How to Use AI: The Four Modes, the Cost Math, and the Gate
If you want to know how to use AI in a way that survives contact with real work, the answer is not a tool list. It is picking the right mode for the job. There are four: ask, draft, transform, and delegate. They differ in how much you hand over and how much you have to check, and almost everyone starts in the weakest one. Below is the whole method, with the cost arithmetic and the one gate that matters.
TL;DR
- Four modes, in ascending order of what you hand over: ask, draft, transform, delegate. Transformation is where beginners get the fastest real return, because your own material is the source of truth and the model has nothing to invent.
- Adoption is not the same as use. The 2026 Stanford AI Index puts generative AI at 53 percent population adoption within three years, while the US Census Bureau measured only 19.8 percent of US businesses actually using it as of May 3, 2026.
- Cost is almost never the blocker. At Anthropic's published rates, 200 emails a day costs about 31 cents on Claude Haiku 4.5 and about $1.55 on Claude Opus 5. Same job, five times the price.
- Two questions decide what you delegate: can I verify it, and can I undo it. Everything irreversible gets a human gate regardless of how good the model has been.
- The repetition rule: the third time you paste the same instructions into a chat window, that prompt is a specification and belongs in code.
The four modes, and why the order matters
Most advice about how to use AI collapses into "try this tool." That skips the actual decision, which is how much of the job you are handing over. Each mode below hands over more and demands a different kind of checking.
Ask is question answering. You supply a question, the model supplies an answer from what it learned in training. This is the mode everyone starts in and the one with the worst risk profile, because you are asking for facts the model may not have and cannot look up, and a confident wrong answer looks exactly like a right one.
Draft is generation. You supply intent and context, the model supplies a first version. Useful, and the failure mode is mild: bad drafts are obvious.
Transform is where the real leverage sits. You supply the source material and ask for a different shape of it. Summaries, extractions, reformatting, classification, translation between formats. The model is not recalling anything, it is restructuring what you gave it, so the hallucination surface shrinks and verification is fast because the input is right there.
Delegate is handing over a goal plus the tools to pursue it. This is the agent mode, and it is genuinely different in kind: you are no longer checking one output, you are accountable for a sequence of actions you did not individually approve.
Start with transform, not ask
Here is the single change that moves a beginner furthest in an afternoon: stop typing questions and start pasting your own material.
Take a meeting transcript and ask for the decisions and their owners. Take forty support tickets and ask for a table of category, urgency, and whether a human is needed. Take a messy spreadsheet export and ask for it as clean rows with the units normalized. In every case the source of truth is in the request, which means the model is restructuring rather than recalling, and you can check the result against the input in seconds.
This is also why the adoption numbers look so strange. The 2026 Stanford AI Index reports generative AI reaching 53 percent population adoption within three years, faster than the personal computer or the internet, and estimates $172 billion in annual US consumer surplus by early 2026. Meanwhile the Census Bureau's Business Trends and Outlook Survey measured business use at 19.8 percent as of May 3, 2026, rising to 37 percent at firms with 250 or more employees and 39.7 percent in the Information sector, but sitting under 20 percent for firms with fewer than 20 employees. Lots of people have tried it. Far fewer have moved past asking it things.
Write the request so it works the first time
Four things separate a request that works from one that produces mush. None of them are tricks.
Give it the material. The most common failure is asking about something the model cannot see. Paste the document. Attach the data. If the answer depends on your Q3 numbers, your Q3 numbers have to be in the request.
Name the output shape. "Summarize this" gets you a paragraph. "Give me five bullets, each naming a decision and its owner, and mark any decision with no owner" gets you something you can act on. Asking for JSON with named fields turns the output into something ordinary code can route and count.
Say what to do when it does not know. Models fill gaps by default. Add the instruction explicitly: if the answer is not in the material provided, write NOT FOUND rather than guessing. This one line removes a large share of fabricated detail.
Show one example. A single worked example of input and the output you would accept does more than three paragraphs of description. If you want more on this, how to write AI prompts covers the full anatomy.
Pick the model by the job, and know the arithmetic
Model choice is a cost decision far more often than a capability decision, and the gap is wide enough to matter.
Using Anthropic's published pricing, Claude Haiku 4.5 bills $1 per million input tokens and $5 per million output. Claude Sonnet 5 is $2 and $10. Claude Opus 5 is $5 and $25.
Run a real job through that. Say you classify 200 customer emails a day, roughly 800 input tokens and 150 output tokens each. That is 160,000 input and 30,000 output tokens daily.
- On Haiku 4.5: 0.16 million input at $1 is $0.16, plus 0.03 million output at $5 is $0.15. About 31 cents a day, under $10 a month.
- On Opus 5: $0.80 plus $0.75. About $1.55 a day, roughly $47 a month.
Five times the cost for identical work. If the cheap model handles it, the difference is pure waste. Two more levers: cache hits bill at a tenth of the base input price, so repeated context gets cheap fast, and the Batch API takes 50 percent off both sides for anything that does not need an answer this second. That drops the Haiku job under $5 a month.
The practical rule is to start on the cheapest model that could plausibly work, build your test cases, and only move up when you can point at the specific cases that fail. How to choose an AI model goes deeper on where each tier earns its price.
The gate: what you can verify, what you can undo
Delegation is where using AI stops being a productivity tip and starts being an engineering decision. Two questions decide it.
Can I verify the output? Not in principle, in practice, in less time than doing the task myself. If checking is fast, delegation is a good trade even with a model that is sometimes wrong, because you will catch it. If checking costs as much as doing, you have moved the work rather than removed it.
Can I undo it? This is the harder gate and it is not about model quality at all. Sending a message to a customer, moving money, deleting records, and publishing are one-way doors. Every one-way door gets a human approval step, permanently, regardless of how well the thing has performed so far.
There is a common worry that delegating work stops you learning it. Anthropic's Economic Index report, published June 26, 2026, found the opposite pattern in its data: heavier delegators reported learning at the same rate as everyone else, and users with higher automation shares were more optimistic about their pay and job security rather than less. That is one dataset from one product's users, so hold it loosely. The mechanism that matters is verification. You learn from the parts you actually read.
Your first week, concretely
Day one, pick one task you did at least five times last week. Not the most important one, the most repetitive one.
Day two, do that task in transform mode. Paste the real material, name the output shape, add the do-not-guess line. Keep the prompt in a text file rather than the chat history.
Day three, save ten real examples with the output you would have accepted. This is your test set, and it is the difference between an opinion about whether the thing works and knowing.
Day four, run the prompt against all ten and count the failures. Fix the prompt, not the examples.
Day five, apply the gate. If the task is reversible and fast to check, wire it into something that runs without you. If it is not, keep yourself in the loop and accept that the win is speed rather than autonomy. For the tools worth having in place before any of this, the AI daily driver stack is the setup I actually run.
If you want the conceptual foundation underneath all of this, introduction to AI for builders covers the ten ideas that change what you ship.
The bottom line
Knowing how to use AI is not about which tool you opened. It is about matching the mode to the job and being honest about what you can check. Transformation gives beginners the fastest honest win because the answer is already in the room. Cost is rarely the constraint, and when it is, it is usually because someone reached for the expensive model out of habit. Delegation is worth it exactly when verification is cheap and the action is reversible, and never worth it when it is not. The people getting real value are not using better tools than you. They picked one repetitive task, wrote ten test cases, and stopped guessing.
Pick one repetitive task this week and run it through the four modes. If you want the tool setup that makes it painless, start with the AI daily driver stack, and subscribe to the newsletter for the builds and the arithmetic behind them every week.
What is the fastest way to start using AI if I have never really used it?
Stop asking it questions and start giving it your own material. The fastest useful move is transformation: paste something you already have, a meeting transcript, a long email thread, a spreadsheet export, a support inbox, and ask for a specific shape back. Turn this transcript into five decisions and their owners. Turn these forty tickets into a table of category, urgency, and whether it needs a human. You get value in the first ten minutes because the source of truth is already in front of the model, so it has nothing to invent, and you can check the output against the input in seconds. Question-answering is the mode most beginners start in and it is the one where hallucination hurts most.
How much does it actually cost to use AI for real work?
Far less than people assume, and the model is rarely the expensive part. Using Anthropic's published rates, Claude Haiku 4.5 bills one dollar per million input tokens and five dollars per million output. A job that reads 200 customer emails a day at roughly 800 input and 150 output tokens each burns 160,000 input and 30,000 output tokens, which is about 31 cents a day, or under ten dollars a month. The same job on Claude Opus 5 at five and twenty five dollars runs about a dollar fifty five a day, five times more for identical work. Two levers cut it further: cache hits bill at a tenth of the input price, and the Batch API takes fifty percent off both sides for anything that can wait.
What should I never hand over to AI?
Anything you cannot verify and cannot undo. Those two tests do more work than any policy document. If the output is checkable in less time than doing the task yourself, delegation is a good trade even when the model is sometimes wrong, because you will catch it. If checking costs as much as doing, you have not saved anything, you have moved the work and added risk. The undo test is the harder gate: sending a message to a customer, moving money, deleting records, and publishing anything are all one-way doors. Put a human approval step in front of every one-way door, no matter how reliable the model has been so far.
Is it better to use AI through a chat window or to build something with the API?
Chat is for exploration and one-off work. The API is for anything you will do more than about twenty times. The tell is repetition: the moment you notice yourself pasting the same instructions into a chat window for the third day running, that prompt has become a specification and belongs in code where it can be versioned, tested, and run without you. Chat also quietly hides your costs and your failure rate, because nothing is logged and nothing is counted. Moving a repeated task into a small script does not require a framework or an agent. It usually takes one file, one prompt, and a loop over your inputs.
Does using AI to do work for me stop me from learning?
The available evidence says no, which surprised a lot of people. Anthropic's Economic Index report published June 26, 2026 found that heavier delegators report learning at the same rate as everyone else, and that users with higher automation shares were more optimistic about their pay, job security, and employability rather than less. That is one dataset drawn from one product's users, so treat it as a signal rather than a settled finding. The practical version is simpler: you learn from the parts you verify. If you delegate a task and read the output critically before accepting it, you are still engaging with the work. If you accept output unread, the delegation is where the learning stops.
OpusJake is Jake Schincariol's operating system for building with AI: agents, workflows, prompts, and the free resources behind them. Get the next move every week.