Field Notes
Stop buying genius for jobs that need accuracy.
That sounds strange until you look at how a lot of people use AI. They find the strongest model they can access, point every task at it, and hope the bill does not start looking like a utilities company.
I get the instinct. Stronger feels safer.
But once you have done the earlier work in Arc 3: AI Engineering, the problem changes.
You are no longer staring at one vague blob called "use AI."
You have lanes.
The prompt defines the job. The context defines the room. The harness defines the tools, permissions, memory, stop signs, and receipts. The loop defines how the work repeats, improves, and gets checked.
At that point, model engineering becomes a routing problem.
You look at each lane and ask a blunt question:
How much intelligence does this step actually need?
One lane may only need classification or date extraction.
Another may need research, math, judgment, taste, or a human to say, "Yes, send it."
Those lanes should not automatically use the same brain.
I wrote about this earlier in Specified Intelligence. If you need to tighten one screw, you do not need the most powerful drill in the store.
The job defines the intelligence.
Model engineering applies that idea inside a real workflow.
The target is the lowest sufficient intelligence for each step.
Lowest does not mean cheap for the sake of cheap. It means the least expensive, fastest, most reliable model that can produce the exact result you need with the risk level you can accept.
If the lane is simple and easy to verify, do not burn frontier reasoning on it.
If the lane touches money, customers, legal language, strategy, reputation, complex math, or a hard research question, pay for the better brain.
Cheap mistakes get expensive when they are attached to the wrong workflow.
The practical version is simple.
Classify the support ticket with a workhorse model.
Pull the customer record with a tool-connected agent.
Use a stronger model to reason through the edge case.
Use another model to review the draft against the policy.
Put the human approval gate in front of refunds, public claims, pricing moves, account access, or anything that can create real damage.
That is how a workflow becomes both smarter and cheaper.
The model race still matters. I watch the leaderboards too.
Artificial Analysis gives you a map: quality, speed, latency, price, and context window. OpenRouter gives you a way to route across models without rebuilding the whole system every time the market changes.
Treat those tools like scouting reports.
They can tell you who looks fast, strong, expensive, cheap, or balanced.
They cannot tell you which model you should trust with your refund workflow, your investor memo, your guest message, or your pricing recommendation until you test the model against that work.
Build the lanes first.
Test models inside the lanes.
Route the work to the lowest level of intelligence that reliably gets the result.
"Best model" is too broad to be useful.
Best for what: long-context research, math, tool calls, writing in your voice, returning clean code or structured data, or running cheaply 20,000 times a month?
Those are different questions.
Different questions should create different routes.
There is a catch.
Model routing only works when you own the AI layer.
If your entire system lives inside one subscription tool, you do not really have routes.
You have one brain.
Claude for everything.
ChatGPT for everything.
Whatever that platform gives you this month for everything.
That can be fine for casual use. It is weak architecture for a real operating system.
If a platform removes a lower-tier model, changes the price, changes the context window, weakens tool access, or falls behind another model, your workflow has to absorb that decision.
You cannot automatically move the review lane to Kimi, the extraction lane to GLM, or the research lane to a new GPT model because it is now better for the job.
You are locked into the house brain.
That is why the open AI stack matters.
Open source is an operating advantage, not a purity argument.
Open means your context, memory, files, tools, evaluation rules, and workflow body are not trapped inside one vendor's app.
When you own that layer, OpenRouter becomes useful. Artificial Analysis becomes useful. New model releases become opportunities instead of migrations.
Some lanes may run better and cheaper through models like GLM, Kimi, DeepSeek, or the next strong workhorse that appears.
Depending on volume, that can beat the $20 subscription habit while giving you more control.
Keep using Claude and ChatGPT where they win.
Stop building the business as if one subscription brain is the whole AI strategy.
Training comes later.
People hear "train a model on your data" and picture a custom company brain that finally understands everything.
Maybe, eventually.
But training before the workflow is clear teaches the machine your mess.
Documentation, clean examples, pass/fail criteria, and a real definition of done come first.
Training becomes useful when the job is narrow enough to describe, repeated enough to matter, and documented enough that the model has something clean to learn from.
The value comes from the same place every other layer comes from:
the work has been made visible.
Playbook
Here is how I would use this.
Pick one AI-assisted workflow.
Do not start with the model.
Start with the lanes.
1. List the steps
Write the workflow like a checklist.
For customer support, that might be:
- classify the message;
- find the account;
- summarize the issue;
- draft the response;
- check the response against policy;
- escalate refunds, legal language, angry customers, or account access issues.
Now you have lanes.
2. Score each lane
For each lane, ask what kind of intelligence it needs.
- Reasoning: does it need to connect several ideas?
- Judgment: can the answer be technically correct and still wrong for the business?
- Math: does bad calculation create real cost?
- Research: does it need to search, compare, or source claims?
- Taste: does voice, positioning, or nuance matter?
- Consequence: what happens if it fails?
The cost of being wrong should set the intelligence budget.
3. Pick the lowest sufficient brain
For simple, high-volume, easy-to-check work, use a cheaper workhorse.
For judgment-heavy work, use a stronger operator model.
For strategy, complex math, research, public claims, or work that is hard to verify, bring in the best reasoning you can justify.
The standard is:
What is the cheapest model I would trust with the cost of being wrong here?
4. Add the review route
The model doing the work does not always need to be the model reviewing the work.
For important lanes, separate them.
One model drafts.
Another model reviews.
The harness checks required fields, citations, policy rules, or source receipts.
The human approves the parts that still need ownership.
That is how you avoid confusing autonomy with accountability.
5. Test against real work
Leaderboards help you narrow the field.
Your workflow is the test.
Run the same task through two or three routes.
Check accuracy, speed, cost, consistency, formatting, source use, and how often a human has to clean it up.
The cheapest model is only cheaper if the cleanup cost stays low.
6. Keep the route swappable
Do not marry the model.
The market is moving too fast.
Use routers, clean interfaces, and plain names for each lane so the workflow can swap models without rebuilding the whole thing.
Keep the company brain and workflow body outside the model app.
That is the difference between using AI and owning the AI layer.
7. Train only after the lane is real
Save approved outputs.
Save rejected outputs.
Write down why they passed or failed.
Capture edge cases.
Build a small evaluation set.
Then talk about training.
Clean training data is usually operational discipline wearing a technical hat.
Annoying, but true.
Orientation
Operators need to learn how to direct work instead of collecting tools.
Prompt engineering makes the ask clear.
Context engineering gives the model the room.
Harness engineering gives it tools, rules, memory, permissions, and receipts.
Loop engineering makes the work repeatable.
Model engineering decides which brain belongs in each lane.
Stack those together and you are starting to build an operating system.
That is what I am teaching live this Thursday inside the AI Strategy Academy webinar. We are going deeper on AI workflows, model routing, and the path into the next Agent Bootcamp cohort. Save your seat here if you want the live version.
For this week, pick one workflow where every AI step still goes to the same model.
Break it into lanes.
Mark the steps that truly need reasoning, judgment, math, research, taste, or human approval.
Then route the rest to the lowest sufficient intelligence you would trust.
That is where the cost savings start.
More important, that is where the system starts making sense.
Brian