Why AI Founders Are Racing to Own the Context Layer

What happened

A September Forbes Council Post by Vishal Marria, founder and CEO of the London data-intelligence company Quantexa, has become a useful lens for a much bigger argument running through enterprise-AI coverage this year: that the real bottleneck in corporate AI adoption isn't the model, it's the data feeding it 16. Marria's piece opens with a story from his years working with bank investigators — individual accounts and transactions that looked clean in isolation but revealed a fraud network once linked through shared addresses, phone numbers and ownership records 1. His argument is that most AI strategies invest heavily in defining what data means (semantic layers) and retrieving relevant documents (RAG) while neglecting a third, harder problem: whether the records scattered across different systems actually refer to the same real-world person, company or account 1910.

That argument arrives at a moment when a wide range of outlets, research firms and infrastructure vendors are converging on the same diagnosis, even if they use different vocabulary to describe it.

The context-layer debate

Several technical explainers published around the same period try to pin down what a "context layer" actually is, and they largely agree on the shape of the problem even as they disagree on the label. Tellius describes it as the architectural tier between enterprise data and AI agents, combining semantic definitions, entity resolution, governance, lineage and memory — and explicitly distinguishes it from a semantic layer, which only standardizes metric definitions 9. SurrealDB frames the same territory as a three-layer stack — knowledge graph, semantic layer, context layer — where the context layer is the dynamic retrieval mechanism that assembles the right fragment of data at the moment an agent needs it 10. DataHub goes further, arguing that the term has become genuinely contested: the semantic-layer vendor calls context "metric definitions," the knowledge-graph vendor calls it "entity relationships," and the RAG startup calls it "a vector database and retrieval pipeline" 20. Quantexa's own materials add a fourth voice, describing the same landscape as "contextual decision intelligence" and insisting no single vendor owns the whole stack 181314.

That last point is the tell. Marria isn't just diagnosing an industry problem — he's also describing the business Quantexa has run since 2016, when he and co-founders including Jamie Hutton, Imam Hoque, Laura Hutton and Felix Hoddinott built the company around entity resolution and knowledge graphs for banks, insurers and government agencies chasing fraud and financial crime 1314. The company's fundraising history tracks the market's growing appetite for this pitch: TechCrunch reported a $129 million Series E in 2023 at a $1.8 billion valuation, up from an $800–900 million valuation on a $153 million Series D in 2021 15. Tech Funding News reported a further $175 million Series F in 2025 at a $2.6 billion valuation, with Quantexa crossing $100 million in annual recurring revenue and adding 23 customers in 2024 16. Quantexa's own QuanCon25 materials cite a Forrester-commissioned study claiming a three-year 228% ROI for its platform 17 — a company-sourced figure, not an independently verified one.

The adoption data behind the urgency

The timing of Marria's argument lines up with McKinsey's 2025 state-of-AI research, which several outlets cite to explain why enterprises are suddenly preoccupied with data foundations rather than model capability. McKinsey found that workflow redesign, more than any other of 25 attributes tested, correlated with whether organizations saw bottom-line impact from generative AI — yet only 21% of organizations using gen AI say they've fundamentally redesigned even some workflows, and more than 80% report no tangible enterprise-level EBIT impact 7. A separate summary of the same McKinsey data, published by CoLab, sharpens that further: roughly 79-80% of organizations report regular gen AI use in at least one function, but only 5.5% — 109 of 1,933 respondents — qualify as high performers where AI drives more than 5% of EBIT 8. CoLab also flags what it calls the "agentic gap": fewer than 10% of organizations are scaling AI agents in any function, and 73% aren't using agents at all in product development 8. Read together, these figures support the broader claim running through the context-layer pieces: adoption is not the bottleneck anymore. Reliability, governance and integration into real workflows are.

The founder-strategy angle

For startup builders, this debate has direct implications, and TechCrunch's founder-focused coverage fills in that side of the picture. One piece argues that distribution, not product quality, is now the deciding factor for AI startups, because building software has gotten so easy that differentiation has to come from go-to-market execution, hiring judgment and trusted advisers rather than a better demo 11. Another lays out a more operational playbook: build an accurate cost model before scaling, choose deliberately between renting cloud AI capacity and hosting models, prioritize clean training data over chasing GPU hardware, and — critically — go vertical rather than horizontal, since a narrow, well-served industry problem is harder for a fast-following competitor to copy than a general-purpose wrapper around someone else's model 12. A third piece on AI agents joining startup teams describes founders increasingly deciding what work should go to a human hire versus an AI agent from day one, reframing team-building itself as a context-and-delegation problem 19.

These threads point toward the same conclusion Marria reaches from the enterprise side: a startup whose only asset is a chat interface wrapped around a foundation model is exposed, because model vendors can absorb that layer quickly. Defensibility increasingly comes from things that are hard to replicate overnight — customer-specific data mappings, resolved identities, audit trails, and integration into the actual system where a decision gets made, not a dashboard that sits unopened 120.

Where the reporting agrees

Across the Forbes column, the architecture explainers, the McKinsey-derived research and the TechCrunch founder advice, there is strong convergence on a few points. First, retrieval-augmented generation alone is insufficient for production AI — every technical source describes RAG as one component of a larger context architecture, not a synonym for it 91020. Second, the industry has moved from an adoption problem to a value-realization problem, with McKinsey's figures cited consistently across sources to show that most organizations using gen AI still aren't seeing it move their bottom line 78. Third, vertical focus and workflow-specific data are seen as more durable competitive advantages than general-purpose product features, a view shared by TechCrunch's startup-economics piece and echoed implicitly in Marria's advice to start with one consequential decision rather than a universal platform 112.

Where it doesn't

The clearest divergence is definitional, not factual: outlets and vendors do not agree on what a "context layer" actually is. Tellius insists a semantic layer is only one component of a true context layer and that conflating the two leaves "a semantic layer and a gap" 9. SurrealDB treats context, semantic and knowledge layers as three distinct tiers in a stack 10. DataHub argues the disagreement is structural and self-interested — each vendor defines the term to match what it already sells, whether that's a semantic layer, a knowledge graph, a governance framework or a retrieval pipeline 20. Quantexa, notably, sidesteps the definitional fight altogether by arguing no single vendor owns the category, which is itself a strategic position rather than a neutral one 18.

There's also a gap in evidentiary weight. Marria's Forbes column offers no new data, benchmark or customer study — it is explicitly labeled a fee-based Council Post reflecting one executive's expertise, not independent reporting 16. Quantexa's own performance claims, including the 228% ROI figure and 90%-plus accuracy claims from QuanCon25, are company-commissioned and promotional rather than independently verified 17. By contrast, the McKinsey adoption figures, while cited secondhand through CoLab's summary, come from a large survey (1,933 respondents) and are corroborated across two separate write-ups 78.

The reading the evidence supports

Taken together, the sources support treating Marria's argument as a legitimate description of a real production problem — unresolved identity and fragmented relational data genuinely do undermine agentic AI in ways that better models can't fix — while also recognizing it as advocacy for Quantexa's existing business model rather than disinterested analysis. The McKinsey-derived adoption numbers independently validate the symptom Marria describes: enterprises are struggling to convert AI pilots into bottom-line results. But nothing in the broader reporting independently validates his specific prescription — that entity resolution and knowledge graphs, Quantexa's core products, are the necessary cure. The more durable and better-supported claim, corroborated across the architecture explainers and the TechCrunch founder-strategy pieces, is the general one: AI startups and enterprise buyers alike are shifting their attention from model quality to data foundations, workflow integration and governance, and founders who build defensibility there — rather than in a thin layer atop someone else's model — are better positioned for the next phase of enterprise AI.