Why is my AI bill exploding, and which steps actually need a model? A 6-step test

One of six steps needed the model In the video, Nick walks a competitor-monitoring workflow end to end. Six steps came out of it, and exactly one of them needed the model's judgment. The other five -- a trigger, a string match, a database read, formatting, and posting to Slack -- all have right answers you can write down as rules.

Zapier modeled 24 workflows and saw 71% average savings Zapier ran 24 common agent workflows through the same test. The average savings was 71%. That is not a tuning trick or a cheaper model. It is the result of pulling model calls out of steps that never needed judgment in the first place.

Big spending stories open the video at the 24-second mark In the video, Nick talks about Uber, Salesforce, and a $150,000 month. The key-moments table marks that stretch at the 24-second point. The point is scale, not blame. These are large firms with budget reviews, and the pattern still caught them.

What are the six steps in Nick's workflow test?

These are the steps as Nick lays them out in the video, with the timestamp where each one starts. Five of the six have a right answer you can write down as a rule. One needs judgment.

Six rows, one model call. The timestamps point to the moment in the video where Nick covers each step.

Where are the key moments in the video?

If you want to watch rather than read, these are the eight marked moments and the seconds at which they start.

Eight marked moments across a 282-second video.

Why is my AI bill so high?

Most of your AI bill is going to work that does not need AI at all, not work that could be optimized. That is the whole point of the video, and it lands in the first few seconds. The waste is not in the model you picked or the prompt you wrote. It is in the steps you handed to a model that never called for one.

Nick has a name for the pattern: token maxing. You throw the model at everything and assume the value shows up later. Sometimes it does. Often it does not. You end up paying a model to predict its way through if-then statements, which plain automation would do for free.

The reason this hides so well is that each call looks small. A few cents here, a few cents there. But a workflow runs many times a month, and five of its six steps are being billed when they could cost nothing. The bill grows with volume, not with value.

So the first move is not to shop for a cheaper model. It is to look at one workflow and ask what each step is really doing. In Nick's own example, the model was doing clerical work at a premium price, and only one step in six called for real judgment.

In the video, Nick also notes that the cost compounds. Teams hand an agent work that plain automation could do, then build a second agent to check the first one's work. Now you are paying twice for one outcome.

What is token maxing?

Token maxing is the pattern of throwing the model at everything and assuming the value shows up later. Nick's point is that it does not always show up. You end up paying a model to predict its way through if-then statements, which plain automation would do for free.

It is an easy habit to fall into. Model calls are simple to write. You describe the task in English and the model attempts it. Writing a rule takes more thought up front, even when the rule is three lines long and never fails.

The table of steps shows what token maxing looks like in practice. Step 2 asks whether an item matches a competitor you already track. That is a string match, and a model has no advantage here. Nick's words: in fact it can get it wrong. So this one is also a waste of tokens.

There is one exception to hold onto. Token maxing is not the same as using AI too much. Step 5 is a real model call and Nick is happy to pay for it. The problem is not the spending. The problem is spending on steps that have a right answer.

That is why the fix starts with a list, not a budget cut. You are not trying to use less AI. You are trying to use it where it earns its keep.

What is value maxing?

Value maxing is spending tokens only where a model's judgment actually earns its cost. It comes down to three rules: default to deterministic, spend tokens only on judgment, and keep your model options open. The question stops being how do we spend less on AI and becomes what is my AI spend actually achieving.

Default to deterministic means you start every step as a rule and only promote it to a model call when a rule cannot do the job. That is the opposite of how most teams build. Most start with a model and trim later, if they ever trim at all.

Spend tokens only on judgment is the heart of the test. Judgment means there is no single right answer you could write down. Reading a competitor's announcement and saying what it means for your company in three sentences is judgment. Trimming a date field is not.

Keep your model options open means you do not wire your whole workflow to one provider. If a cheaper or better model shows up next quarter, you want to swap it in at one step rather than rebuild everything.

In the video, the three rules land at the 203-second mark, after the worked example. That order matters. The rules sound abstract until you have seen six steps on a page and counted how many needed a model.

How do I reduce my AI API costs?

Take one workflow, go through it step by step, and mark which steps actually need a model. Anything with a right answer you can write down as a rule is an if-then, so it should cost zero tokens. When you remove model calls and replace them with old-fashioned deterministic automation, you save money. Nick's guess is that fewer steps need the model than you think.

Start with the workflow you run most often, not the one that sounds most advanced. Volume is where the money is. A clever workflow that runs twice a week costs less than a dull one that fires every ten minutes.

Write the steps out in plain words, one line each. Do not describe the tools. Describe what happens. A new item arrives. We check if it matches a name on our list. We pull a record. We tidy the fields. We ask what it means. We post it.

Now mark each line. Rule or judgment. Be strict. If you can imagine writing the answer as a lookup table, a comparison, or a sort, it is a rule. In the worked example, five of the six lines turned out to be rules.

Then rebuild. Keep the model call where you marked judgment and swap the rest for plain automation steps. You do not need to rip out the whole system at once. Pull out the cheapest step to replace first and watch the bill for a week.

Which tasks actually need an LLM and which should be deterministic?

Spend tokens on judgment: summarizing, classifying ambiguity, drafting, and deciding when the rules do not genuinely apply. Keep everything else deterministic, including triggers, string matches, database reads, and formatting. If the task has a right answer you can write down as a rule, a model is the wrong tool for it.

The steps table gives you four clean examples of deterministic work. Step 1 is a trigger: a new item shows up in the monitoring feed, and Nick says it should require zero tokens. Step 3 is a database read. Step 4 is cleaning dates and trimming fields. Step 6 is delivery to Slack.

The one judgment step is step 5. Read this and tell me what it means for our company, in three sentences. There is no lookup table for that. Two thoughtful people would write different sentences and both could be right. That is the signal you are looking at real judgment.

The exception to name is the fourth item on the judgment list: deciding when the rules do not genuinely apply. Some workflows have messy edges. If your incoming data is so varied that no rule covers it, a model call can be the honest choice. Just confirm the mess is real before you pay for it.

Are AI agents more reliable than regular automation?

Not on rule-based work. An LLM predicts a likely answer, it does not execute fixed logic. So on a task with a right answer it will occasionally miss, and it does so at a premium price. Deterministic automation is both cheaper and more reliable for that kind of step.

This is the part of the video people skip, and it may matter more than the cost. The reliability problem sits at the 71-second mark, right after the spending stories. Cost is the symptom. Flakiness is the disease.

Think about step 2 again. A string match either matches or it does not. A rule gets it right every single time, forever. A model gets it right almost every time, and the almost is what wakes you up at night. Nick is blunt about it: in fact it can get it wrong.

There is a second cost hidden in that flakiness. In the video, Nick describes teams that hand an agent work plain automation could do, then build a second agent to check the first one's work. Now you are paying twice for one outcome. The second agent exists only because the first one is not trusted on rule work.

For any small team running these jobs, the practical read is simple. Put rules where rules belong and you get fewer broken mornings, not just a smaller bill.

How much can you save by replacing agent steps with automation?

Zapier modeled 24 common agent workflows this way and the average savings was 71%. Your mileage will depend on how complex your workflows are and which models you are using. But the direction holds.

That figure comes from running the same step-by-step test across a set of workflows people really build. It is not a lab benchmark. Twenty-four workflows is a small sample, and an average hides the spread. Some will have saved more and some less.

What makes the number believable is the arithmetic of the worked example. Six steps, one model call. If five of six steps stop billing, a large cut is not a surprise. The savings come from removal, not from tuning.

Hold two cautions in mind. First, the number is an average, so do not promise it to anyone before you run your own test. Second, it is specific to agent workflows, which are repeated, multi-step jobs. A one-off research task you run by hand is a different animal.

In the video, the 24-workflow result lands at the 179-second mark, right after the six-step count. The order is deliberate. You see one workflow first, then the pattern across many.

Why are companies overspending on AI tokens?

Uber burned through its entire AI budget in the first four months of the year. Salesforce is looking at $300 million in Anthropic tokens this year, during a hiring freeze, and one engineer reportedly put up a $150,000 Claude Code bill in a single month. Those stories open the video at the 24-second mark.

The cost also compounds: teams hand an agent work that plain automation could do, then build a second agent to check the first one's work. Now you are paying twice for one outcome. The key-moments table marks that double-cost problem at the 58-second mark.

These are not small, careless shops. They have budget reviews and engineers who read invoices. The pattern still caught them, which is the reason Nick opens with those stories.

Part of it is speed. A model call ships today. A rule takes an afternoon of thought. When a team is under pressure, the model call wins, and nobody goes back to check whether that step ever needed judgment.

Part of it is that nobody owns the question. The person who wrote the workflow is not the person who sees the bill. By the time the bill looks strange, the workflow has run over and over. The waste is structural, and the fix is a test you can run in an hour.

How much does it cost to run an AI agent?

Far less than most teams pay, because most of the bill goes to steps that never needed a model. The test in this video is to break one workflow into its actual steps and mark which need judgment. In the competitor-monitoring example, six steps came out and exactly one did. Zapier modelled 24 common agent workflows the same way and the average saving was 71%.

There is no single price for an agent, and anyone quoting you one is guessing. The cost depends on how many steps call a model, how long the inputs are, how often the workflow fires, and which model you picked.

That is why the step count is the useful number instead of a dollar figure. If one step in six calls a model, your cost per run is roughly one step's cost, not six. If all six call a model, you are paying six times over for one answer.

The video does not give a per-run dollar cost for the example workflow. It gives the step split and the 71% average across 24 workflows. We are not going to fill that gap with a made-up estimate.

What you can do is measure your own. Run the test, replace the rule steps, then compare one month to the next. That comparison is more useful than any published price.

How do I know if a step needs an LLM?

If the task has a right answer you can write down as a rule, a model is the wrong tool for it. That one line is the whole test. Everything else is practice applying it.

Here is a fast way to check. Ask yourself if two careful people doing the step by hand would produce the same output. If yes, it is a rule. A date gets trimmed the same way by everyone. A name either appears on a list or it does not.

Now ask if the step requires reading something and forming an opinion. Step 5 in the table asks what a competitor's news means for our company, in three sentences. Two people would write different sentences. That is judgment, and it is the part Nick is happy to pay for.

Watch for a trap in the middle. Classifying things sounds like judgment but often is not. Sorting emails by sender domain is a rule. Sorting them by how upset the writer sounds is judgment. Nick's judgment list names classifying ambiguity, and the key word is ambiguity.

Write your marks down before you touch any code. A step list with rule or judgment next to each line is the artifact that saves you money. The rebuild is the easy part once the list exists.

Which AI tools save the most time?

The question is flipping around: the tools that save time are the ones you stop paying a model to do. A trigger, a string match, a database read and formatting are all faster and more reliable as plain automation, and they cost nothing to run.

Speed is the quiet benefit of going deterministic. A rule returns in a blink. A model call waits on a network round trip and a generation. Strip four model calls out of a six-step workflow and the whole thing finishes in a fraction of the time.

There is a reliability dividend too. A rule does not have a bad day. It does not need a retry, a second agent checking it, or a person reading the output to confirm the format held.

This video is sponsored by Zapier, and Zapier has published the framework behind the 24-workflow test. You can get the Zapier framework guide through the link in the sources below if you want their version of the step-by-step test.

Our advice for a reader with limited hours is to pick the loudest workflow and run the test once. One honest step list tells you more about your tools than a week of comparison shopping.

What should my team do this week?

Pick one workflow, write out its steps, and mark each one rule or judgment -- the same move that turned up 6 steps and 1 model call in Nick's example. That is a one-hour job, not a project.

Start with the workflow that runs most often. Write each step as a plain sentence with no tool names. Then mark the lines. Then count how many model calls you are paying for versus how many you marked as judgment.

If the counts do not match, you have found your savings. Replace the cheapest rule step first with plain automation and watch for a week. Nothing should break, because rules do not guess.

Then repeat on the next workflow. The three rules carry over: default to deterministic, spend tokens only on judgment, and keep your model options open. Those rules appear at the 203-second mark of the video if you want to hear them in Nick's words.

The frame to hold onto is the one from the pull quote. The question isn't how do we spend less on AI, it's what is my AI spend actually achieving. A smaller bill is a side effect of answering that honestly.

Method and sources: The data on this page comes from one video by Nick Standlea, published August 5, 2026, and current as of October 2026. The step table copies the six steps of the competitor-monitoring example exactly as Nick lays them out, with the timestamp where each begins; the second table lists the eight marked moments in the 282-second video. The savings figure of 71% is an average across 24 common agent workflows that Zapier modeled, not a guarantee for any single workflow. The spending figures for Uber, Salesforce, and the $150,000 month are as stated in the video; the video does not give a per-run dollar cost for the example workflow, so we have not printed one. This video is sponsored by Zapier, and the framework guide link appears in the sources.