A new open convention for packaging knowledge could help AI agents use regulatory content more reliably while leaving the harder work of legal meaning governance and accountability to the organizations that provide it
Google’s Open Knowledge Format deserves serious attention from the compliance sector because it gives AI agents a portable way to inspect curated knowledge together with its sources, review signals and lifecycle. Its value will depend on disciplined content providers, however, because an open file convention cannot replace legal interpretation, domain context, professional verification or accountable maintenance.
A new format for agent knowledge
Google Cloud introduced the Open Knowledge Format, or OKF, on June 12, 2026, as an open convention for packaging knowledge so that people, software and AI agents can inspect and exchange it more easily. Version 0.2 followed in July with additional fields for provenance, verification, freshness, lifecycle and attested computation. The rapid revision reflects a wider change in the AI market, where retrieving information is becoming easier than proving that it is current, attributable and suitable for a consequential decision.
The format is deliberately modest. An OKF bundle consists mainly of Markdown files with YAML metadata, ordinary links and a small set of conventions for indexes and logs. The specification does not require a proprietary database, a particular model, a cloud service or a central registry, while Git is recommended as one convenient way to distribute bundles and preserve a readable history. A company can therefore prepare knowledge in a familiar form without committing its content to one AI provider or retrieval platform.
This simplicity explains much of the interest around OKF. Organizations have spent years placing valuable knowledge inside websites, document systems, intranets, ticketing platforms and specialist portals, yet an AI agent often encounters the result as a collection of pages with inconsistent structure and limited evidence about authorship or validity. OKF offers a common envelope in which a concept can travel together with its description, source references, status and trust signals. The format does not determine whether the concept is correct, but it gives a receiving system more information with which to make that judgment.
For compliance, that distinction becomes especially important. An agent answering a product question can sometimes recover from an incomplete source, but an agent interpreting a fiscal, tax or regulatory requirement may influence software behavior, reporting, documentation or a customer’s legal position. The value of an open knowledge format therefore depends less on how easily it can store text than on whether it helps preserve the chain between a statement, its source, its review history and the conditions under which it applies.
Version zero point two makes trust visible
The first release of OKF established the basic file structure and a small vocabulary for describing concepts, while version 0.2 added the signals that make the format relevant to professional knowledge. A concept can identify its sources, record how it was generated, show whether it has been verified, declare whether it is a draft or a stable item and specify when it should be treated as stale. These fields allow an agent to inspect important facts before loading the full content or relying on it in a response.
Provenance begins with the sources field, which can point to a resource and give it a stable identifier, title, author and modification date. Claims in the Markdown body can refer to individual sources through footnotes, allowing a reader or agent to connect a particular statement with the evidence behind it. OKF avoids assigning a universal credibility score, because the authority of a source depends on the task and domain. An official regulation, an implementation guide and an expert interpretation may all be useful, although they should not be treated as interchangeable.
The generated and verified fields separate authorship from confirmation. A concept may have been created by a person, a model or an automated process, while a different person or system later verifies it against the cited material. Google describes three advisory trust levels that can be inferred from this information: unverified, machine-confirmed and human-reviewed. The distinction does not guarantee accuracy, but it helps consumers avoid treating polished machine-generated prose as reviewed professional knowledge.
Freshness is represented through an absolute stale_after timestamp, while the status field distinguishes draft, stable and deprecated concepts. These are simple controls, yet they address a weakness that appears repeatedly in AI systems built on document collections. A response can be factually faithful to a source and still be wrong for the present moment because a later rule replaced it, an implementation deadline moved or the underlying guidance changed. Explicit lifecycle signals give software a deterministic reason to exclude or flag material before generation begins.
Version 0.2 also introduces Attested Computation, a concept through which a knowledge bundle can identify an approved computation, its parameters, the executor and the receipt expected from execution. A separate attester can then check whether the approved process ran and whether the displayed result matches the authoritative output. Several technical details remain deferred, and OKF does not itself execute the calculation, but the proposal is important because it connects knowledge with evidence that a sanctioned method was followed rather than relying entirely on a plausible explanation.
Compliance knowledge creates a harder test
Compliance exposes the difference between information retrieval and dependable knowledge more sharply than most business domains. A regulatory document may be authentic while a summary of it is incomplete, and a summary may be accurate while the rule is not yet effective. A binding obligation may apply only to certain taxpayers, transaction types, sales channels or implementation periods, while an exception elsewhere in the same framework changes the practical result. An agent that retrieves the right paragraph can still reach the wrong conclusion when it lacks this surrounding context.
The problem becomes more difficult across countries because regulatory concepts rarely align through vocabulary alone. Similar business events may be described through different legal traditions, administrative processes and technical terminology, while identical words can carry different consequences in separate jurisdictions. Compliance intelligence must therefore preserve the country context, legal status, temporal position and implementation meaning of a rule instead of relying only on semantic similarity between passages.
The authority of each source also requires careful treatment. An official law, a ministry announcement, a tax authority FAQ, a technical specification and a professional interpretation can all influence implementation, although each occupies a different position in the evidence chain. The provenance fields in OKF can help keep those sources visible, but the organization producing the knowledge must still explain their role and resolve conflicts where possible. A format can carry authority signals; it cannot manufacture authority when the underlying editorial work is weak.
Time adds another layer of difficulty because legal change rarely follows a single date. Publication, entry into force, technical availability, mandatory application, transitional relief and enforcement may occur at different moments. OKF’s stale_after field is useful for deciding when knowledge deserves another review, but it does not model the full temporal structure of regulation. Compliance applications will still need domain-specific ways to express when a rule becomes relevant and when a previous interpretation ceases to apply.
These limits do not make OKF unsuitable for compliance. They show why an open container should be judged by the quality of the domain knowledge placed inside it. The format can improve portability and make trust signals easier to inspect, while legal specialists, compliance providers and responsible organizations remain accountable for interpretation, applicability and maintenance.
How compliance portals may evolve
Specialist compliance portals like our https://www.fiscal-requirements.com were originally designed for human navigation. They organize official sources, news, explanations and technical guidance so that professionals can follow developments and understand their operational impact. AI agents create a second audience with different needs, because they require content that can be located, interpreted and cited without reconstructing the meaning of every web page from its visual layout.
OKF offers one possible bridge between those two audiences. A portal can continue to provide the readable experience, expert context and research discipline that human users expect while making selected knowledge available in a portable form that an agent can inspect. The public opportunity is broader than converting pages into Markdown, because useful agent knowledge also needs evidence about origin, review and freshness. In this sense, OKF supports the evolution from a document portal toward compliance intelligence without prescribing the internal design of that evolution.
The Fiscal Requirements Portal illustrates why this direction matters. Its public value rests on country-aware content, official sources, regulatory updates and expert interpretation of implementation consequences. As portals of this kind develop, open formats may allow approved knowledge to serve both professionals and AI systems while the underlying research and validation remain controlled by the knowledge provider. The standard does not replace the portal, because the portal is where the content gains context and responsibility before it reaches another system.
This relationship also prevents an exaggerated view of automation. An agent-readable format does not turn every legal text into an executable rule, and a collection of well-structured files does not automatically become compliance intelligence. It creates a cleaner boundary through which maintained knowledge can move. The intelligence still comes from the connection between authoritative material, expert interpretation, operational context and a process that keeps those elements current.
What the format does not solve
OKF’s flexibility is one of its strengths and one of its most important limitations. Only the type field is mandatory, producers may add their own fields and consumers are expected to tolerate unfamiliar concept types. A minimal file can therefore comply with the specification while offering little value for a serious decision. Technical conformance proves that the content follows an exchange convention, but it says nothing by itself about completeness, professional quality or legal reliability.
The status vocabulary provides another example. Draft, stable and deprecated are useful descriptions of a knowledge object’s lifecycle, although they cannot express whether a law is proposed, adopted, published, effective, suspended, repealed or subject to a transition. Those distinctions belong to the compliance domain and require additional structure. The same principle applies to applicability, jurisdiction, exceptions and the difference between official material and interpretive guidance.
Relationships in OKF are ordinary Markdown links rather than formally typed legal connections. A link may indicate that one concept explains another, replaces it, depends on it or creates an exception, but the base format does not require software to distinguish among those meanings. That choice keeps OKF simple and readable, while leaving specialist domains responsible for the richer semantics needed for automated reasoning.
The trust tiers should also be understood as advisory signals. Human-reviewed is more informative than an unlabeled block of generated text, but the label does not reveal what the reviewer examined, which professional role the reviewer held or whether the review covered legal accuracy, technical implementation or editorial clarity. Organizations using OKF for consequential material will need review practices that give those signals a precise and auditable meaning.
Access control also sits outside the specification. Google notes that repositories alone do not provide enterprise discovery, identity, permissions or policy governance, which is why its larger example connects OKF with Knowledge Catalog and identity controls. This separation is sensible because an exchange format should not dictate how every organization secures information, although it means that businesses cannot treat an OKF bundle as a complete governance solution.
OKF alongside other agent standards
OKF enters an environment that already includes several proposals for making digital information easier for AI systems to discover and use. These technologies are sometimes discussed as competitors even though they address different layers. Understanding those distinctions is especially important in compliance, where a convenient connection must not be mistaken for trustworthy content.
The llms.txt proposal gives website owners a simple way to direct language models toward selected resources and cleaner versions of pages. It mainly helps an agent discover where useful content is located. OKF goes further by packaging knowledge concepts with metadata, links and trust signals, while remaining independent of the public website through which the material may first have been found. A site could use both conventions, with llms.txt supporting discovery and OKF supporting portable representation.[5, 7]
The Model Context Protocol, or MCP, addresses connection and interaction. It defines how an AI application can discover resources and tools offered by an external system, while a server can notify clients when resources change or allow controlled operations to be invoked. OKF can provide the form of the knowledge delivered through such a connection. MCP helps an agent reach a system and understand the available actions; OKF helps the receiving agent understand the structure and trust signals of the content.
APIs and structured web data continue to serve their established purposes. An API may deliver precise records under authentication, while Schema.org markup helps search systems understand entities and pages. An XML sitemap identifies indexable locations, and a conventional document remains useful for human reading. OKF does not eliminate any of these surfaces, because the same governed knowledge may need different representations for different consumers.
Retrieval technology belongs to another layer again. Embeddings can identify semantically related passages, and graph structures can represent explicit relationships, but neither establishes that a regulatory statement remains effective or applies to the case at hand. OKF can carry curated content into those environments while preserving useful context, although the retrieval and reasoning systems must still interpret that context responsibly.
The role of legal and compliance standards
OKF explicitly avoids replacing domain schemas, which is essential for understanding its place in regulatory technology. The law contains structures that a general knowledge format cannot express through a few common fields. Obligations, permissions, prohibitions, exceptions, jurisdiction, temporal conditions and conflicts between rules require a richer representation when software is expected to reason over them.
Akoma Ntoso provides a structured vocabulary for legislative, judicial and parliamentary documents, preserving the components and references within formal legal texts. LegalRuleML addresses another level by representing legal statements, their sources, temporal characteristics, jurisdiction and defeasible reasoning. These standards pursue more specialized objectives than OKF, and their complexity reflects the difficulty of translating law into structures that machines can evaluate.
These formats can therefore complement one another in practice. An OKF concept may provide a readable explanation, provenance and lifecycle information while linking to a structured legal document, a formal rule or another domain representation. This allows a general-purpose agent to understand the concept without preventing a specialist system from using more precise legal semantics where the task demands them.
Keeping the layers separate can also support better explainability. A formal rule may produce a result, but a business user still needs to see the official basis, the interpretation applied and the conditions that affected the outcome. A portable knowledge concept can connect those layers in language that remains inspectable by people while preserving references to the machine-readable structures beneath it.
Trust still depends on institutions and process
The most important trust mechanism in compliance will remain the institution responsible for the knowledge. An open format can display provenance, verification and freshness, but it cannot determine whether a reviewer has the necessary expertise, whether all relevant sources were considered or whether a difficult ambiguity was disclosed honestly. Those judgments arise from professional standards, editorial discipline and accountability.
This is why human review should not be reduced to a ceremonial approval attached to machine-generated text. Effective review requires access to the cited source, an understanding of its legal status and enough operational knowledge to identify what the interpretation changes in practice. Where the evidence is incomplete or contradictory, the knowledge should preserve that uncertainty instead of turning it into an artificial conclusion.
Machine verification also has a legitimate role. Software can confirm that links resolve, timestamps remain valid, required fields exist and quoted passages match a retained source. Those checks improve consistency and allow specialists to concentrate on meaning, although they cannot decide whether an authority’s guidance conflicts with the governing law or whether an exception changes a retailer’s obligations.
OKF is useful because it keeps these different forms of trust visible instead of collapsing them into one quality score. The consumer can see who or what generated the content, whether another process checked it and when the material may require renewed attention. The value of those signals rises when the producing organization defines them clearly and applies them consistently.
Compliance intelligence begins with usable context
The longer-term importance of OKF lies in the idea that knowledge should be maintained as an addressable artifact rather than left inside documents and prompts. A regulatory concept that retains its sources, lifecycle and connections can support search, explanation, comparison and implementation guidance without losing the evidence that gives it authority. This is one of the foundations of compliance intelligence.
Compliance intelligence goes beyond placing a chatbot in front of a document archive. It seeks to connect official material with the country context, expert validation and operational meaning needed for a dependable answer. The agent may provide a natural-language interface, but the quality of the result depends on the knowledge layer beneath it. OKF contributes a portable representation for that layer while leaving the deeper domain model to the provider.
Attested Computation points toward a later stage in which an agent’s use of approved logic can be verified. In compliance, a similar principle could eventually help connect explanations with sanctioned calculations, validation routines or policy checks. The current specification does not provide a legal execution environment, and the step from readable knowledge to executable regulation remains substantial, but the emphasis on receipts and deterministic verification is directionally important.
The future will therefore contain several levels of machine usability. Some knowledge will remain explanatory, some will become structured enough for comparison, and selected rules may eventually support controlled execution or testing. OKF can help these levels exchange context without pretending that every paragraph of law can be reduced to the same technical object.
A measured view of adoption
Organizations evaluating OKF should begin with a realistic understanding of its maturity. Version 0.2 remains young, the reference tools are described as proofs of concept and parts of the attestation approach are still deferred. The format may develop rapidly, but its long-term adoption will depend on whether producers publish useful bundles, consumers implement the trust signals consistently and domain communities define the additional conventions their work requires.
Compliance is likely to become an important test because the domain rewards portability while punishing missing context. A shared envelope could reduce the repeated conversion of the same curated material for different agent systems, yet a weak bundle would merely distribute uncertainty more efficiently. The strongest implementations will preserve a clear distinction between the open representation and the professional process that produced the knowledge.
Organizations should also avoid equating openness with unrestricted access. The specification allows knowledge to move between systems, while identity, permissions and commercial terms remain separate responsibilities. This is consistent with the broader history of enterprise standards, where common formats often increase interoperability without removing the need for controlled access or accountable ownership.
The most useful early measure will be whether OKF improves the quality and traceability of agent responses. A successful use of the format should make it easier to identify the supporting source, recognize stale material, distinguish reviewed knowledge from generated content and understand when the available evidence does not justify a confident answer. File counts and ingestion speed matter far less than those observable improvements.
The format matters because the problem is real
Google has not created a complete model for law, compliance or enterprise knowledge, and the OKF specification does not claim otherwise. It has proposed a simple, inspectable way to package concepts together with signals that agents increasingly need. That contribution is valuable because the present AI ecosystem has many methods for retrieving text and too few common ways to show why a particular piece of knowledge deserves trust.
For compliance portals, the format offers a credible direction for evolution without diminishing the role of the portal itself. Expert research, country context, source validation and interpretation remain the source of value, while an open representation may allow selected knowledge to reach the agents and applications that increasingly participate in business decisions. The portal becomes more useful when its knowledge can travel with its evidence, yet it remains responsible for the discipline behind that evidence.
OKF should therefore be understood as an enabling layer in the future of compliance intelligence. Its success will depend on disciplined content providers, meaningful domain extensions and consumers that respect provenance, lifecycle and uncertainty. If those conditions develop, the format could help move AI agents from convenient access to regulatory information toward a more accountable use of maintained compliance knowledge.
Sources and references
All web sources were checked on 12 September 2026 unless another access date is stated. The discussion of compliance portals draws only on public information and the author’s published work.
[1] Google Cloud, Introducing the Open Knowledge Format, 12 June 2026.
[2] GoogleCloudPlatform, Open Knowledge Format Version 0.2 Specification, accessed 12 September 2026.
[3] Google Cloud, Open Knowledge format v0.2 tackles agentic trust, 24 July 2026.
[4] Google Cloud, Using OKF with Knowledge Catalog to serve context for agents, 27 August 2026.
[5] Semrush, Google Launches Open Knowledge Format an AI Standard, 23 June 2026.
[6] Model Context Protocol, MCP introduction and Resources specification, 28 July 2026 specification.
[7] llms txt, The llms.txt file version 2, modified 10 August 2026.
[8] OASIS Open, LegalRuleML Core Specification Version 1.0, 30 August 2021.
[9] Fiscal Solutions, Fiscal Requirements Portal, accessed 12 September 2026.
[10] OASIS Open, Akoma Ntoso Version 1.0 Part 1 XML Vocabulary, 29 August 2018.
