firm-level measure is

$20$3

This construction assigns larger weight to fields in which the firm is more patent-active. It therefore measures not whether AI exists somewhere in the economy, but whether the firm’s own technological domains are embedded in fields that have become more integrated with AI.

The methodology is designed to capture external technological integration rather than internal strategic relabeling. That distinction becomes central in later decomposition tests, where the paper separates external “technological evolvement” from firms’ own “strategic pivot” behavior (Zeng et al., 13 Aug 2025).

3. Complementarity with firm-level AI capability and performance

The empirical argument proceeds in two stages. The first is a demand test of input complementarity:

$20$4

The dependent variable is measured by four indicators: log frequency of AI words in MD&A (“AI strategy”), share of AI-related job ads (“AI talent”), log AI-specific assets in footnotes (“AI assets”), and their standardized sum (“AI capability”). Controls are firm size, leverage, and Tobin’s $20$5. A positive $20$6 supports the claim that firms in high-readiness domains invest more in AI (Zeng et al., 13 Aug 2025).

The second stage is an output complementarity test:

$20$7

Here, $20$8 denotes productivity outcomes such as revenue per employee or TFP, and innovation output is tested through a Poisson version of the same specification using trademarks. A positive $20$9 implies that AI capability produces larger gains in high-readiness domains (Zeng et al., 13 Aug 2025).

The reported findings are strong and internally consistent. In the demand test, $30$0 and is significant at $30$1 across AI strategy, talent, assets, and the composite index. In the performance test, $30$2 and significant. Moving $30$3 from $30$4 increases labor productivity by approximately $30$5 of a standard deviation and TFP by approximately $30$6 of a standard deviation, holding AI capability constant. At the same time, $30$7 becomes negative once the interaction is included, implying that AI capability alone in low-readiness domains can hurt performance. Innovation output shows the same complementarity pattern (Zeng et al., 13 Aug 2025).

A further decomposition separates readiness changes attributable to external technological evolution from those attributable to firms’ own strategic pivots. The external component drives interaction effects approximately $30$8 larger and remains highly significant, whereas the internal strategic-pivot variation is insignificant. This directly addresses the concern that firms might mechanically boost readiness by shifting patenting into AI-adjacent fields without any real change in the domain’s external technological environment.

The results are reinforced by robustness and instrumental-variable exercises. Using the cumulative number of active AI-promoting policies in the firm’s province, lagged one year, as an instrument for AI capability, the first stage yields an $30$9-statistic above $j$0 and a positive coefficient on policy count at $j$1. In second-stage $j$2, the estimated $j$3 rises by approximately $j$4–$j$5 relative to OLS, which the authors interpret as evidence that OLS likely underestimates the true complementarity. Placebo regressions on firms with zero AI talent and zero AI assets show no significant direct effect of the instrument on productivity. The same qualitative pattern survives when readiness is measured by co-occurrence share rather than frequency, and when firms with any AI patents are excluded (Zeng et al., 13 Aug 2025).

4. Temporal dynamics and the sources of rising readiness

Domain AI readiness is presented as a dynamic rather than static construct. Time-series analysis of IPC4 co-occurrence patterns indicates substantial changes in how AI diffuses across technical fields. From $j$6, co-occurrence in IPC Section A (Human Necessities) grew approximately $j$7, Section G (Physics) approximately $j$8, and Section H (Electricity) approximately $j$9. The associated decile analysis shows that many mid-decile IPC4 codes migrated sharply upward, indicating a re-ranking of fields by AI integration rather than a simple deepening within already AI-dense areas (Zeng et al., 13 Aug 2025).

The paper attributes this rise primarily to external technological evolvement. Three forms of evidence are used. First, OpenAlex publication trends show that, between 2016 and 2022, “Medicine” and “Biology” AI papers grew fastest among non-AI topics. Second, in Poisson regressions of co-occurrence counts, IPC4 codes G16B (bioinformatics) and G16H (healthcare informatics) grew $$f_{j,t} \equiv \text{number of patents in year } t \text{ whose IPC4 list contains } j \text{ and at least one code from } A_t.$$0 faster than other G-section codes in AI patents, with a positive and significant interaction of Bio $$f_{j,t} \equiv \text{number of patents in year } t \text{ whose IPC4 list contains } j \text{ and at least one code from } A_t.$$1 Year trend. Third, enterprise AI patenting in a given IPC4 is positively predicted by lagged research-institution patenting in the same IPC4, while no similar effect appears for individual inventors (Zeng et al., 13 Aug 2025).

These patterns matter because they shift the interpretation of readiness away from firm voluntarism. The results imply that readiness is not mainly created by isolated corporate repositioning. It emerges from academic breakthroughs, shared data ecosystems, standards, and knowledge spillovers that alter the technological substrate of a domain. In that sense, domain AI readiness is historically cumulative and partially collective.

A related implication appears in other AI-readiness literatures. Scientific-data papers similarly describe readiness as something produced through infrastructure, provenance, and standardization rather than simply declared by downstream users. REDI, for example, defines data readiness for scientific AI as the degree to which a raw scientific dataset has been “domain-appropriately cleaned, validated, feature-engineered, and semantically enriched so that it can be directly consumed by large-scale AI workflows,” and operationalizes it through the five-stage IPTSO pipeline (Wilkinson et al., 2 Jul 2026). This parallel does not collapse the two concepts, but it underscores the same underlying logic: AI uptake depends on prior integration work within the relevant domain.

5. Relation to other readiness frameworks

The term “AI readiness” is used across several adjacent literatures, but the object being assessed differs substantially. Domain AI readiness evaluates the external technological integration of an industrial field. Other frameworks evaluate datasets, institutions, public administrations, or operational systems. This diversity is visible across scientific AI, healthcare, public-sector deployment, autonomy, and national benchmarking (Wilkinson et al., 2 Jul 2026, Legara et al., 17 May 2026, Dubey et al., 2024, Hobbs et al., 2022, Alalaq, 26 Mar 2025).

Construct	What is assessed	Core formulation
Domain AI readiness	Technological integration of an industrial domain with AI	IPC4 co-occurrence score $$f_{j,t} \equiv \text{number of patents in year } t \text{ whose IPC4 list contains } j \text{ and at least one code from } A_t.$$2 and firm-level aggregation $$f_{j,t} \equiv \text{number of patents in year } t \text{ whose IPC4 list contains } j \text{ and at least one code from } A_t.$$3 (Zeng et al., 13 Aug 2025)
Data readiness for scientific AI	Degree to which raw scientific data can be directly consumed by large-scale AI workflows	DRAI levels and IPTSO / ingest-to-shard pipelines (Wilkinson et al., 2 Jul 2026, Brewer et al., 30 Jul 2025)
Institutional Alignment Readiness	Whether the receiving institution can responsibly deploy an AI system	Institutional and operational compatibility, data ecosystem maturity, human oversight capacity, fiscal sustainability, regulatory alignment (Legara et al., 17 May 2026)
AI readiness in healthcare	Responsible and effective adoption by medical practitioners	Data, technical, interpretability, and organizational readiness (Dubey et al., 2024)

In scientific AI, readiness is often attached to data transformation pipelines. REDI introduces levels from RAW to FULLY_AI_READY and tracks completeness, consistency, and transformation quality across ingest, preprocess, transform, structure, and output (Wilkinson et al., 2 Jul 2026). “Data Readiness for Scientific AI at Scale” uses a two-dimensional framework with Data Readiness Levels and Data Processing Stages, moving from Raw to Fully AI-ready and from ingest to shard (Brewer et al., 30 Jul 2025). AIDRIN and its later extension evaluate data quality, bias, fairness, privacy, and FAIR compliance, and in AIDRIN 2.0 the metrics are reorganized under six pillars including Data Quality, Understandability/Usability (FAIR), Impact on AI, Governance, Structure, and Fairness (Hiniduma et al., 2024, Hiniduma et al., 22 May 2025). SciHorizon-DataEVA generalizes this into Governance Trustworthiness, Data Quality, AI Compatibility, and Scientific Adaptability (Liu et al., 29 Apr 2026).

In public deployment settings, the key limiting factor may be the receiving institution rather than the model or the domain. Institutional Alignment Readiness evaluates institutional and operational compatibility, data ecosystem maturity, human oversight capacity, fiscal sustainability, and regulatory alignment readiness, and is explicitly motivated by cases in which technically viable systems could not advance to broader rollout for institutional reasons (Legara et al., 17 May 2026). Healthcare AI readiness adds yet another axis by tying adoption to interpretability and organizational trust, with storytelling XAI proposed as a mechanism to address audience-specific explanation needs (Dubey et al., 2024).

Operational and mission settings treat readiness as a staged maturity problem. STARL separates autonomy readiness from trust readiness in space systems (Hobbs et al., 2022). TRL4ML defines TRL $$f_{j,t} \equiv \text{number of patents in year } t \text{ whose IPC4 list contains } j \text{ and at least one code from } A_t.$$4–$$f_{j,t} \equiv \text{number of patents in year } t \text{ whose IPC4 list contains } j \text{ and at least one code from } A_t.$$5 for AI/ML systems, with formal artifacts, reviews, and the sim-to-real risk formulation $$f_{j,t} \equiv \text{number of patents in year } t \text{ whose IPC4 list contains } j \text{ and at least one code from } A_t.$$6 (Lavin et al., 2020). Browne and Bailey’s AI Readiness Framework introduces five gates—Alignment, Justified Confidence, Governance, Data Readiness, and Human Readiness—across ARL $$f_{j,t} \equiv \text{number of patents in year } t \text{ whose IPC4 list contains } j \text{ and at least one code from } A_t.$$7–$$f_{j,t} \equiv \text{number of patents in year } t \text{ whose IPC4 list contains } j \text{ and at least one code from } A_t.$$8 (Browne et al., 15 Apr 2025).

At national and public-sector scales, readiness is expressed through composite indices. The Oxford Insights Government AI Readiness Index uses Government Environment, Technology Ecosystem, and Data & Infrastructure (Alalaq, 26 Mar 2025). A GCC-specific AI Adoption Index identifies Infrastructure & Resources, Organizational Readiness, and Policy & Regulatory Environment, with the combined model explaining $$f_{j,t} \equiv \text{number of patents in year } t \text{ whose IPC4 list contains } j \text{ and at least one code from } A_t.$$9 of the variance in AI outcomes (Albous et al., 5 Sep 2025). These frameworks are not substitutes for domain AI readiness, because they assess country-level or public-sector enabling conditions rather than domain-specific technological integration.

Taken together, this literature suggests that “readiness” is not a single scalar property. It is a family of level-specific constructs attached to different units of analysis: domain, dataset, institution, organization, nation, or operational system. Domain AI readiness occupies the meso-level position within that family.

6. Implications, misconceptions, and research directions

The immediate managerial implication of domain AI readiness is that AI adoption should be conditioned on domain preparedness rather than on internal enthusiasm alone. The paper explicitly recommends “Strategic Waiting”: delaying heavy AI investment until domain readiness reaches a critical threshold. It also recommends a “Domain Audit,” in which firms assess $f_{j,t}$0 across their core technologies via patent co-occurrence analysis. By contrast, “Shallow” pivots—simply filing patents in new fields to chase AI—do not yield productivity gains, consistent with the finding that internal strategic-pivot variation is insignificant (Zeng et al., 13 Aug 2025).

The policy implications are similarly domain-specific. Beyond general AI R&D grants, the paper recommends targeting data governance, interoperability standards, and talent pipelines in high-potential domains. It also recommends investment in shared data infrastructures, open scientific repositories, and regulatory frameworks that lower barriers for all firms in the domain, as well as using patent co-occurrence metrics to identify lagging sectors and tailor interventions accordingly (Zeng et al., 13 Aug 2025). These recommendations align with neighboring literatures in which FAIR compliance, provenance, machine-readable schemas, and governance infrastructure are treated as prerequisites for scalable AI use (Wilkinson et al., 2 Jul 2026, Caufield et al., 12 Sep 2025).

Several misconceptions are directly contradicted by the evidence. The first is that AI capability is uniformly valuable. In the reported estimates, its marginal effect becomes negative in low-readiness domains once the interaction term is included. The second is that readiness can be manufactured internally by strategic pivoting. The decomposition results attribute the stronger effects to external technological evolvement instead. The third is that readiness is exhausted by technical performance. Adjacent work in healthcare and public systems shows that interpretability readiness, organizational readiness, human oversight capacity, and regulatory alignment can independently constrain responsible adoption even when models are accurate (Dubey et al., 2024, Legara et al., 17 May 2026).

A plausible implication is that robust AI deployment requires alignment across multiple layers. At the domain level, the field must already be technologically integrated with AI; at the data level, assets must be cleaned, structured, and provenance-backed; at the institutional level, approvals, oversight, and budgets must exist. The supplied literature does not collapse these layers into a unified formula, but it repeatedly shows that failure at any one layer can block realized returns. Domain AI readiness is therefore best understood not as a generic innovation slogan, but as a measurable meso-level condition linking external technological evolution to firm-level productivity and innovation outcomes (Zeng et al., 13 Aug 2025).