By Kauts Shukla Founder and CEO of Dview Technologies
The initial phase of enterprise AI involved integrating language models into dashboards, while the subsequent stage standardized the underlying infrastructure connecting agents and systems. However, neither approach bridged the divide between widespread software adoption and tangible financial returns. The outcome of the coming five years depends on an overlooked component: a machine-readable framework that defines business concepts, access permissions, and permitted agent actions.
Enterprise analysts frequently face a familiar routine. Upon noticing a decline in German revenue via a revenue dashboard, an analyst must consult a separate product-mix dashboard, execute SQL queries to analyze regional data, review past quarterly reports for context, and contact the Berlin office. This workflow typically requires forty-five minutes merely to generate a hypothesis rather than a definitive answer.
By 2026, the tooling landscape shifted to include built-in copilots and agents across nearly all applications. While operations operate more quickly in specific respects, the overall structure remains largely unchanged, leaving human workers to serve as the primary integration layer.
Data reflecting the state of enterprise AI in 2026 highlights this dynamic. McKinsey’s November 2025 State of AI survey, which gathered responses from 1,993 participants across 105 countries, indicated that 88% of organizations deploy AI within at least one business function, up from 78% the previous year. Yet, only 39% of respondents linked any enterprise-level EBIT impact to these deployments, with the majority noting gains of under 5%.
How we got here
Enterprise software historically advanced through successive layers designed to reduce operational friction. Enterprise Resource Planning (ERP) systems digitized business records, Customer Relationship Management (CRM) tools monitored clients, data warehouses preserved historical records, and cloud computing provided elastic storage. Meanwhile, modern data stacks automated ingestion, transformation, and orchestration. Throughout this evolution, human workers handled the final stage of analysis.
The emergence of the semantic layer represented an early acknowledgment that raw data differs from actual knowledge. Net revenue functions not merely as a spreadsheet column, but as a defined metric governed by specific business rules and historical decisions. Semantic layers formalized these metrics for reuse, serving as an initial, early attempt to make business context machine-readable.
When language models subsequently made software interfaces more flexible, the industry’s initial reaction involved directing them toward existing dashboards and search boxes. This phase concluded not due to technical failure, but because it bypassed the root cause of analytical delays. Dashboards were never the primary obstacle; rather, they acted as symptoms of fragmented information.
Because dashboards present precalculated answers to past questions, they remain static while businesses evolve dynamically. The aforementioned analyst required forty-five minutes not because of rendering speeds, but because the necessary insights were scattered across nine locations utilizing distinct access models, three conflicting definitions of customer churn, and policy documents left unread since previous audits. Providing a language model with a natural-language interface atop such fragmentation frequently yields a rapid path to an incorrect conclusion.
Addressing the consequences of these limitations, Gartner projected in June 2025 that over 40% of agentic AI initiatives would face cancellation by late 2027 due to excessive costs, unclear business value, or insufficient risk controls.
Despite two years of rapid model advancements, practical value remains concentrated among a small group of firms. This imbalance stems from architectural challenges rather than technological limitations.
AI-assisted is not AI-native
This distinction remains central to the discussion. AI-assisted analytics preserves legacy architectures by attaching a language model to the front end, leaving dashboards, semantic models, and data warehouses intact. In these setups, models translate natural language into SQL queries, summarize charts, or propose follow-up questions without imparting a deeper operational understanding to the system.
Conversely, AI-native analytics reverses this dynamic by embedding the model directly into the core architecture. Rather than functioning as a mere interface feature, the system maintains a continuous representation of business definitions, retrieves context dynamically from structured and unstructured sources, plans multi-step inquiries, operates within established policy guidelines, and returns responses supported by verifiable evidence. Dashboards are generated dynamically on demand rather than stored as permanent artifacts.
Practical differences emerge clearly during investigations. When asked why customer churn increased in Germany, an AI-assisted system translates the query into SQL, executes it, and outputs a data table. An AI-native system evaluates the query as causal rather than descriptive, retrieving churn records alongside support tickets, account notes, and current quarterly churn definitions. It then assesses potential drivers, executes comparative evaluations, and provides an explanation complete with citations and defined confidence boundaries.
The core challenges of developing the second type of system reside beneath the user interface.
Context became critical infrastructure, and the market noticed at once
A notable shift over the past year involved the industry moving away from interface debates and focusing instead on context management. During the March 2026 Gartner Data and Analytics Summit in Orlando, analysts characterized context as essential infrastructure that serves as a foundation for AI systems rather than a peripheral accessory. Summit research indicated that while 80% of organizations increased AI spending, roughly 20% achieved measurable returns.
Gartner presently positions semantic frameworks as vital components of cost and trust strategies, forecasting that organizations prioritizing semantics within AI-ready data environments will enhance agentic accuracy by up to 80% and reduce expenditures by up to 60% by 2027.
Another summit projection highlights a key architectural risk: approximately 60% of agentic analytics initiatives that depend exclusively on protocol connectivity without a foundational semantic framework are projected to fail by 2028.
Major platform vendors are reaching similar conclusions regarding these constraints. Initial collaborative releases included Agents Schema, an open standard that designates a single schema within the data warehouse as a shared context layer containing metric definitions, semantic models, data lineage, and documentation in standard tables.
The plumbing standardized faster than anyone expected
Agent connectivity ceased to serve as a competitive differentiator over the past year. Once integration methods become standardized, connectivity offers fewer advantages. Proprietary value instead resides in the specific meaning of those connections, access permissions, and the actions agents may execute independently of human oversight.
The unification of enterprise knowledge
Enterprise information has historically remained decentralized across database tables, operational logs, documents, emails, policy files, images, code repositories, chat transcripts, recorded calls, and ticketing platforms. Organizations tolerated this fragmentation because legacy systems lacked cross-source reasoning capabilities, necessitating siloed tool purchases and manual synthesis by human analysts. Modern retrieval and reasoning technologies now perform this synthesis at machine speed, provided underlying data structures support integration.
Essential components include semantic layers encoding business rules, vector search functions operating over unstructured data, retrieval mechanisms grounded in enterprise content rather than static model memory, knowledge graphs mapping entities and relationships, and models capable of executing multi-step investigations.
Economic incentives reinforce this structural hierarchy. As foundational model capabilities increasingly become commoditized inputs, constructing architectures around proprietary context proves more sustainable than depending entirely on any single model.
What actually changes about the work
Once underlying data substrates are unified and systems can reason across them, interaction models shift significantly. When an analyst submits a complex inquiry regarding quarterly churn increases in Germany alongside recommended actions, the system breaks down the request, collects structured and unstructured evidence, applies organizational definitions and access protocols, evaluates potential root causes, and delivers grounded recommendations accompanied by verifiable source trails. Consequently, the analyst’s role transitions from data collection to evaluation.
Insights from Microsoft’s 2026 Work Trend Index—derived from trillions of anonymized Microsoft 365 signals and surveys involving 20,000 AI users across ten countries—indicate that this operational shift is already occurring.
Although technological readiness has advanced, operating models continue to adapt.
One practical approach involves embedding context directly into analytics layers rather than leaving it distributed across dashboards, documents, and informal channels. Platforms such as Hex enable users to query systems using natural language and construct analyses while data teams oversee underlying business definitions, governance, and contextual frameworks. This approach enhances self-service capabilities without treating isolated model capabilities as comprehensive solutions.
Governance stopped being access control and became decision governance
The advantages of these systems diminish rapidly if a platform fabricates metrics, exposes unauthorized records, or executes unapproved actions. While dashboard-era governance focused primarily on data-layer access controls, agentic environments require governance across retrieval, reasoning, and operational execution.
Permissions must accompany models throughout every retrieval phase so that retrieved text inherits the security constraints of its originating document. Furthermore, responses require clear lineage back to source records to support auditing and user validation. Hallucination mitigation relies on grounded architectural design rather than prompt engineering alone, requiring systems to acknowledge when they lack sufficient information to answer.
Such considerations have become prominent focal points within analytical communities.
While strong base capabilities facilitate product demonstrations, comprehensive governance remains necessary for production deployment.
What comes after the agent wave: reliability engineering
Initial language model advancements confirmed machines’ ability to process language effectively. Current agentic developments address whether systems can successfully complete complex workflows, yielding mixed reliability results.
Addressing these limitations requires reliability disciplines rather than simply larger models. Key requirements include context engineering to ensure models evaluate relevant enterprise segments rather than entire warehouses, evaluation infrastructure designed to test agents against specific business metrics rather than generic leaderboards—prompting the release of new agent evaluation and observability tools—persistent session memory management, and orchestration protocols that prevent failed execution steps from triggering incorrect actions. While foundational labs continue to enhance model capabilities, enterprise value depends heavily on establishing operational reliability.
A believable 2030
Future workflows will likely involve employees stating objectives in natural language, prompting systems to assemble evidence, apply relevant policies and definitions, evaluate the problem, and return actionable recommendations with clear verification trails. Dashboards will transition into ephemeral artifacts generated for specific queries and discarded thereafter. Reports will function as interactive conversations, and institutional knowledge will be invoked dynamically without requiring manual reviews.
In this environment, enterprise operating systems will function as reasoning frameworks centered around living semantic models, entity relationship graphs, unstructured context indexes, and policy engines governing visibility and execution.
Human oversight will remain mandatory for actions involving financial commitments, customer relationship modifications, or regulated processes, shifting the primary unit of work from manual information gathering to active decision-making.
The honest list of what can go wrong
Realizing these operational changes requires addressing several notable challenges.
Accuracy remains an initial hurdle. Reasoning models evaluate premises provided to them, making confident errors potentially more dangerous than acknowledged uncertainty, though grounding mechanisms reduce these risks.
Cost structures represent a second challenge. Although baseline pricing has declined—such as Anthropic reducing Opus-class pricing by two-thirds in November 2025—advanced reasoning tasks executed at maximum capacity have grown more costly. Agentic systems analyzing extensive contexts can generate unexpected expenses, reinforcing Gartner’s emphasis on semantic efficiency as a cost-control measure.
Security presents a third, novel concern. Agents operating across multiple systems remain vulnerable to unauthorized operational instructions, and prompt injection attacks delivered via retrieved content represent a distinct supply chain vulnerability.
Organizations achieving success by 2030 will likely prioritize foundational administrative tasks—such as constructing semantic layers, curating knowledge graphs, integrating access controls directly into retrieval mechanisms, establishing evaluation frameworks, and redesigning workflows around agent systems rather than legacy dashboards.
What stays, and what doesn’t
Certain core elements will persist through this transition. Data warehouses will remain integrated within wrapper layers, semantic models will function as contracts between business operations and models, governance will evolve into decision-control frameworks, and human decision-makers will shift from data gatherers to evaluators.
Conversely, navigational friction, constant tab-switching, manual report searches, and static dashboards risk obsolescence. While charts indicate historical changes, they rarely explain underlying causes or prescribe necessary actions. Although enterprises have historically struggled to access and apply internal intelligence effectively at the right moments, artificial intelligence can bridge this gap for organizations willing to formalize and structure their own operational context. The primary limitation moving forward involves context rather than raw intelligence, making the next phase of enterprise AI less about deploying advanced models and more about making organizations legible to machine systems.
