Operating Safely and Cheaply

Two mistakes end AI operators' good runs faster than anything else: pasting something they can't take back, and not noticing the bill until it's already large. Neither mistake requires bad intentions — both come from not having a simple mental model for what crosses the line and what the meter is actually counting. This lesson gives you both.

What Never Crosses the Boundary

The rule is simpler than most people make it: if you couldn't undo it being public, don't paste it. That single test covers almost everything that actually matters.

Credentials and API keys. A password or token typed into a chat window is a password or token that has left your control. It doesn't matter how careful the vendor is — the safe move is that it never gets typed there in the first place.

Customer or user PII. Names, emails, addresses, anything tied to a real, identifiable person who isn't you. You don't have the right to decide, on their behalf, that this information should leave your systems and land in a third party's logs.

Confidential or NDA'd documents. Anything you don't personally have the authority to disclose. "It seemed fine" is not a defense after the fact.

The reasoning that matters here isn't really about whether a model trains on what you send it — vendors have different policies, and that argument gets litigated online endlessly. The reasoning that matters is more basic: the instant you hit send, that data is logged and stored somewhere you don't control, and there is no version of "actually, take it back" available afterward. Treat the boundary as the last moment before you hit send, not the chat window itself.

What's actually fine to paste: public information, your own already-published writing, synthetic or made-up examples, and general questions with no sensitive specifics attached. That's most of what people actually need AI for day to day — the boundary rule isn't a reason to be scared of the tool, it's a filter you run before you paste, not a wall around the whole practice.

Tokens Are Money

Every call to a model costs something, and the amount is driven by a small, learnable set of factors — not mystery, not "AI is expensive," just arithmetic.

Context size. Everything you paste, attach, or include in a conversation's history gets billed on every single call in that conversation, not just once. A long document pasted at the start of a chat gets re-billed on every follow-up message.

Output length. A longer response costs more than a shorter one. Asking for exactly what you need, instead of "give me everything you know about this," is a real cost lever, not just a style preference.

Model tier. Frontier-tier calls cost meaningfully more per token than fast-and-cheap-tier calls — often several times more. This is the single biggest lever most people never touch.

Retries. A failed attempt that gets automatically or manually re-run bills again, on top of the first attempt. A workflow that silently retries three times on failure is paying for four attempts, not one.

To put actual numbers under all four levers, without pretending any vendor's price sheet stays fixed: a short chat costs a small fraction of a cent, a long, document-heavy session runs a few cents, and it's heavy daily agent use — many calls, large context, run automatically — where real money starts to add up.

Stretch that across a month and the four levers set the total. Light chat use lands in the smallest bracket — a subscription running single-digit dollars, or pennies per call through an API. A document-heavy workflow moves higher, since context gets re-billed on every call, not just the first. Unattended agent runs carry the risk: many calls, large context, running automatically, sometimes retrying on failure unwatched. The bill isn't decided by which model you used — it's decided by how those four levers stacked up, and the expensive failure mode is automation left running, not a person chatting.

The Cheap-First Habit

The habit that controls all four of the above, before you've even sent your first message: default to the fast-and-cheap tier, and only reach for the frontier tier when the job genuinely needs it. Pair that with trimming your context down to what the task actually requires — not the whole document when a paragraph would do — and you've addressed the two biggest cost levers before the model has run a single token. Caching (reusing a previous result instead of re-asking) is the third lever worth building into any repeated workflow: if you're going to ask the same or a near-identical question more than once, that's a sign the answer should be stored, not re-purchased.

When AI Is the Wrong Tool, Full Stop

Not every job belongs in a model at all, and the safest habit is recognizing this before you start, not after a bad answer costs you something. AI is the wrong tool when:

• The answer must be ground-truth verified, not estimated — an exact account balance, a legal filing deadline, a live measurement. A model's best guess is not a substitute for checking the authoritative source.
• The action is real-time and irreversible — authorizing a payment, sending an email that can't be recalled, deleting production data. Autonomy is for drafting the action, not for executing it unsupervised when the cost of a mistake is high.
• A wrong answer carries legal or safety consequences you're not positioned to absorb. Get it verified by whatever the actual authoritative process is, and use AI to draft or research around that process — not to replace it.

None of this means avoid AI for these domains entirely. It means AI drafts, checks, and prepares; the authoritative source or process still makes the final call.

Bottom Line

Run everything through one test before you paste it — could you undo this being public? — and credentials, PII, and confidential documents never make it into a model. Tokens are money, and four things drive the bill: context size, output length, model tier, and retries; controlling the first two before you even hit send is the highest-leverage habit you'll build. Default to the cheap tier and escalate deliberately. And recognize the jobs — ground-truth facts, irreversible real-time actions, high-stakes legal or safety calls — where AI drafts and a verified authoritative source still decides. Next: how the academy itself is built, and exactly where to go from here.