Closing the Enterprise AI Context Gap with Relational AI

/ AI, Application Development, Decision Intelligence, Enterprise Applications, Machine Learning

What’s Happening

Relational AI CEO Molham Aref joined ECI Research’s Paul Nashawaty on the AppDevAngle to make a pointed case: the enterprise AI market has a context problem, not a model problem. Organizations are pouring capital into larger foundation models and more inference capacity, yet business value from AI agents in production-grade decision workflows remains stubbornly elusive. Aref argues that closing this “AI value gap” requires executable, relational context grounded in business semantics, not markdown-wrapped data dumps fed into an LLM prompt. The conversation covered tokenomics, the limits of SQL for decision intelligence, and Relational AI’s approach of combining relational ontologies with post-training to make smaller open-weight models smarter about a specific business.

The Bigger Picture

The AI Value Gap Is a Context Architecture Problem

The phrase “AI value gap” is becoming the enterprise AI industry’s honest reckoning with itself. Vendors spent 2023 and 2024 racing to the top of benchmark leaderboards; 2025 and 2026 are revealing that benchmark-maximized models don’t automatically translate into agents that can price a product, rebalance a supply chain, or make a credible hiring recommendation.

What Relational AI adds to the conversation is specificity about the form that context must take. Text and markdown representations of enterprise data are, in Aref’s framing, a kind of cognitive handicap for agents. Business logic that was once encoded in supply chain applications, pricing engines, and fraud detection systems lives in relational structures, and agents that cannot natively consume and produce relational data are operating without the same inputs a human analyst would have. The implication is architectural: organizations cannot fix the value gap by prompt engineering alone. They need an entirely different context layer.

What This Means for ITDMs

For IT decision-makers, the economic dimension of this problem is becoming urgent. Aref’s reference to one customer consuming 27 billion tokens per day is not an abstraction. Token costs at that scale are a P&L matter, and boards are beginning to ask pointed questions about AI ROI that innovation budgets cannot indefinitely absorb. The distinction Aref draws, between a dollar spent on experimentation and a dollar spent on production that generates revenue, maps directly to how CIOs need to be reframing AI investment conversations with their CFOs.

The implication is a maturity shift. Organizations that have been running parallel AI experiments need to begin consolidating toward production architectures with measurable decision outcomes. ECI Research’s 2025 AI Builder Summit survey found that 44% of enterprise AI leaders have only moderate confidence that AI agents can act autonomously without human intervention. That finding, paired with the token economics Aref describes, suggests the near-term enterprise AI opportunity is not autonomous AGI-style agents. It is well-governed, cost-efficient agents that augment human decision-making in specific, bounded operational domains.

ITDMs evaluating vendors in this space should be asking a direct question: does this platform provide context that is executable and native to our relational data? This would reduce token consumption and improve accuracy.

What This Means for Developers

For developers and architects, Relational AI’s approach surfaces a real tension in AI system design. Most current agent frameworks assume that giving an LLM access to a data store via tool calls is sufficient. Aref’s argument is that without the semantic layer, the agent still lacks the business reasoning capacity to do anything useful with that access. Asking an LLM to sum profit across a filtered relational table by invoking a SQL query is solvable. Asking it to optimize a pricing strategy across product lines, geographies, and margin targets is a categorically different problem, one that requires prediction tools, graph reasoning, and rule-based business logic that SQL cannot express.

The practical architecture Aref describes is a layered stack: a relational ontology that models business concepts and computed relationships, a suite of specialized tools (prediction, graph reasoning, goal-seeking solvers) that the agent calls rather than trying to replicate in-context, and a training data generation loop that uses that ontology to post-train smaller open-weight models. The post-training piece is especially relevant for teams under infrastructure cost pressure. Rather than routing every query through a frontier model, organizations can develop business-specific models that carry enterprise context in their weights rather than burning tokens to reconstruct it on every inference call.

This is a meaningfully different architecture from what most enterprise teams are building today. ECI Research’s 2025 AI Builder Summit found that half of enterprise AI leaders say their organizations still rely primarily on public AI tools like ChatGPT or Copilot. Moving from public tool reliance to a governed, business-contextualized agent architecture is a significant engineering undertaking. But the economics of token consumption at scale make it an increasingly necessary one.

Competitive Positioning

Relational AI is entering a market with well-funded competitors operating at the context and data layer. What distinguishes Relational AI’s positioning here is the explicit focus on post-training as a cost reduction and accuracy improvement mechanism, not just retrieval-augmented generation. That framing aligns with where the frontier labs themselves are heading. The market supports Aref’s argument that the next competitive frontier is business-specific, efficiently trained models, not raw parameter counts.

Looking Ahead

The Next 18 Months: From Experiment to Production Decision Intelligence

The window between now and late 2027 will likely determine which AI architecture patterns become the enterprise standard. Organizations that move quickly to build relational, executable context layers and instrument their agents for cost accountability will enter that window with a durable advantage. Those still routing general-purpose queries through frontier models without semantic context will face either escalating token costs or diminishing returns on AI investment, or both.

The prototype-to-production gap remains one of the hardest challenges in this market. As ECI Research has observed, many organizations can demonstrate promising proofs of concept but cannot operationalize them reliably, with barriers including lack of governance frameworks, performance unpredictability, cost volatility, and integration challenges across legacy and cloud-native systems. Relational AI’s ontology-based approach addresses governance and integration complexity, but enterprises will also need to invest in the data architecture work required to instantiate those ontologies.

A Market Rewarding Context, Not Just Capability

The broader signal from this conversation is that the competitive differentiation in enterprise AI is moving from model capability to contextual intelligence, and from raw experimentation budgets to measurable decision outcomes. Vendors that can offer executable business semantics, cost-efficient inference architectures, and clear paths from AI output to business action will earn enterprise spend. Vendors that cannot will find themselves positioned as expensive infrastructure rather than strategic partners. Relational AI is betting, credibly, that the relational paradigm underlying every enterprise system of record is the right substrate for that next layer of AI value.

Author