.png){kind=logo}
By Kauts Shukla Founder and CEO of Dview Technologies
The first wave of enterprise AI bolted language models onto dashboards. The second wave standardized the plumbing between agents and systems. Neither closed the gap between near-universal adoption and rare bottom-line impact. What decides the next five years is the layer almost nobody puts in the budget: the machine-readable representation of what a business means, who may see it, and which decisions an agent is allowed to make.
There is a particular ritual every enterprise analyst knows. You open the revenue dashboard and see that Germany is down. You then open a second dashboard for product mix, use a SQL tool to analyse the data by region, refer to a third report for last quarter’s context, and connect with someone in the Berlin office. Forty-five minutes later, you have a hypothesis, not an answer.
Here is what changed by 2026, and what did not. Every one of those tools now has a copilot built in. Some of them have agents. The process is faster in certain ways, but the structure remains largely the same. You are still the integration layer.
That is the honest state of enterprise AI in 2026, and the numbers are unusually clear about it. McKinsey’s State of AI survey, published in November 2025 with 1,993 respondents across 105 countries, found that 88% of organizations now use AI in at least one business function, up from 78% a year earlier. However, in the same survey, only 39% were able to attribute any enterprise-level EBIT impact to AI, with most reporting gains of less than 5%.
Enterprise software has evolved in layers, each abstracting away a previous layer of pain. ERP digitized the record of the business. CRM tracked its customers. Warehouses stored its history, the cloud made that history elastic, and the modern data stack industrialized ingestion, transformation, and orchestration. Every layer assumed a human would do the final mile of reasoning.
The semantic layer was the first honest admission that data alone is not knowledge. Net revenue is not a column. It is a definition, a set of business rules, and a history of decisions about what counts. Semantic layers formalized that so it could be reused. In retrospect they were the first attempt to make enterprise context machine-readable, built for a consumer that had not arrived yet.
Then language models made the interface plastic, and the industry’s first instinct was to point them at dashboards and search boxes. That phase is largely over, and it ended not because it failed technically but because it never touched the thing that made the work slow. The dashboard was never the problem. The dashboard was a symptom.
A dashboard is a precomputed answer to a question someone once thought to ask. It is frozen at design time while the business moves at runtime. That critique was correct in 2024 and it remains correct. The reason the analyst needed forty-five minutes was never rendering. It was that the knowledge required to answer the question lived in nine places with nine access models, three conflicting definitions of churn, and a policy PDF nobody had read since the last audit. Give a language model a natural language front end to that mess and you get a faster path to a confident wrong answer.
Gartner put a number on the consequence in June 2025, predicting that more than 40% of agentic AI projects will be canceled by the end of 2027 on cost, unclear business value, or inadequate risk controls.
Two years of extraordinary model progress, and the value is concentrated in a thin slice of firms. That is not a capability failure. It is an architecture failure, and it is worth being precise about where it sits.
This distinction still carries the argument, so it is worth slowing down for. AI-assisted analytics keeps the old architecture and stitches a model onto the front. The dashboard stays, the semantic model stays, the warehouse stays. A model translates natural language into SQL, summarizes a chart, or suggests a next question. The system gains no new understanding of the business. It gains a nicer door.
AI-native analytics inverts the relationship. The model is not a feature on top of the stack; it is a participant in it. The system maintains a continuous representation of what the business means, retrieves context on demand across structured and unstructured sources, plans a multi-step investigation, executes it under policy, and returns an answer with a traceable chain of evidence. The dashboard becomes one possible rendering of that answer, generated for the moment rather than stored as a permanent artifact.
The practical difference shows up immediately. Ask an AI-assisted system why churn increased in Germany and it translates the question into SQL, runs it, and hands you a table. An AI-native system recognizes the question as causal rather than descriptive. It pulls churn data, retrieves support tickets and account notes, consults the definition of churn in force this quarter, reasons over candidate drivers, runs the comparisons that discriminate between them, and returns an explanation with citations and a confidence boundary. One returns data. The other returns an argument you can audit.
Almost everything hard about building the second kind of system lives below the interface.
The most important shift of the past year is that the industry stopped arguing about interfaces and started fighting over context. At the Gartner Data and Analytics Summit in Orlando in March 2026, analysts opened by calling context the new critical infrastructure, the brain for AI rather than an accessory to it. Their conference research is pointed: four in five organizations increased AI investment, while roughly one in five can show measurable return.
Gartner now treats semantics as a cost and trust strategy, forecasting that organizations prioritizing semantics in AI-ready data will improve agentic accuracy by up to 80 percent and cut costs by up to 60 percent by 2027.
One prediction from that summit deserves to be pinned above every architecture review: roughly 60 percent of agentic analytics projects that rely solely on protocol connectivity, with no semantic foundation underneath, are expected to fail by 2028. Connectivity without meaning is a very efficient way to distribute confusion.
The platform vendors are converging on the same conclusion, which usually means the constraint is real. Their first joint release included Agents Schema, an open standard that designates one schema in the warehouse as the shared context layer, holding metric definitions, semantic models, lineage, and documentation in plain tables.
The other big change of the past year is that agent connectivity stopped being a differentiator. When the integration surface becomes a standard, integration stops being an advantage. What stays proprietary to your organisation is what the connections mean, who may traverse them, and which actions an agent may take without a human. That is not a protocol. That is your semantics, your policy, and your evidence trail.
Enterprise knowledge has never lived in one place. It sits in tables, logs, documents, email, policy PDFs, images, code, chat threads, recorded calls, and tickets. For decades we tolerated the fragmentation because no system could reason across all of it, so we bought one tool per silo and hired people to stitch the results together in their heads. That stitching is what retrieval plus reasoning can now do at machine speed, provided the pieces are represented well enough to be joined.
The ingredients are familiar: a semantic layer encoding business definitions, vector search over unstructured corpora, retrieval grounded in enterprise content rather than model memory, a knowledge graph of entities and relationships, and a model that can plan a multi-step investigation instead of answering in one shot.
There is an economic argument for that ordering.
If frontier capability is becoming a substitutable commodity input, betting your architecture on one model is a strange bet, and betting it on your own context is the obvious one.
Once the substrate is unified and the system can reason over it, the interaction model changes in a way that is easy to describe and hard to internalise. The analyst asks one question. Why did churn increase in Germany this quarter, and what should we do about it? The system decomposes it, gathers structured and unstructured evidence, applies the organization’s definitions and access rules, reasons over candidate causes, and returns a grounded recommendation with a trail back to source records. The analyst’s job shifts from gathering to judging.
Microsoft’s 2026 Work Trend Index, drawing on trillions of anonymized Microsoft 365 signals and a survey of 20,000 AI users across ten countries, shows this happening in telemetry rather than in theory.
The technology has been ready for a while. The operating model has not.
Every benefit described above collapses the moment a system fabricates a metric, surfaces a record the user is not cleared for, or takes an action nobody authorized. In the dashboard era, governance was access control at the data layer. In the agent era it has to extend into retrieval, reasoning, and action.
Permissions must follow the model into every retrieval step, so a passage inherits the constraints of the document it came from. Every answer needs lineage back to source records, for audit and for the user’s own calibration. Hallucination control is grounding architecture rather than prompt hygiene, which means the system must be able to say it cannot answer.
This is now the loudest theme in the analyst community rather than the quietest.
Capability gets you a demo. Governance gets you into production.
The language model wave answered whether machines could understand language. Yes, decisively. The agent wave is answering whether they can complete work. The answer there is more interesting: often, but not dependably.
It is not a bigger model. It is a reliability discipline: context engineering, so the model sees the right slice of the enterprise rather than the whole warehouse; evaluation infrastructure that tests agents against your business instead of a leaderboard, which is why every major platform shipped agent evaluation and observability tooling this year; memory that survives across sessions without accumulating garbage; and orchestration that keeps a failed step from becoming a wrong action. Frontier labs will keep raising the ceiling. Enterprises get paid for raising the floor.
So what does this look like when the work is done? Employees will not open software in the way we mean today. They will state an intent, in the language they think in, and the system will assemble evidence, apply definitions and policy, reason through the question, and return an answer with a recommendation and a trail. Dashboards will not disappear so much as become ephemeral, generated for the question and discarded when it moves. Reports become conversations. Knowledge becomes executable, invoked and acted on without a human re-reading it every time.
The enterprise operating system in that world is less a database than a reasoning system with four things at its center: a living semantic model of the business, a graph of entities and relationships, an index over unstructured context, and a policy engine governing what may be seen and done.
The human stays in the loop for anything that commits money, changes a customer relationship, or alters a governed process. What changes is the unit of work. People stop gathering and start deciding.
None of this is inevitable, and pretending otherwise is the fastest way to delay it.
Accuracy remains the first hard problem. Reasoning models reason, but they reason from whatever premises you hand them, and a confident wrong answer is more dangerous than an admission of ignorance. Grounding reduces this without eliminating it.
Cost is second, and its shape has changed. The floor keeps falling, with Anthropic cutting Opus-class pricing by two thirds in November 2025, while frontier reasoning at maximum effort has grown more expensive rather than less. Agentic systems that reason at length over large contexts produce bills nobody modeled, which is exactly why Gartner frames good semantics as a cost control.
Security is third, and genuinely new territory. An agent that can act across systems can be instructed to act wrongly across systems, and prompt injection through retrieved content is a supply chain problem wearing a language costume.
The organisations that get to 2030 first will not be the ones with the best models. Models are converging into a shared and substitutable substrate. They will be the ones that did the unglamorous work: building the semantic layer, curating the knowledge graph, wiring access control into retrieval rather than around it, standing up evaluation harnesses, and redesigning the workflow around the agent instead of around the dashboard it replaced.
Some things survive the transition. The warehouse is not going away, it is being wrapped. The semantic layer is not going away, it is becoming the contract between the model and the business. Governance is not going away, it is becoming the control system for decisions rather than a gate on tables. The human decision-maker is not going away, they are being promoted from gatherer to judge.
What disappears is the navigation tax, the endless switching between tabs, the time spent searching through reports, and dashboards that become outdated almost as soon as they are created. A chart may show what changed, but rarely explains why it happened or what action should follow. Enterprises have always been rich in data and knowledge, yet struggled to access and apply that intelligence at the right moment. Every generation of enterprise tools has narrowed this gap, but the responsibility of interpretation has remained with people. AI can now close that gap, but only for organisations willing to define and structure their own context. The bottleneck is no longer intelligence, which is increasingly accessible across platforms; it is context, which cannot be bought because it is unique to every business. The next phase of enterprise AI will be less about building smarter models and more about making organizations understandable to machines that are finally capable of learning from them.
