Enterprise AI Agents Grapple with Context and Semantics Amid Shift to Actionable Outcomes

The artificial intelligence industry is rapidly evolving beyond the initial awe of generative AI 1.0, with enterprises now demanding concrete, measurable outcomes from advanced agentic systems. This shift marks the “third inning” of the modern AI era, moving past academic discovery and initial hype towards practical implementation and tangible productivity gains. The central premise underpinning this evolution, as highlighted by industry experts, is the critical importance of data within its proper context, enabling AI agents to act, coordinate, and learn across complex organizational landscapes riddled with diverse data, workflows, policies, and nuanced semantics.

This transition is not without its complexities, as market watchers scrutinize vague narratives and enterprises grow impatient for visible results and control. The modern data stack of the 2010s, with its cloud-centric approaches and separated compute and storage, now appears insufficient for the demands of agentic systems that must navigate a mess of structured and unstructured data, conflicting policies, and multiple identities. As detailed in a recent SiliconANGLE report, the industry, while still excited, remains conflicted, facing the challenge of reliably deploying AI agents to deliver on the promises of productivity.

The Evolving Landscape of AI Adoption and Investment

Spending momentum in the machine learning and AI space remains highly elevated, despite a modest deceleration, according to quarterly spending intentions surveys by Enterprise Technology Research (ETR) involving over 1,700 IT decision-makers. This sustained investment underscores the strategic importance enterprises place on AI, with leading benefits consistently cited as productivity, decision support, and business transformation. Notably, there has been a significant uptick in citations for “future headcount avoidance,” moving from 21% of respondents in mid-2024 to 30% recently, signaling a growing expectation for AI to streamline operations and optimize human capital.

However, the path to consistent positive outcomes is not always smooth. A Workday survey revealed that while 85% of respondents reported AI saving one to seven hours a week, more than a third of this time saving was lost to correcting errors, rewriting, and verifying content. Only 14% consistently reported positive outcomes, indicating a persistent gap between AI’s potential and its reliable delivery within enterprise settings. This suggests a need for more refined AI capabilities, particularly in ensuring accuracy and reducing the burden of human oversight.

Navigating the Competitive AI Ecosystem

The competitive landscape among AI players continues to shift dynamically. OpenAI Group PBC and Microsoft Corp. maintain leadership in account penetration and spending velocity, despite recent negative sentiment. Conversely, Meta Platforms Inc.’s Llama has seen a decline in momentum, while Anthropic PBC is thriving. Amazon Web Services Inc. and Google LLC remain strong contenders, with Google actively closing the gap on AWS. Data management specialists like Snowflake Inc. and Databricks Inc., previously off the ML/AI radar, are now pivotal players, demonstrating significant spending velocity and account penetration, particularly as the linchpin of AI increasingly becomes data.

This fluid environment compels customers to make strategic bets amidst considerable market noise. Legacy players are actively fighting for relevance, while new entrants and evolving technologies continually reshape the ecosystem. The constant flux underscores the challenge for enterprises to identify and implement the most effective AI solutions that align with their specific needs and deliver tangible value.

2025 Predictions: A Mixed Review for Agentic AI’s Early Promises

A review of 2025 predictions by a panel of industry experts from The Cube Collective and Data Gang revealed a mixed performance, highlighting both progress and persistent challenges in the deployment of agentic AI:

• Sanjeev Mohan on the Rise of AI Agents: Rated yellow/green. Mohan noted that while agents have proliferated, the term has suffered from “semantic drift” and over-rebranding. Tangible wins were observed in narrow domains like coding (with Anthropic’s Claude reportedly developing 100% of its new additions) and customer service. However, the more ambitious goal of agents operating autonomously within complex, multi-step enterprise workflows remains unfulfilled, leading Mohan to conclude, “we are not there yet.” His prediction for widespread personal agents was also delayed due to technological difficulty.
• Tony Baer on the “Renaissance of Data”: Rated yellow. Baer argued that data moved “front and center” for vendors, evidenced by the rapid adoption of the Model Context Protocol (MCP) and a strong year for Postgres platforms. However, customer execution lagged, with data governance and AI governance remaining largely siloed, and persistent issues with data quality and integration, as highlighted by an AWS study. Knowledge graphs also showed less adoption than anticipated.
• Carl Olofson on the Rise of Knowledge Graphs: Revised to yellow-green. Despite their perceived value, enterprise adoption of knowledge graphs showed only a marginal increase, with early-stage activity declining. The primary gating factor was identified as the complexity of assembling and preparing inputs for implementation, constraining momentum.
• Dave Menninger on ‘LLMs to LAMs’ (Large Action Models): Rated red/yellow. Menninger’s prediction of a distinct category of ‘LAMs’ did not materialize. However, the underlying problem—LLMs’ deficiency in planning and executing tasks—was indeed addressed. Mainstream models were extended with reasoning, task orchestration, and techniques like chain-of-thought prompting, indicating that the problem was real, but the solution form-factor differed.
• Brad Shimmin on ‘Security is the Buzzkill’: Rated green. The panel unanimously agreed that security has become a critical gravitational force shaping the possibilities of AI in enterprise markets, validating Shimmin’s warning.

2026 Predictions: The Battle for Context and Semantics

The focus for 2026 pivots to the crucial role of context and semantics in scaling AI agents:

• Sanjeev Mohan: ‘Context is Mandatory’ but 2026 will be wasted debating ‘Context Graphs’: Mohan predicts that while large language and action models absolutely require context, the industry will squander 2026 debating ‘context graphs’ without achieving broad success. He critiques the absence of accepted standards, consistent implementation models, and stable definitions, drawing parallels to the data mesh concept. Mohan distinguishes between knowledge graphs (traversing transactional data) and the deeper “why” of context (from web searches, logs, emails), arguing that true context emerges when these signals are related back to traditional enterprise data.
• Tony Baer’s Response: Context is Slippery, AI May Be Required to Help Build It: Baer concurred with Mohan, describing context as “slippery” and hard to define operationally. He suggested that AI itself might play a role in guiding the assembly and organization of context, akin to how language models inspect databases. The panel reinforced the gap between conceptual discussions and practical implementation, with some arguing for manual “4D maps” of enterprises (Palantir, Celonis approaches) and others for AI to shoulder more of the heavy lifting.
• Tony Baer: Semantic Layer is ‘More Doable’ and Will Solidify into ‘Semantic Spheres of Influence’: As a pragmatic counterpoint, Baer predicts a renewed focus on semantic layers and metric stores, which he deems “more doable” than context graphs due to decades of precedent (e.g., BusinessObjects Universe). He argues that while semantics don’t solve context end-to-end, they go ‘a good bit of the way’ by providing explicit meaning and relevance beyond the ‘walled gardens’ of traditional BI. Baer forecasts that major enterprise application vendors (SAP, ServiceNow, Salesforce, Workday) will extend their influence through harmonized data models and ‘data products,’ shared while remaining resident within their systems. He also highlights the Open Semantics Interchange (OSI) framework, based on dbt, as a mechanism for portable and interpretable semantics across different systems, suggesting its emergence signals the centrality of semantics to AI systems.

The ongoing debates and predictions underscore a critical inflection point for enterprise AI. The industry is maturing, shifting its focus from demonstrating raw generative capabilities to solving the intricate challenges of integration, data quality, and contextual understanding. The divergent approaches to ‘context graphs’ versus ‘semantic layers’ reflect a fundamental disagreement on the most effective and actionable path to achieving truly intelligent and autonomous agentic systems that can reliably deliver on their immense promise.