The Horizon: Where Agent-Operated Analytics Is Headed

Two true stories at once

The bull case is real and well-sourced. Gartner projects that a third of enterprise software will embed agentic AI by 2028, up from under 1% in 2024, and that 15% of day-to-day work decisions will be made autonomously by then.¹ The reality check is just as well-sourced. An MIT study found that about 95% of enterprise generative-AI pilots delivered no measurable profit impact;² S&P Global found the share of firms abandoning most of their AI initiatives jumped from 17% to 42% in a single year;³ and RAND put the AI-project failure rate above 80%, roughly twice that of ordinary IT projects.⁴

Why most fail — and who doesn't

The failures are not mysterious. The MIT work diagnosed a “learning gap”: systems that never retain feedback or improve. It also found a sharp split in how organizations succeed — buying from specialized vendors worked about 67% of the time, while internal builds succeeded roughly half as often.² McKinsey's numbers point the same way: adoption is wide but shallow, with only 23% of organizations scaling even one agentic system and rarely beyond a function or two.⁵ The winners share a profile — narrow, workflow-embedded deployments with a clear owner and a tight feedback loop. The losers share one too: broad, unscoped “let the AI figure it out” ambitions.

The durable lesson: own the contract

If one architectural idea survives the hype cycle, it is the one from §17.2: the semantic layer is the contract that lets every agent in a chain resolve business logic identically. A governed metric definition — what “active customer” means, how “margin” is computed — is what keeps a fleet of agents consistent, auditable, and correct.⁸ Models will keep changing; the contract is the asset that compounds. The firms that win the agentic transition will be the ones that treated their definitions as infrastructure long before the agents arrived.

The reliability ceiling

It also pays to be precise about what agents still cannot do. Even text-to-SQL, the most mature data-agent capability, trails human experts: on the BIRD benchmark the best systems reach about 82% execution accuracy against a 93% human baseline — an 11-point gap that, on a metric that drives decisions, is enormous.⁶ That ceiling is why autonomy is being deployed cautiously and measured. Anthropic's analysis of real agent usage found that 80% of agent tool calls came from agents with at least one safeguard in place and 73% appeared to keep a human in the loop, while only about 0.8% of actions looked irreversible.⁷ Autonomy is expanding — but unevenly, and with humans deliberately kept close to anything that cannot be undone.

The analyst's job moves up a level

For the reader of this book, the most consequential shift is to the work itself. As agents absorb the mechanical tasks — writing the query, building the dashboard, running the model — the human moves “above the loop.” The scarce, durable skills become framing the question, owning the semantic layer the agents depend on, and judging whether an answer can be trusted.

What it all converges on

Standardization is quietly doing the unglamorous work that makes the rest possible: MCP for agent-to-data access and A2A for agent-to-agent collaboration, both now under neutral governance, are turning a fragmented field into an interoperable stack.¹¹ And the products keep getting more autonomous — Databricks' Genie and Snowflake's Intelligence for governed analytics, OpenAI's ChatGPT agent operating its own computer and asking permission before consequential actions.⁹¹⁰ Put it together and the picture is not a robot replacing the analyst. It is the data-to-decision loop this book opened with — now with an agent able to turn each crank, and a human above it who still owns the question, the definitions, and the call.