AI Readiness in Education Cleared One Bar. Agentic AI Has Raised Another. 

In December 2025, we argued that digital delivery and AI readiness are different conditions. A lesson can sit inside a polished adaptive platform and still be arranged for teacher navigation rather than machine reasoning. The distinction has become more important, although the reason has shifted. Content is now being read, selected and assembled by systems that can take several bounded steps before an educator or editor reviews the result. 

Earlier AI-enabled education workflows were mainly retrieval and response workflows. A system found relevant material and generated an answer. Current agentic approaches add planning, tool use, checking and hand-offs between steps. Their level of autonomy varies considerably, and review points remain essential in educational settings. Yet even a carefully bounded agent requires content that describes what each component is, where it belongs, what it depends on and how it may be used. Findable content is a starting point. Content that can support a sequenced instructional decision requires more explicit structure. 

This follow-up extends our previous argument. In this article we consider recent publisher activity, the maturing standards and research base, and the practical implications for content teams preparing curriculum for retrieval, generation and controlled orchestration. 

What Education Publishers are Shipping into Teaching-Learning Workflows 

The transition is visible in the products publishers and education companies are putting into existing teaching and learning workflows. The common thread is not a stand-alone chatbot. It is the use of trusted, curriculum-aligned material inside the places where educators already plan, teach and differentiate. 

Company	What moved	Why it counts
Pearson	Launched AI modules across more than 20 disciplines, from business to health sciences, with Credly badges as a portable credential (announcement)	Research with AWS framed the problem as an AI-readiness gap between education and work, which is the same gap content structure has to close
Cambridge University Press & Assessment	Joined ARIAM in June, a responsible-AI coalition whose early signatories also include Disney, the BBC, the New York Times, and Adobe (NISO coverage)	Governance is formalizing around how AI systems use published content, which only becomes urgent once that use is real and at scale
HMH	Built curriculum-aligned AI tools directly into HMH Ed rather than shipping them as a separate product	Positions AI as an extension of the curriculum’s existing learning science instead of a layer bolted on beside it
Discovery Education	Launched a conversational AI interface inside its Google Classroom add-on on August 31, built on Gemini, rolling out to select districts this fall	Teachers describe a need in plain language and refine by reading level, standards, or class time, which only works if the metadata is already there
Elsevier	Launched Nora AI inside its healthcare eBooks on VitalSource Bookshelf, answering questions with source citations and follow-up prompts	Grounding every answer in the specific title being read is a content architecture commitment before it is an AI feature

These moves differ in audience and maturity, but they point in the same direction. AI features depend on material that can be identified, scoped and governed in a way the underlying workflow can trust. The feature may look conversational, but the useful work happens before the prompt. It depends on the resource, the metadata, the permissions and the pedagogical constraints that determine what can be returned or generated. 

Why Fluent AI Output is not Instructional Design 

The pedagogical case has not changed. Decisions about sequence, scaffolding, tone across grade levels, the placement of examples and the handling of misconceptions require curricular judgment. Generative systems can produce fluent explanations and agents can execute defined tasks, but neither creates a sound instructional design merely by producing plausible text. 

An agent that selects material for a review packet, proposes a sub-lesson or prepares a differentiated activity is applying a design that has already been expressed somewhere. That design may sit in a curriculum map, an instructional model, an editorial policy, a knowledge graph or a metadata schema. As systems take more steps, the consequences of missing instructions become more visible. A vague asset can produce an acceptable one-off response. In a multi-step workflow, the same ambiguity can affect selection, sequencing, adaptation and validation. 

A fluent explanation of convection, for example, does not establish that the explanation fits a particular curriculum’s scope and sequence, a stated grade-level expectation or a known misconception revealed in assessment data. The model may infer some of this from nearby material. It cannot reliably substitute inference for content that carries the necessary instructional signals. Better models improve fluency and retrieval. They do not remove the need to express pedagogical intent. 

Digital, AI-ready and Agent-ready Content 

The earlier comparison between digitally delivered lessons and AI-structured content now benefits from a third category. Agent-ready content is not a claim that an agent should make every educational decision without oversight. It describes content that gives a bounded orchestration layer sufficient information to retrieve, sequence, check and hand off instructional components within defined rules. 

The distinction also helps with investment decisions. Content already segmented at concept level and enriched for retrieval is closer to agent-ready than legacy digital content. The remaining work concentrates on relationships, permissions, provenance, validation status and dependencies. Those details allow an orchestration layer to determine whether it may use a component, what must precede it and what checks are required before it passes work to the next step. 

Feature	Digitally delivered lesson	AI-structured content	Agent-ready content
Structure	Lessons or units with multiple learning objects; sequencing follows rules or adaptive logic	Concept-level components with explicit boundaries and relationships	Concept-level components with task, dependency and validation markers usable within defined workflow rules and documented limits
Metadata	Standards, difficulty, prerequisites and basic skill mapping	Pedagogical intent, misconceptions, cognitive demand and semantic relationships	The same pedagogical fields, plus provenance, validation status and permitted-use information
Adaptation	Pre-authored paths triggered by performance or rules	Real-time retrieval and assembly supported by model reasoning	Multi-step planning and checking across content units, with review and control points
Reusability	Reusable inside a platform or object library	Interoperable and suitable for controlled generative recombination	Composable across governed tools and workflows without pre-mapping every combination

IEEE 2881-2025 and Machine-readable Learning Resource Metadata 

IEEE 2881-2025 has moved from a proposal worth watching to an active standard that publishers can build against. Formally published on 3 October 2025, it sets out a vocabulary for learning resources and learning events, designed for machine-readable representation through RDF. The distinction matters because a resource and an event carry different kinds of information. Earlier metadata practice has often brought them together without making the relationship explicit. 

The standard will not solve a content transformation programme by itself. It does, however, offer a more stable reference point for teams deciding how to represent resources, relationships and activity data. The working group’s open-source schemas and application-profile work give implementers a route from high-level terminology to a practical data model. Publishers need not wait for universal adoption before using that direction to shape an internal model. 

The value is not compliance theatre. A shared vocabulary reduces the number of local interpretations that content, product and data teams must maintain. It also makes it easier to preserve meaning when material moves between systems. It gives editorial, product and data teams a common basis for specifying change requests, validating transformed assets and resolving questions that would otherwise be handled as one-off exceptions. For a publisher with a large backlist, that is a useful discipline even before an agentic workflow enters production. 

Where Retrieval Needs a Curriculum Knowledge Graph 

The earlier distinction between retrieval-augmented generation and knowledge-augmented generation remains useful, provided it is not treated as a choice between opposing approaches. Retrieval is effective when a system must locate relevant material quickly. Knowledge structures help when the system must reason over prerequisites, conceptual relationships, constraints and curriculum rules. Educational applications increasingly combine both. 

Recent research offers working examples. A 2026 Curriculum-KAG paper describes interdisciplinary study-plan synthesis using a vector index alongside a curriculum knowledge graph that encodes prerequisites, subject domains and regulatory constraints. An IEEE-published framework for educational content management similarly combines adaptive chunking, vector embeddings and document-grounded responses. These are research examples rather than proofs of a universal production pattern, but they show the direction of travel. Semantic similarity alone does not provide the relationships required for curriculum-aligned planning. 

For content teams, the implication is practical. Topic and difficulty tags are necessary but insufficient. Systems also need conceptual relationships, prerequisite structures, instructional approach, cognitive demand, representation and known misconceptions. Those fields give retrieval a stronger basis for selection and give structured reasoning a basis for checking whether a proposed assembly is defensible. 

Knowledge Graphs and Model Context Protocol 

The most concrete commercial signal comes from publishers that are treating their knowledge assets as usable through external AI interfaces. In its August 2026 earnings call, McGraw Hill described an agentic AI strategy built around purpose-built knowledge graphs, models of different sizes and Model Context Protocol interfaces. The company reported a proprietary education ontology informed by learning interactions and described plans to make selected content available to agents through standardised interfaces. 

The details matter more than the terminology. A publisher is no longer only offering access to material that a person reads inside a publisher-controlled interface. It may also provide controlled access to content that another application’s AI system queries as part of a workflow. That use case requires clear boundaries. An external system cannot resolve a vague label through informal knowledge of a product team’s conventions. It needs machine-readable information about a component’s scope, source, permitted use and relationship to the rest of the curriculum. 

Trust is central to this approach. McGraw Hill’s reported educator research distinguished between AI embedded in familiar educational platforms and general-purpose chatbots. Discovery Education’s recent release similarly places licensed, standards-aligned content and district governance settings inside a teacher’s usual planning workflow. Grounding and governance are therefore not add-ons. They are properties that must be available to the system at the point a decision is made. 

Four Steps to Make Curriculum Content Agent-ready 

The three actions proposed in 2025 still apply. An AI-readiness audit identifies whether concepts are segmented finely enough for retrieval and whether objectives and relationships are expressed in machine-readable form. Concept-level modularisation turns composite lessons into units that can be selected and tagged independently. Metadata enrichment captures prerequisites, cognitive demand, misconceptions and instructional approach. 

A fourth action now belongs with them. Design for orchestration asks whether a component includes sufficient provenance, validation, dependency and permitted-use information for a bounded agent workflow to act on it correctly. The question is not whether a system can generate an answer. It is whether a system can make an allowed, traceable and instructionally sound decision as one step in a sequence. 

For example, an agent preparing targeted revision material should be able to identify the relevant concept, confirm the prerequisite knowledge, select an appropriate representation, avoid known misconceptions and record the source material used. It should also know when a decision exceeds its remit and requires educator or editorial review. That is what structured content makes possible. The necessary control points are explicit rather than left to model interpretation. 

Publishers that have completed the first three actions are already partway through this work. Their content has clearer boundaries, stronger metadata and a better basis for quality assurance. Extending that foundation to provenance, validation and dependencies is a focused continuation, not a separate reinvention programme. 

Where Instructional Quality Still Comes From 

Agentic AI has raised the standard for education content, but it has not changed the source of instructional quality. That remains the work of curriculum, editorial and learning-design teams. Their decisions need to become sufficiently explicit that a system can retrieve, use and check material without losing the intent those teams established. 

The practical task is to make content legible to the systems that will work with it, while retaining the human review points required for educational quality and accountability. That work starts with an audit of the material and its metadata, then develops through modularisation, enrichment and governance. It is the foundation on which retrieval, generation and agent-led workflows can be evaluated responsibly, in a repeatable and auditable way.