§17.1

The Agentic Turn

Part VI — the data-to-decision loop, now operated by AI agents.

Every chapter so far has assumed a human at the controls. A person framed the question, wrote the SQL, fit the model, read the chart, and signed the memo. The methods changed; the operator did not. This final part asks what changes when the operator is an AI agent — when a system can pull data from the warehouse, run an analysis, watch a model for drift, and draft the recommendation, calling on the human only at the moments that matter. This is not science fiction. By 2026 it ships as product. The job of Part VI is to survey what is real, what is hype, and what it means for how a data-driven organization is run.

The word agent has been stretched to cover everything from a chatbot to a science-fiction android, so the first task is to pin it down. The distinction that has held up — and that the major labs now share — is about who controls the workflow.

Workflow or agent?

Anthropic's widely cited guidance draws the line cleanly: a workflow is a system where models and tools are orchestrated through predefined code paths, while an agent is a system where the model dynamically directs its own process and tool use, keeping control over how it accomplishes the task.¹ OpenAI's framing agrees from the other side — an agent is a system that independently accomplishes tasks on your behalf; an app that calls a model but does not let it control execution (a simple chatbot, a single-turn classifier) is not an agent at all.²

Two ways to put a model to work

Workflow

The developer fixes the steps in code. The model fills the blanks at each station.

Extract→Classify→Summarize→Route

Predictable, testable, bounded. The path never changes.

Agent

The model chooses the next step, calls a tool, reads the result, and decides again — until the goal is met.

Model decides⇄Tool / environment

Flexible and open-ended. The path is discovered at run time.

Anthropic draws the line here: a workflow orchestrates the model through predefined code paths; an agentlets the model dynamically direct its own process and tool use. Both are “agentic systems” — the difference is who holds the steering wheel.

Figure 1. The dividing line is who chooses the next step. In a workflow the developer fixes the path; in an agent the model discovers it at run time.

The difference matters for managers because it changes what you are buying and what can go wrong. A workflow is bounded and testable: it does the same five steps every time. An agent is flexible and open-ended: it can handle a question you didn't anticipate — and it can also wander somewhere you didn't intend. Most production systems are still workflows with a little agency sprinkled in. The frontier is moving toward genuine agents, and the rest of this part is about that frontier.

What a data agent is made of

Strip away the marketing and a data agent has a recognizable anatomy. At the center is a language model — the reasoning core. Around it sit four augmentations: tools (the ability to run SQL, search, or execute code), memory (state that persists across turns), retrieval (access to documents and the firm's semantic layer), and planning (the ability to decompose a goal and reflect on progress). What turns these parts into an agent is the loop: the model plans, acts on the world through a tool, observes the result as ground truth, reflects, and decides what to do next — stopping when the goal is met or when it is time to ask a human.¹

Anatomy of a data agent — an augmented model inside a control loop

The core is a language model augmented with tools, memory, retrieval, and planning. What makes it an agent is the loop: it acts on the world, reads the result back as ground truth, and decides what to do next — stopping when the goal is met or when a human is asked to approve.

Figure 2. A data agent is an augmented model running a control loop. Snowflake describes its Cortex data agents with exactly this shape — a model plus tools, running plan → act → observe → reflect — operating inside the warehouse's governance perimeter.

This is not a hypothetical diagram. Snowflake describes its Cortex data agents in precisely these terms: a model (Anthropic's Claude) combined with a text-to-SQL tool and a search tool, running a four-step planning → tool use → reflection → monitor-and-iterate loop inside the data platform's existing security boundary.³ The remaining chapters of this part take the four augmentations in turn — querying data (§17.2), running predictive workflows (§17.3), connecting to tools (§17.4), and governing the whole thing (§17.5).

The autonomy dial

The most important managerial idea in this part is that autonomy is a setting, not a property. The same model can be deployed to merely suggest, to draft-and-wait, to act only after a human approves, or to run a routine loop unsupervised. Autonomy is co-determined by the model, the human oversight wrapped around it, and the product design — which means it is something you choose, deliberately, per task.

Levels of autonomy — the dial a manager actually sets

AssistHuman does the work; model suggests

Human does the work; model suggests

DraftModel proposes; human edits and runs

Model proposes; human edits and runs

Act-with-approvalModel executes after a human gate

Model executes after a human gate

Supervised autonomyModel runs the loop; human monitors

Model runs the loop; human monitors

Delegated autonomyModel owns the task end to end

Model owns the task end to end

Autonomy is not a property of the model — it is a setting the deploying team chooses, co-determined by the model, the human oversight around it, and the product design. The same model can sit at rung 2 for a pricing change and rung 4 for a routine data refresh.

Figure 3. The autonomy dial. A pricing change might sit at rung 2; a nightly data refresh at rung 4. The dial is set by the deploying team, not by the model.

Where the market actually is

It is easy to mistake the volume of agent announcements for the volume of agent deployment. The honest picture is a wide experiment-to-scale gap. McKinsey's 2025 survey found that while 88% of organizations use AI regularly and 62% are at least experimenting with agents, only 23% are scaling an agent in even one function — and fewer than 10% are scaling agents in any single function. Just 39% report enterprise-level profit impact.⁴

The experiment-to-scale gap (McKinsey, State of AI 2025)

Using AI regularly88%

At least experimenting with agents62%

Scaling an agent in ≥1 function23%

Scaling agents in any single function10%

Almost everyone is trying agents; almost no one has them running the business yet. Gartner expects the same wave to thin out — more than 40% of agentic-AI projects are forecast to be canceled by the end of 2027 over cost, unclear value, and weak controls.

Figure 4. Almost everyone is experimenting; almost no one is running agents at scale. The shape of every credible 2025 survey.

The trajectory, though, is steep. Gartner expects 40% of enterprise applications to embed task-specific AI agents by the end of 2026, up from under 5% a year earlier;⁵ Deloitte projected the share of generative-AI adopters running agentic pilots to climb from a quarter in 2025 toward half by 2027.⁸ And the data platforms are already reporting scale: within three months of launch, Snowflake said more than a thousand customers had deployed over fifteen thousand AI agents on its Intelligence product.⁷

The same forces that drive adoption also drive disappointment: Gartner forecasts that more than 40% of agentic-AI projects will be canceled by the end of 2027, undone by cost, unclear value, and weak controls.⁶ That is not a reason to stay out. It is a reason to bring discipline — the kind this book has spent seventeen chapters building.

Sources

Verified June 2026