Seven enterprise AI-agent adoption issues, raised by Aaron Levie (CEO, Box) in an X post cross-posted to LinkedIn on July 8, 2026, mapped one-to-one to Semantic Web / Linked Data solution principles — with Virtuoso as the exemplar implementation of each.
On July 8, 2026, Aaron Levie — CEO of Box, the intelligent content management platform, and a prolific commentator on enterprise AI-agent adoption who writes frequently on operating-model redesign, the enterprise “company brain” knowledge-base pattern, multi-model routing, and the role of Forward Deployed Engineers in the applied-AI layer (self-described on LinkedIn as having “an unhealthy obsession with enterprise software,” based in the San Francisco Bay Area) — posted a set of observations to X after meetings with roughly two dozen enterprise IT leaders about AI agents, then cross-posted the identical content to LinkedIn the same day (owl:sameAs-linked in the companion RDF).
“Just coming off of meetings with a couple dozen enterprise IT leaders discussing AI agents. Here are a few of the common themes that stand out” — operating model challenges, data fragmentation, data moats, outcome-based adoption metrics, multi-model routing, talent shortage, and workflow-transforming use-cases. This page maps each of those seven themes to a Semantic Web / Linked Data solution principle, with Virtuoso as the exemplar implementation.
The seven common themes Aaron Levie observed across enterprise IT leadership conversations about AI agents.
Most companies have orgs that have always operated in silos, but agents work best when tied to a process that cuts across those silos. The open question: how do you deploy centrally managed agents that work across organizational boundaries — who manages them, and how do they get adopted?
Data remains highly fragmented and not in standard formats, or is not available to the right people and agents. This spans both structured data (product metrics, revenue figures) and unstructured data (roadmaps, customer contracts), undermining answer accuracy and business-practice conformance.
If every company has access to roughly the same model-level superintelligence, the context fed into those models becomes the future source of proprietary value. Capturing that context and getting it into a usable format becomes critical.
Tokens are generally agreed not to be the right metric. Consensus is leaning toward business outcomes (more revenue, more shipped product), but measuring outcomes requires getting close to each individual workflow — making top-down management harder.
Enterprises are converging on a multi-model world, with growing interest in routing layers that send workloads to different frontier or open-weight models for cost/performance reasons — and in deciding what to give models directly versus what to keep as swappable horizontal systems and context.
Talent for driving AI adoption and implementation remains a major issue. Many enterprises conclude they must train this talent internally because of a shortage of externally trained candidates — which is also framed as a major opportunity for those who get good at deploying and managing agents.
The best AI use-cases fundamentally change the work being done, rather than just replacing an existing process more efficiently. Because this varies by industry, companies must work through their own versions individually — but this remains the highest-upside category of use.
Each issue matched 1:1 against a Semantic Web / Linked Data solution principle.
Levie's problem: agents work best when tied to a process, but a process cuts across silos an org chart was built to keep apart. Here is the fix. A federated Linked Data graph doesn't force finance, legal, and support to merge their databases — it denotes every department using a hyperlink into one shared graph, and lets a single Agent Skill query across all three without touching a system of record. Skills aren't welded to one backend; per Idehen's loosely coupled data spaces architecture, the same Skill runs against a database today and a knowledge graph tomorrow. The graph becomes the operating model your org chart never was — centrally managed, cross-boundary, and owned by no single department.
Levie's issue is that enterprise data is highly fragmented. Enterprise knowledge graphs, deployed as Semantic Webs — courtesy of Linked Data principles — address this problem head-on. Here is what it costs you to adopt one: nothing you already built. RDF Views (R2RML) map your existing relational schema into a standard graph live, without migrating a row. Full-text and vector indexing do the same for the unstructured content sitting outside your databases — contracts, roadmaps, decks. One Agent Skill, one graph-query surface, both kinds of data, no ETL project, no new tool bolted on for every source.
Levie says context, not model access, is the future moat. Agree. Here is how you build one that survives every model release: capture proprietary context as RDF sentences where hyperlinks denote subject, predicate, and object — which makes each fact easy to look up by both humans and AI agents — tracked with a provenance-oriented ontology so every fact is traceable to where it came from. Standards-based, not vendor-locked — when the next frontier model ships, your moat needs no re-platforming, because it was never sitting inside any model's weights. The graph is the asset. The model is a rental.
Levie's consensus: tokens are the wrong metric, business outcomes are the right one — but outcomes require getting close to each individual workflow, which is what makes them hard to manage top-down. A provenance-oriented ontology closes that gap. Every agent action becomes a recorded activity linked to the business entity it touched — a contract reviewed, a report shipped, a deal closed. Aggregate that provenance graph with one graph query and you get revenue-influenced, cycle-time-compressed, product-shipped numbers, per workflow, on demand — not a token bill that tells you nothing about the value delivered.
Levie expects enterprises to live in a multi-model world, routing work to whichever model is cheapest or best for the job. That only works if context doesn't have to move house every time the model does. Hyperlinks put your context in the graph, not the prompt — the loosely coupled 'Model' layer in Idehen's agentic Model-View-Controller pattern, separate from the 'Controller' agents that route work to it. Swap one model for another on Tuesday; the graph doesn't notice, and neither does your context.
Levie's shortage is real: too few people trained to deploy and manage enterprise AI agents, so most companies train internally or fall behind. The fix isn't more hiring — it's turning what your best people already know into something you never have to re-teach. Agent Skills are reusable, discoverable, portable units of procedural knowledge, packaged as a SKILL.md file and installable across Claude Code, Codex, Cursor, and any agent framework that speaks the open standard — three separately valuable layers: agent access, agent skills, and agent tooling. The FDE/IRE Enterprise Skills & Lifecycle Ontology (see the Talent section below) governs exactly this: nine named AI Literacy Skills, a Proposed → Validated → Deployed → Deprecated lifecycle, and two roles — Forward Deployed Engineers who deploy Skills into customer environments, Inward Redeployed Engineers who codify tacit knowledge into new ones. Train the Skill once. Deploy it everywhere. Levie frames the other side of this shortage as opportunity: enterprises are actively looking for whoever gets good at deploying and managing these agents — precisely the capability a governed Skill practice builds.
Levie's best use-cases fundamentally change the work, not just speed up an existing step. That happens when an agent can see the whole process, not one application's database. A knowledge graph exposes every entity and relationship across a workflow, so a loosely coupled agent — the Controller in Idehen's agentic Model-View-Controller pattern — can eliminate handoffs and approvals instead of merely accelerating one legacy stage. Uber's Agentic Pods proved it at scale: cross-functional pods redesigned whole workflows and shipped 2,500+ agent Skills in the process. The workflow, not the task, is the unit of automation once the whole graph is visible.
The full seven-dimension alignment matrix: issue, Semantic Web solution, and the specific Virtuoso capability (or, for talent, the FDE/IRE ontology exemplar) that implements it.
| Dimension | Issue | Semantic Web Solution | Virtuoso / Exemplar |
|---|---|---|---|
| Operating model | Operating-model silos Siloed orgs vs. cross-boundary agents. |
Fire your org chart as the operating model. Hire the graph. One federated RDF graph. |
Federated Graph Query Federated Graph Query |
| Data fragmentation | Data fragmentation Non-standard, scattered structured + unstructured data. |
One format. Every silo. Zero rewrites. Enterprise knowledge graph, deployed as a Semantic Web, as the universal format. |
RDF Views / R2RML RDF Views/R2RML + full-text/vector indexing |
| Data moats | Core data moats Proprietary context as future value. |
Everyone gets the same model. Not everyone gets your graph. Provenance-tracked enterprise Knowledge Graph. |
ACID RDF Quad Store ACID quad store with named-graph provenance |
| Adoption metrics | AI adoption metrics Tokens are the wrong metric. |
Stop counting tokens. Start counting outcomes, per workflow, on demand. Provenance-oriented activity-to-outcome linkage queried via a graph query language. |
ACID RDF Quad Store ACID quad store as the provenance substrate |
| Multi-model routing | Multi-model routing Enterprises will live in a multi-model world. |
Change the model on Tuesday. Don't touch the context. Hyperlinks decouple context from any one model. |
WebDAV-Hosted, Hyperlinked Graphs Content-negotiated, hyperlinked DAV-hosted graphs |
| Talent shortage | Talent shortage for AI adoption Scarce externally trained AI-adoption talent. |
You don't need more headcount. You need reusable Skills. Governed FDE/IRE skill taxonomy. |
FDE/IRE Ontology (Talent Exemplar) ire-fde-ontology-roles-claude_sonnet46-1 Knowledge Graph |
| Workflow transformation | Workflow-transforming use-cases Best use-cases change the work, not just accelerate it. |
Don't automate the step. Redesign the whole workflow. Graph-scoped, whole-process visibility instead of single-step automation. |
Sponger Middleware Sponger middleware bringing whole-process silos into one graph |
OpenLink Software's Universal Server — a hybrid RDF quad store, relational engine, RDF Views/R2RML mapping layer, graph-query federation endpoint, and full-text/vector search engine — used here as the exemplar Semantic Web implementation. Six capabilities map directly onto the seven issues above.
Maps existing relational schemas into RDF graphs live, without duplicating or migrating data — directly resolves the data-fragmentation and non-standard-format issue.
Queries across multiple internal and external endpoints in one request, letting a single agent traverse organizational silos without a rebuilt central warehouse.
Named-graph partitioning keeps provenance, department, and workflow context attached to every triple — the substrate for outcome-linked, auditable metrics.
Unstructured content (contracts, roadmaps, tickets) is indexed and embedded alongside structured triples in the same engine, closing the structured/unstructured fragmentation gap in one store.
Every entity minted in Virtuoso's DAV space is a resolvable IRI (as used throughout this document and the FDE/IRE ontology it cites), making the enterprise graph directly linkable and agent-navigable rather than locked in an opaque database.
Cartridge-based extraction that lifts non-RDF sources (documents, feeds, APIs) into RDF on the fly, giving legacy silos a path into the shared graph without a rewrite.
Direct exemplar for the talent-shortage issue: a governed role and skill taxonomy for AI-native engineering, published as a hyperlinked Knowledge Graph — not a proposal, an already-published artifact.
Unified knowledge model for Forward Deployed Engineers, Inward Redeployed Engineers, 9 AI Literacy Skills, skill-lifecycle governance, and provenance-oriented tacit-knowledge codification — with a real-world showcase from Uber's Agentic Pods.
A hyperlink that, when followed over HTTP, returns a description of the entity it names — the mechanism by which Linked Data entities cross system and organizational boundaries. Technically an Internationalized Resource Identifier (IRI), but functions exactly like the hyperlinks on any web page.
A W3C-standard mapping that exposes an existing relational database as an RDF graph without copying or migrating the underlying data.
A graph query language feature letting one query span multiple independent endpoints — the query-time equivalent of crossing organizational silos.
A provenance-oriented ontology records which activity generated a piece of data, by whom, and from what — used here to link agent activity to measurable business outcomes.
A triple store extended with a fourth, named-graph element per statement, used to attach provenance, source, or department context to every fact.
Levie's term for the proprietary context an enterprise feeds to AI models — durable competitive value once base model capability converges across vendors.
Outward-facing engineering role embedded in customer environments to deploy and adapt AI systems, per the FDE/IRE ontology.
See also: FDE role definition (FDE/IRE ontology)
Inward-facing role that extracts tacit organizational knowledge and codifies it into structured, AI-executable Skill artifacts, per the FDE/IRE ontology.
See also: IRE role definition (FDE/IRE ontology)
Five-step path for an enterprise to move from Levie's observed issues to a Linked-Data-grounded operating model, using Virtuoso as the reference implementation.
Use R2RML/RDF Views to expose each department's existing relational schema as RDF, without moving or duplicating data. This produces the first cross-boundary graph agents can traverse, addressing the operating-model and data-fragmentation issues simultaneously.
Route contracts, roadmaps, and other unstructured content through full-text/vector indexing and Sponger-style extraction into the same quad store as the structured RDF Views, so agents query one interface for both data classes.
Record each agent action as an activity linked to the business entities it touched. This replaces token-based reporting with per-workflow outcome metrics queryable via a graph query language.
Store durable context and identifiers in the hyperlinked RDF graph rather than inside any single model's prompt state, so a multi-model routing layer can swap models without reformatting context.
Apply the FDE/IRE Enterprise Skills & Lifecycle Ontology's nine AI Literacy Skills and Proposed→Validated→Deployed→Deprecated governance to internal training, turning ad hoc upskilling into an auditable program.
Interactive force-directed view of every entity and relationship in the companion RDF. Click a node or edge label to open its full description via URIBurner.
This page was generated from a companion RDF-Turtle knowledge graph (schema.org vocabulary plus a provenance-oriented ontology) mapping Aaron Levie's July 8, 2026 observations about enterprise AI agents to Semantic Web / Linked Data solution principles, with Virtuoso positioned as the exemplar implementation of each, then rendering that graph as this interactive infographic and companion Markdown document.