Everyone Is Building It, Nobody Has Named It

When Google acquired Cask, I moved from a startup where shipping was an act of courage to a company where shipping was an act of process. And I will admit something founders are not supposed to admit: the process was magnificent.

Inside Google Cloud, no change of consequence traveled alone. It moved wrapped in machinery: review, tests, canary, monitoring, rollback, postmortem. The individual engineer was not necessarily better than engineers I had worked with anywhere else. The engineer's loop was better. Failure was cheap, visible, and educational, and so the whole system compounded. That experience fused with everything I saw before and after it into one conviction: the durable advantage is never the intelligence at the center. It is the loop around it.

Which is why the current moment fascinates me. Part 3 ended with a teaser: a dozen companies, different logos, different category names, all building components of the loop this series has been describing, mostly without saying so. This part is the map.

Recall the destination from Part 1: a decision-and-outcome observability layer. The market has no single name for it, but watch what six supposedly different product categories are converging toward:

That is the enterprise equivalent of what Google built around its engineers, and of what the industry built around coding models: a repository, tools, execution, tests, diagnostics, and retries. Now the evidence, category by category.

Process mining used to answer a historian's question: what actually happened in our process? The next generation asks an operator's question: what is happening now, what should happen next, and can an AI or a workflow intervene? ServiceNow now markets process mining explicitly as fact-based visibility for people and AI agents, bottlenecks and rework included, and can mine the execution and audit logs of AI agents themselves, so organizations can see what their agents actually do in production.¹ Microsoft's stack spans organization-level event logs down to desktop task mining, and this March added object-centric process mining so interacting business objects stop being flattened into one case ID.² UiPath's newest case-management product is built around long-lived, exception-heavy cases where humans and agents exercise judgment at decision points, with history retained toward an auditable outcome.³

Read that last sentence against Part 2 of this series. Long-lived cases, judgment at decision points, history retained toward an auditable outcome. That is a decision episode with a vendor logo on it.

There is a persistent belief that human work is simply too messy to instrument. Industry practice quietly disproved it years ago: commercial task-mining tools capture clicks, keystrokes, hotkeys, and screenshots, and merge task variants into representations of how work is actually performed.⁴ Capture is solved. Semantic compression is not. An organization does not need to learn that an employee clicked Save. It needs to learn: at this decision point, after seeing this state, people with this objective tend to choose this path, producing these immediate results and these longer-term outcomes. That is the transformation from activity telemetry to experience, and it is exactly the transformation the click-capture products have not made.

Meanwhile, the context side is industrializing. Microsoft positions Work IQ as a workplace intelligence layer that gives agents semantic organizational context with permission-aware governance.⁵ Palantir goes further conceptually: its ontology puts data, logic, actions, and security in one decision-centric model, and explicitly argues that decision lineage should feed future decisions.⁶ These companies share no product history and no architecture team. They share a direction: agent intelligence is being surrounded by organizational context, process state, governed actions, execution telemetry, and observability.

There is even a literal version of the analog transition. In late August, Anthropic previewed its Model Hardware Standard, letting AI agents operate programmable scientific and industrial devices, microscopes and robotic arms included.⁷ The protocol matters less than the direction: physical instruments exposing state, actions, and sensor feedback to agents makes physical experimentation resemble a software loop. Which produces this series' favorite paradox: a robotic arm or a warehouse scanner often gives cleaner machine-readable feedback than a strategy meeting. The hardest analog domain was never the physical world. It is human judgment.

Put the whole map together and the gap between the industry's talking points and its build sheets becomes the most revealing chart in enterprise AI:

Six categories, one architecture, no name. Which raises the question every operator reading this has been muttering since Part 1: fine, but where does this actually work first? Because the loop does not close evenly. Some workflows are practically begging for it. Others will eat your roadmap and give you nothing back. I have a scoring function for that, and a ranking that surprised me when I built it.