From Skill to Workflow:
Rebuilding the Business Loop

There is a useful lesson in the history of electricity. By the late eighteenth century, factories had access to electric power, yet for decades it barely moved productivity. Owners used electricity to light the floor or to speed up one machine, while the rest of the plant ran exactly as before. The tool was new; the workflow was old.

Real gains only arrived in the early twentieth century, when factories were rebuilt around electric power from the ground up. The entire process was redesigned, and only then did the new technology deliver the leap everyone had expected.

AI is at the same point. We now hold a large stock of intelligence assets — Skills that solve individual problems well. But if those Skills keep running inside the old, human-driven process, where a person manually invokes each one to draft an email, run a search, or translate a passage, the work never truly accelerates. The last step of AI adoption is not adding more Skills. It is redesigning the whole flow.

The word "workflow" is overloaded, so it helps to pin down a working definition. A Workflow is an executable business loop. The form does not matter — it can be a folder, a document, a slide deck, or a record on a platform. What matters is that the content describes a loop someone (or some Agent) can actually run.

A complete Workflow should contain at least six elements:

1. Goal: the business outcome the loop exists to produce.
2. Input: what the loop needs to start.
3. Process knowledge: how the work is carried out today.
4. Data: the assets and states accumulated along the way.
5. Review points: where a human must inspect or approve.
6. Next action: what happens once the loop produces its output.

If you can describe these six clearly, the Workflow itself is clear. It sounds like something that can be executed and reviewed, not a vague aspiration. A Workflow can be large or small; as long as it has a next step, it can later be composed with others into a bigger loop.

These three terms are easy to confuse, so it is worth separating them clearly:

• Agent is the executor. Given an input, it continuously gathers context, finds the right tools, and keeps acting until the goal is reached. The most important work with an Agent is alignment — agreeing on what "done well" means.
• Skill is a single-point capability distilled from doing a task. Its defining property is that it is portable and reusable. A prompt library, an image-generation routine, a scoring rule — anything reusable for a specific scenario can be a Skill.
• Workflow is an executable business loop: a knowledge set that defines a whole process, including the goals, steps, the Skills involved, the review points, and the output.

The relationship is clean. Both Skills and Workflows are reusable resources that an Agent — or a person — can call. A Workflow operates at the business level; a Skill operates at the single-capability level. Skills decide what AI can do. A Workflow decides how a business runs.

Any team that wants a mature Workflow should complete at least these five steps, in order:

1. Business value. Before discussing what you do, discuss why it matters. The goal is never to "kill a task" — it is to achieve a purpose. Outreach exists to win clients; a filing process exists to raise compliance accuracy and improve the client's wait time and experience. Settle the value first, because everything downstream is organized around it.
2. Current process knowledge. Map how people do the work today. But hold this loosely: it is context, not a script. The point is not to make the Agent copy human steps one by one. It is to give it background so it can answer a better question — "people used to do it this way; how should it be done now?"
3. Data flow and states. Switch perspective. Behind every process is the movement of information, resources, and data. Ask what data the loop accumulates, how it transforms, and which data are the real assets. This sounds technical, but it is not — it is just asking what is actually flowing.
4. AI redesign and Skill design. Bring the process and the data to the AI and let it redesign the execution and split out the Skills it needs. Once the process and data are clear, Skill design meets almost no resistance, because each Skill has a clean input and output.
5. Metrics and success. Close the loop. Define what good looks like, so people can watch indicators rather than steps. This is what makes the whole thing a loop instead of a one-off run.

The thread running through all five: start from value, let AI redesign the middle, and leave the human watching metrics and approving at key nodes.

The data view deserves its own emphasis, because it is the step most people skip — and it is the one that decides whether the AI redesign can actually land.

When you only watch the process, you are forever chasing progress: which client reached which step, what happened next. When you watch the data, the picture changes. If you can build a small set of tables that capture the loop's state, you no longer have to follow every step. You glance at the tables and you understand the whole system at once.

So for each Workflow, ask: which data do we want to grow, and which data must be accurate? Those are your assets. Good data modeling turns a hard-to-track process into a few legible states — and gives every downstream Skill a precise place to read from and write to.

Take a growth outreach loop as a worked example. Its business goal is to expand the client base. The human process, roughly, is: discover market signals (who is advertising, where), qualify which leads are worth pursuing, find the right contact and their email, send a cold email, and follow up on replies.

Seen through the data view, four kinds of data flow through this loop:

• Signals — raw leads: an ad, a conference touch, a call, an email.
• Companies — qualified accounts, each with a score, a stage, and a state.
• Contacts — the verified people and emails behind each company.
• Outreach — the record of every message and its current status.

Once the data is clear, the AI redesign becomes a set of time-triggered routines: pull active ad signals in the early morning, qualify the new signals, find contacts for the high- and medium-value accounts, draft cold emails, check for replies and update status, and produce a morning brief. Approved drafts are sent later in the day.

And the human? Reading the morning brief, verifying that an uncertain contact is real, approving a draft before it sends, and taking over once a lead converts — quoting, handing off, and closing. How signals are scored or scraped no longer needs the operator's attention; that is a Skill detail to be tuned separately.

A Workflow becomes "runnable" once the first four steps are done. But runnable is not the same as good — you still do not know how well it runs. That is what the fifth step closes.

Set explicit metrics and success criteria. For an outreach loop, that might be: surface at least a set number of new signals per week, qualify a minimum number of new accounts, keep follow-ups at zero misses, and stay under a latency target. Each Workflow gets its own.

The shift this enables is the whole point. People stop staring at the process and start watching the indicators. When the human no longer tracks every step but watches the metrics — and steps in only at the review nodes — efficiency rises sharply. This is also why value comes first: the end state is an AI-redesigned loop where the person watches numbers and approves at key points.

• Building Skills before the Workflow is clear. Without value, process, and data settled, Skill development has no clean input or output to aim at.
• Forcing AI to follow the old process. Asking "how do I make the old platform do this" is the wrong question. Ask instead what business problem the step really solves — the method may change, but the flow of value does not. AI can often find a newer path than the one inherited from years ago.
• No central data and state management. Without a single legible source of truth, teams fall back to person-to-person updates and lose the loop.

Mapping the current process is genuinely tedious — modern organizations are so specialized that often only a lead understands the full upstream and downstream. That work is worth doing, but it is background knowledge, not the destination.

1. Start from value, not steps. Open every Workflow discussion with what it is for and why it matters.
2. Map the current process in detail — then treat it as context, not as a script to copy.
3. Model the data flow. Name the few tables that capture the loop's state, and decide which data should grow and which must be accurate.
4. Let AI redesign the middle. Hand over the process and data, ask the open question, and let the Agent split out the Skills.
5. Keep human review nodes. Anything that sends, publishes, submits, or changes a critical state needs a person to confirm.
6. Define metrics and watch them. Close the loop with success criteria, then manage by indicators instead of steps.

A useful reframe is "loop engineering": the thing worth designing is not just a single capability, but the business loop itself — where the human sits, where the value is, how it is measured, how people and AI interact. Skills let AI do useful work. A Workflow lets AI operate a business loop with context, checkpoints, and metrics. The Agent is what uses both to reach the goal.