Key Takeaways
- Telling AI “how” forces it to mimic your workflow, which is often the slowest and most expensive path.
- AI has native tools (APIs, data connectors, direct system access) that are faster, more reliable, and far more token-efficient than replaying clicks.
- Specify the outcome, the constraints, and what good looks like. Let the AI choose the path.
- When the path is the product (regulated, audited, safety-critical), the how still matters. Specify it then.
The finance manager had the process memorized. Every Monday morning: open the CRM, click through six screens, export the pipeline report, open the spreadsheet, filter for at-risk accounts, copy the results into an email to the sales leaders. Forty-seven clicks, twelve minutes, zero thought required.
When her company rolled out an AI assistant, she knew exactly what to do. She recorded her workflow and asked the AI to replay it every Monday.
It worked. Slowly. Each run burned thousands of tokens to move a pointer across screens, and the whole thing broke the first time the CRM updated its interface.
She had taught the AI her habits. The AI had obeyed perfectly. That was the problem.
Expertise Builds Patterns
Expertise is pattern recognition earned the hard way. Years of repetition compress into instinct. The expert stops thinking about the steps because the steps have become reflex.
Psychologists have a name for the trap inside that gift. The Einstellung effect: once the mind finds a method that works, it keeps reaching for it even when a better one is available. The better the expert, the stronger the pattern, and the harder it is to see around.
This is exactly what happens when experts start directing AI. They do not describe the problem. They dictate their process. They hand the AI a map of their own habits and call it instructions.
The AI complies. It replays the forty-seven clicks. It never asks whether there was a shorter road, because nobody told it to look for one.
Mimicry Is the Most Expensive Path
Modern AI can watch a screen, move a pointer, and click through software exactly the way a person would. That capability feels like magic the first time you see it.
It is also, in most cases, the worst way to use the tool.
Every click is a round trip. Every screen must be interpreted, every button located, every pause endured. A task that takes a database a fraction of a second becomes a minutes-long puppet show, billed by the token. And the moment a vendor redesigns a page, the recording breaks.
None of this is the AI’s fault. It was given a human’s workflow and executed it faithfully. Mimicry is reliable at one thing: reproducing your inefficiencies at machine cost.
The goal is not to make AI a faster version of you. It is to let AI be something you are not.
Free the AI to Choose Its Tools
Here is what the finance manager never tried.
Instead of the recording, she could have written one paragraph: “Every Monday at 8 a.m., send the sales leaders the at-risk accounts, the deal size, and one line on why each is at risk. Pull from the live CRM. Flag anything over $100K separately.”
Given that assignment, the AI would never touch the interface. It would query the CRM directly through a data connector, assemble the answer in seconds, and use a fraction of the tokens. No clicks. No screens. Nothing to break when the layout changes.
AI systems now connect directly to the tools organizations already run: CRMs, databases, document stores, ticketing systems. These connections are faster than screen replay, more reliable than screen replay, and dramatically cheaper than screen replay. They are also invisible to anyone whose mental model of “using software” ends at the mouse.
Your workflow was designed for your hands. The AI has different hands.
Specify the What and the Guardrails
None of this means giving AI vague instructions and hoping for the best. “Handle the report” is not an assignment. It is a wish.
A good assignment has three parts. The outcome: what should exist when the work is done. The constraints: what data it may use, what it must never touch, what rules it must follow. And the definition of good: how the result will be judged, what gets flagged, what requires a human decision.
That is the entire job of the person directing the AI. Not the steps. The destination, the boundaries, and the standard.
This is where experience becomes an asset again. Nobody knows the constraints of the work better than the expert: which sources are trustworthy, which exceptions matter, where the judgment calls hide. Put that knowledge into the assignment. Leave the pathfinding to the machine.
Tell it what. Tell it what matters. Then get out of its way.
Sometimes the How Is the Task
There is an exception, and it matters.
Sometimes the path is the product. A regulated filing must follow prescribed steps. An audited control must execute in a defined sequence. A safety-critical procedure must be performed exactly as written. In those cases, the how is not a habit. It is the requirement. Specify it completely.
The question to ask is simple: if the AI found a different route to the same result, would anything be lost? If the answer is yes (compliance, evidence, safety), dictate the path. If the answer is no, dictate the destination.
Most knowledge work falls in the second category. Most experts still assign it like the first.
The Shift
The future of working with AI is not about learning its features. It is about unlearning the reflex to hand it your playbook.
Your experience remains the asset. It belongs in the assignment, not in the steps.