An AI employee audited my AI bill and roasted me harder than my followers ever have.
I was burning $4,700 a year on tools I barely opened.
My dumbest habit was asking ChatGPT what Grok thought, and paying for both.
This is the system that fixed it 👇 https://x.com/i/article/2120534427755302912
How I Cut My AI Bill by Mass-Unsubscribing and Hiring One AI Employee (From Scratch)
Most AI power users pay for three or four model subscriptions to do one person's job. I did too, until I tracked a full workday and saw that I spent more time choosing a model than using one. So I hired one AI employee and cancelled most of the rest. The system that replaced the habit fits on one screen.
I'm @leopardracer, and my account covers local AI inference, Claude workflows and AI productivity systems.
How I stopped picking models, hired Viktor as an AI employee in Slack, and built a system around gears instead of subscriptions. It comes with the formulas, the decision rule, the checkup question and a cheat sheet you can run on your own stack tonight. Viktor does the work, and you keep the decisions.
Every time a new model drops, the timeline says it changes everything. So you subscribe, open another tab, and paste the same prompt into three windows to see which answer wins. Somewhere in there you lose the thread and start over.
I did this for a long time: one model for writing, another for code, one more for the timeline, and a second window for the first model once it forgot what we were doing. I thought that was smart power-user behavior.
Then I tracked a full workday and wrote down every time I hesitated between models: each tab switch, each re-paste, each "let me try this one too." By the evening the tally said I had spent more time choosing an AI than using one, mostly on tabs, re-pasting, and rebuilding context after every switch.
I was managing AI more than I was using it, and the overhead was eating the time it was supposed to save.
This article is the system I built to stop: the mistake, the fix, the formulas, and the one question I asked Viktor that showed me where the waste was.
Part 0. The line item that is not on your statement
Your AI bill has two parts.
The first part is on your card statement. You can see it, cancel it, and complain about it on the timeline.
The second part shows up nowhere. It is the moment before every task when you stop and decide which model gets it.
In a normal workday it goes like this. A Slack message needs a careful reply, so you open a tab and wonder which model to use, since one is good at tone and another is better with context. You pick one, paste the message, get an answer, and close the tab.
Twenty minutes later you need to debug something. The code model is the obvious pick, but the other one still has yesterday's project context, so you open both, paste the same code into each, compare, keep the better answer, and close the tabs.
An hour later a report needs summarizing, and it starts again.
Each time you choose, open, paste, compare and close, and none of those steps is the actual work.
I call it the decision tax.
You do not need my inputs. Plug in your own, and be honest about n_d, because it is always higher than you think. Count it tomorrow with a tally mark on a sticky note every time you hesitate between models. You will be surprised by lunch.
The shape of the formula matters more than your result. The tax grows with every decision you make, so getting faster at deciding barely moves it. The fix is to make fewer decisions.
Part 1. The bill before
This is what I was paying for:
Lined up, the pattern was embarrassing. I paid full price for eight separate AI tools and used each one for a thin slice of my week. ChatGPT stayed open mostly for code completions I could have done elsewhere. Grok I opened to see what the timeline was saying, then left it idle in a tab for hours.
Every AI power user I know keeps at least one "just in case" subscription.
Try this before you read on: open your card statement, list every AI line, and next to each one write the last thing you actually used it for. The lines you cannot fill in are the ones to cancel.
Every major model can write, code, research and reason, so most of those subscriptions pay for a feeling of optimization. "I use the best model for each task" sounds smart, but picking the best one every time often costs more than the quality gap between models on most tasks. The bigger cost is the choosing.
Part 2. Pick the gear, not the model
I hired Viktor.
Viktor is an AI employee. He lives in Slack and Microsoft Teams, connects to 9,800+ tools, and 70,000+ teams have hired him. He takes real jobs on your team, and every action that matters waits for your yes.
What changed how I work was a mental model I borrowed from driving.
When you drive, you pick a gear for the road and leave the engine alone. Viktor works the same way:
With this setup, "which model?" comes up in three places only: once for the team, once for each automation that deserves its own model, and the rare time you type !deep or !fast.
My setup:
- Team mode is Smart, because most of my day is content drafts, research summaries and scheduling, and Smart handles that without burning through credits.
- I picked the model myself for two automations. Content repurposing runs on a specific model because I want a certain writing feel when Viktor turns threads into articles. Competitor tracking runs on a different family because it parses structured data from several sources, and I want the one that is most reliable at that.
- I type !deep to review a brand deal contract, to write a piece that goes out under my name without edits, or to debug a production issue in one of my projects.
- I type !fast for Slack thread summaries, quick translations, table formatting, and pulling a number out of a long document.
One thing people get wrong: Viktor does not pick the best model for each task by himself. I set the gear and he does the work. I like it that way, because the decision gets made once, on purpose, by a person, instead of a hundred times a week by a tired one.
An AI that silently picks models behind the scenes looks good in a demo. In practice you cannot tell why something worked or failed, and you cannot tune it. With Viktor every gear choice is mine, so I know what ran and why, and when the checkup shows where the credits went, I can decide what to change.
Part 3. The test: real work, not demos
Demos and launch videos are made to sell, so I tested on my own work.
I gave Viktor real tasks from my week across research, writing, code and data, and noted which gear each one ran on and what came back.
Tasks where I shifted gear:
Three things stood out:
1. !fast handled more than I expected: summaries, translations, pulling figures, formatting. I had been sending Smart or higher at tasks that did not need it. The Slack summary surprised me most, because the !fast output looked the same as Smart's and arrived faster for fewer credits. The task was simply easy.
1. !deep earned its cost on code and contracts. The async bug would have taken me an hour of print-statement debugging, and Viktor found it on !deep in one pass. The contract review caught clauses I would have read straight past. When being wrong is expensive, the heavier gear pays for itself.
1. Smart was the right gear for content, which I did not expect. I assumed writing needed the heaviest model, but the first draft is scaffolding and my voice comes from the edits. A heavier first draft meant paying more for text I was going to rewrite anyway.
I was overspending most on the tasks that mattered least. The high-volume, low-stakes work burned the most credits because I never questioned the default.
That gave me the rule I now use for every task:
Shift up only when a better answer is worth more than the heavier gear costs.
It sounds obvious, but most people run everything in the highest gear because it feels safer. That is like driving to the corner shop in first gear at full revs.
Applied to normal work, it looks like this:
When I sorted my own week this way, most of it landed in the top half of the table, which means I had been driving everywhere in the highest gear.
Part 4. The checkup: one question that changed everything
The checkup is what sold me. I typed one message in Slack:
"Where are my credits going?"
Viktor came back with where the credits were actually going and a list of fixes he proposed. Nothing changed until I said yes.
Most software is built to sell you more seats and more features.
Viktor told me where I was overspending, proposed the fixes, and waited for my approval. That is what you want from someone on your team, and it runs against how most software makes money.
Paying a senior analyst to staple paper
Every automation has two kinds of steps. Some need judgment, like reading, deciding, writing and summarizing. Others never change: pull the same report, format the same table, post to the same channel at the same time, forever.
Running the second kind through a model every time is like hiring a senior analyst to staple paper all day. They will do it perfectly, and you will still be overpaying.
When Viktor spots steps like that, he can offer to move them into a script, so the model keeps the judgment and the script does the stapling.
Look at (1 - p). The more repetitive the automation, the bigger p gets, and that share of its cost goes away while the output stays the same.
Across a whole team, each automation has three levers: how often it runs (f), which gear it runs in (g), and how much of it is fixed steps (p). Most people only adjust the first one, and the checkup shows all three.
In my case, Viktor flagged formatting steps in my content pipeline that went through a model on every run without ever changing: pull the same RSS feeds, strip the HTML, format a Slack block. He offered to move them into a script and I approved. The summarizing and writing still run through a model because they need judgment, and the mechanical glue no longer does.
Part 5. Where the money went
I put my Usage page next to the old stack.
I cancelled ChatGPT Plus, Google AI Pro, Cursor Pro, Copilot Pro+, Perplexity Pro and Jasper, and downgraded X Premium+ ($40) to regular Premium ($8), which is all I need to publish Articles. That is $200/mo in subscriptions and another ~$250/mo in API credits I had been keeping out of habit. Once Viktor was doing that work in Slack, the separate tabs stopped making sense.
I kept Claude Max ($200/mo) for long personal writing sessions and code work outside of Viktor, where I sit with one window for an hour and want that specific feel. That one is a personal choice, and Viktor handles the rest.
The new stack:
What jumped out on the Usage page was that $658 became $263, and the same week of work cost about 60% less overall once I stopped defaulting to the highest gear. The tasks were the same and I could not tell the output apart. The credits went further because each job ran in the gear it needed, and the 12% I had been burning on Slack summaries at Smart dropped to almost nothing at !fast.
Part 6. The eight traps I fell into (so you do not have to)
1. Subscribing to every launch. Each new model brings a new tab and a new invoice, and you end up collecting models instead of testing them.
1. Using the highest gear by default. It feels safe and costs the most, while most of your week is Slack summaries and table formatting and contract law is rare.
1. Re-pasting context everywhere. Every copy into another window costs the decision tax twice, once to choose and once to rebuild the context.
1. Paying a model to do a script's job. A step that is the same every time does not need intelligence every time.
1. Never asking where the money goes. Most people never look, and one question to Viktor answered it for me.
1. Thinking you need access to every model. What you need is the output. Viktor works with several model families behind one conversation in Slack, which is very different from paying a separate subscription for each.
1. Tuning the prompt when the gear is the problem. I spent hours rewriting prompts to get better output from a lighter model when one !deep would have done it. Prompt work is useful, and using it to avoid the right gear is procrastination.
1. Treating every task as high stakes. A Slack summary is not a legal review, and a table reformat is not a launch email. Most of a workday is low stakes and does not need the heaviest gear, which is what the table in Part 3 shows.
Part 7. When I still pick a model myself
There are still jobs I route by hand:
- Long creative writing sessions where I want one specific model family's voice and I am sitting with it for an hour, iterating, not delegating.
- Experimental code where I want to compare raw outputs side by side before committing to an approach.
The rule from Part 3 covers both. A better answer is worth enough to me there that I want my own hand on the gear, and those jobs are now the exception.
Anyone who says one setup removes every decision forever is selling something. My goal is fewer decisions, each one made on purpose.
Part 8. The math
The whole model fits in one line:
Two terms move G, and most people only look at the first. Cancelling a subscription moves ΔC. Deciding less before each task moves ΔT, and for most people I know that is the bigger term, though they never write it down.
Write it down for your own stack. If G comes out negative, keep what you have and close this tab, no hard feelings.
If it comes out positive, you have been paying a tax on indecision, and you can stop tonight. Set a team mode, give Viktor one job, and ask him where the credits are going.
The cheat sheet (screenshot this)
One real thing from this week
This week I asked Viktor for a competitor breakdown for a brand pitch. He pulled 6 competitors' pricing pages, compared the positioning across all of them, flagged the gaps, and formatted it all as a table I could paste into the deck. It took one message, one !deep pass, and 4 minutes.
I used to do this by hand across several browser tabs and a separate model window: copy one competitor's pricing, paste it, switch tabs, copy the next, re-read the brief, notice a missing feature, go back. Six competitors used to take about 2 hours, and the logistics wore me out more than the analysis did.
Viktor did it in 4 minutes while I made coffee. The table was clean, the gaps were marked, and I went straight into the pitch without another hour of formatting.
Using AI usually means you handle the logistics and the model does the thinking. With Viktor as an AI employee, he handles both and I approve.
Try the gearbox yourself
Set one team mode, give him one real job, then ask him where your credits are going.
Hire @viktor_com for your team, free to start: https://ref.viktor.com/leopardracer-x
Use code LEOPARDRACER100 for $100 in free Viktor credits.
Paid Partnership