Field Notes
I keep seeing people treat AI like the product is the model.
Which model is best? Which one writes better? Which one codes faster? Which one is cheaper? Which one should my team use?
Those are useful questions. They are just not the first question.
The model is the engine. The workflow is the machine.
That distinction matters because engines keep changing. One week Claude feels unbeatable. The next week GPT catches up. Then an open-source model shows up that is good enough for 80% of the work at a fraction of the cost. I wrote about that in Don't Marry Claude: if your whole strategy depends on one model staying ahead forever, the leaderboard owns you.
But if the workflow is your strategy, the model becomes swappable.
I had this realization while thinking through the AI infrastructure I actually need for my own business. The version I need to run the company without everything living in my head.
At first, it is tempting to think the asset is the skill. A prompt library. A set of instructions. A few reusable automations.
A skill tells the AI how to behave.
A workflow tells the system how the work gets done.
That means the prompt, the tools, the data it is allowed to read, the approvals it needs before touching anything sensitive, the handoff points, the model choice, the verification step, and the output format.
That full bundle is the asset.
I have been speaking about this with some of my clients recently: general intelligence is not always the best intelligence. The best intelligence fits the problem.
If you run a law firm, you probably do not care that a model got better at coding. You care whether it got better at law, at understanding previous cases, and at reasoning through how those cases apply to the matter in front of you.
That is a scoped use case. The rest of what the expensive model can do may be impressive, but it is not always useful. If the job is writing a basic email, you do not need the strongest model in the stack. If the job is weighing precedent against a current case, you may need a higher level of reasoning for that specific problem.
That is the point: route intelligence by the work, not by the hype around the model.
Most businesses do not need a god model that can answer every question in the world. They need specified intelligence for how their business actually works.
I saw this play out again in a recent working session with my brother. We were talking through how agents connect to real systems, and the simple metaphor that finally landed was a house.
An API key is the front door. Endpoints are the rooms. A CLI wrapper is a shortcut that says, go to this exact room and bring back this exact thing. An MCP server is more like a scoped toolbox. It gives the agent specific tools it is allowed to use without handing it the whole house.
Quick pause if that language is new: an AI agent is software that can pursue a goal and use tools to complete tasks. I gave the espresso-machine version in The Agentic Internet. This post is the inside-the-business version.
During that same session, I showed him a presentation app with an MCP server. I messaged an agent from Telegram and asked it to find empty canvas space, create a thank-you slide, send me a screenshot, and export the deck as a PDF.
It sounds like a toy example until you look closer.
The agent did not win because it was the smartest brain in the room. It worked because the workflow was clear: tool access, permission boundary, task, output, review.
That is why I think workflows are the new primitive.
The work will not be organized around people clicking through software one task at a time. It will be organized around workflows that agents can run, humans can review, and teams can improve.
Playbook
The practical version is simple.
Pick one repeatable workflow in your business and write it down like you were handing it to a smart employee on day one.
Do not start with job titles. Start with the loop that keeps happening.
A simple workflow has seven parts:
Trigger. What starts the work?
A new email. A booked call. A signed contract. A nightly cron. A customer complaint. A voice note.
Context. What does the system need to know before acting?
Customer history. Prior decisions. Project status. Pricing rules. Brand voice. Sensitive constraints.
Tools. What can the workflow use?
Email. Calendar. CRM. Database. Browser. Stripe. Ghost. OwnerRez. The systems actually part of the job.
Rules. What is allowed, and what requires approval?
Draft but do not send. Recommend but do not change pricing. Read but do not delete. Escalate money, legal, privacy, or reputation issues.
Model route. What level of intelligence does each step need?
Cheap model for extraction. Script for deterministic checks. Local model for private data. Strong model for judgment. Expensive model for final review.
Output. What should come back?
A draft email. A meeting brief. A pricing recommendation. A clean checklist. A posted article. A decision memo with receipts.
Verification. How do we know it worked?
Read back the draft. Check the sent item. Confirm the file exists. Compare the numbers. Test one item before batching. Show the receipt.
That is the difference between a prompt and a workflow.
A prompt says, “Help me triage my inbox.”
A workflow says, “Every two hours, scan the inbox, suppress resolved threads, identify urgent items, draft safe replies, escalate money or legal issues, and show me the evidence before anything leaves the building.”
One is a request.
The other is infrastructure.
If the workflow is clear enough, you can route different pieces to different models.
The intelligence is not just in the brain.
It is in the harness around the brain.
That is the piece most companies are missing.
They are buying access to intelligence, but they are not capturing how their work actually works.
The companies that do capture it will build workflow libraries that compound. Every client call improves the sales workflow. Every mistake improves the verification workflow.
Over time, that library becomes more valuable than the tool running it.
Orientation
This connects the first three posts.
In Welcome to the New World, I wrote about the layer of work underneath us changing.
In Don't Marry Claude, the point was sovereignty. Own enough of your system that you are not trapped every time the model race changes.
In The Agentic Internet, the outside interface changed. Customers and software will increasingly interact through agents.
Workflows are the inside layer.
If workflows are the new primitive, the next uncomfortable truth is that most operators are not blocked by the AI.
They are blocked by themselves.
The judgment calls, the exceptions, the weird edge cases, the “I just know what to do here” moments.
That is the real bottleneck.
The companies that pull ahead will turn tribal knowledge into executable workflows faster than everyone else.
Next time, we are getting into You Are the Bottleneck. Once execution gets cheaper, the bottleneck moves upstream to judgment, taste, context, and the ability to define the work clearly enough that someone else, human or agent, can run it.
Drop a comment below and tell me: what workflow in your business still only exists in your head?
I read every one.
— Brian