Should an AI workflow use more than one model? How to stay model-agnostic with Zapier

Six steps, and only one of them is about picking a model The video lays out 6 steps. Just one of them, step 3, is about choosing the model for the task. The other five are about the workflow around it: separating the layers, building one real end-to-end job, swapping and comparing, judging on reliability, and setting guardrails. Step 1 says the model is one replaceable part, not the foundation. The rest of the list is the part that stays put when the model changes.

The swap happens at the 126-second mark, and nothing else gets rebuilt In the video, the key moment at 126 seconds is switching models without rebuilding. When a model gets cheaper, faster, or more reliable, Nick switches it, tests the agent again, compares the output, and keeps the system moving. The data, tools, apps, approvals, and rules all stay where they are because they never lived inside the model.

Guardrails come last in the list and first in a real business Step 6, at the 163-second mark, is guardrails. Nick does not let an agent randomly send emails, expose private information, edit client files, or make decisions without his approval. In Zapier he builds permissions, approvals, and rules that limit what the agent can do. He calls that the line between playing with AI and using it inside a business you care about.

The 6 steps as Nick lays them out in the video

These are the steps from the video, in order, with the timestamp where each one starts. Read them top to bottom and you have the method.

Six steps in all. Timestamps are seconds into the video.

Key moments with timestamps

Seven marked moments in the video, in order, if you want to jump straight to one idea.

Seven moments. The video runs 207 seconds.

Should AI workflows use more than one model?

Yes, and the reason is simpler than it sounds. The best model keeps changing. That is the first moment in the video, at the 0-second mark, and it sets up everything after it.

If your whole business runs on one model, you are not building a system. You are placing a bet on one company. You are betting that its price, its policies, its performance, and your access all stay the same. The second key moment, at 25 seconds, says it plainly: one model is a bet, not a system.

The fix is not to run five models at once for the sake of it. The fix is to keep the workflow in a layer that is separate from the model. Then the model you use can change without hurting the rest. Today one agent runs on Claude. Tomorrow the same agent can run on GPT or Gemini.

For a small team, this is a question about risk more than taste. A model you like can change its price, its policies, or your access. If the rest of your setup is glued to it, each of those changes becomes your problem. If the rest of your setup sits outside it, those changes become a choice you make on your own clock.

This guide is sponsored by Zapier, the tool Nick builds his agents in. You can try Zapier free. The steps below work as a way of thinking even if you build somewhere else.

What does a model-agnostic AI workflow actually mean?

It means the model is one replaceable part, not the foundation. That is step 1 in the table, and it starts at the 63-second mark.

Think of your workflow as two layers. The top layer is the job: the data, the tools, the apps, the approvals, and the business rules. The bottom layer is the model that handles the step where judgment is needed. Model-agnostic means the top layer does not care which model sits underneath.

Nick builds through Zapier so his data, tools, apps, approvals, and business rules do not have to live inside any one AI company's ecosystem. That is the practical version of the idea. His data is not locked inside one chat tool. His approval rules and app connections sit in the workflow layer instead.

The test is easy to run on your own setup. Ask what has to change if you swap the model tomorrow. If everything else stays put, you are model-agnostic. If your data, your apps, and your rules all have to move with it, you are not.

For a family business or a small team, the payoff is calm. You are not forced to rush when the news changes. You read about a new model, you try it, and you keep the rest.

How do I switch AI models without rebuilding my workflow?

Keep the workflow, data, tools, approvals, and rules in a layer outside the model. That is step 1, and step 4 at the 126-second mark is where the swap itself happens.

The order in step 4 is good to copy word for word. You switch the model. You test the agent again. You compare the output. Then you keep the system moving. Four small moves, and none of them is a rebuild.

The trigger for a swap is also named: something new comes out, or gets cheaper, or faster, or more reliable. Any of those is a fine reason to try. None of them is a reason to tear down the job around it.

Comparing output is the step people skip. The video is clear that you test the agent again and compare the output before you move on. You do not have to take a swap on faith. You read what both models gave you and then pick.

This is why the key moment at 126 seconds is titled switching models without rebuilding. The savings are not in the swap. They are in everything you did not have to touch.

What is a good first AI agent to build for a business?

Something repetitive you already do and can check before it ships. Step 2, at the 82-second mark, says to build one real workflow end to end.

Nick's example is concrete. His agent reads the transcript of his latest podcast episode. It helps him write an article based on what was said in the interview. It puts a draft together in his LinkedIn account. He makes a few edits so it sounds like him, then hits post.

Notice what the agent does and does not do. It does the reading, the drafting, and the placing. It does not do the posting. The agent does the work and he keeps the judgment. That split is a good rule for a first build.

Start with something you actually do every week, the way step 2 describes it. A task that comes around often gives you more chances to see the output and fix it. A task you rarely run gives you very little to go on.

Build it end to end before you make it fancy. One real job, one model step, one output you can review. The point of step 2 is that the whole path runs, not that any one piece is perfect.

Should I use Claude or GPT or Gemini for my AI agent?

Pick per task rather than picking one forever. Step 3, at the 108-second mark, says to choose the model that fits the task today.

Nick runs his article agent on Claude right now because he likes how Claude thinks through the research and the writing. That is a judgment about one job, not a lifetime pledge. Another task might belong on GPT or Gemini.

There is no table in this video that ranks models, and that is on purpose. Any ranking would go stale. What stays true is the method: try the task on more than one model, read the output, and choose with your own eyes.

Step 5, at 137 seconds, puts the same idea in business terms. Nick wants to be able to say Claude is better for this task today, GPT is better for that one, and change his mind when that changes. Changing your mind is only cheap if the workflow lets you.

So the honest answer to which model is best is that it depends on the task and on the month. Build so that answer can move. Then the question stops being stressful and becomes a small test you run now and then.

What is AI agent orchestration, and why does it matter here?

Orchestration is the layer that decides which step runs, in what order, and which model -- if any -- each step needs. It is the top layer from step 1, the one that holds your data, tools, apps, approvals, and business rules.

The argument in this video is that the orchestration layer, not the model, is where the durability lives. Models change fast. The shape of your job changes slowly. A good orchestration layer keeps the slow part safe from the fast part.

If your workflow is wired to one provider, you inherit that provider's outages, price changes, and deprecations. None of those are in your control. All of them land on your calendar anyway.

In plain terms, orchestration is the plumbing. In Nick's example it gets the transcript, passes it along, and puts a draft together in his LinkedIn account. Only one step in that chain needs a model at all. The rest are steps with right answers.

That is also a cost point. Steps with right answers should not call a model. Triggers, lookups, string matches, and formatting all belong in the plumbing. Reserve the model for the genuinely ambiguous step, which here is the writing.

What is the best AI workflow automation platform?

The one that lets you change models without rebuilding the workflow. That is the test proposed here: not which platform has the best model today, but which one makes the model a swappable component.

Nick uses Zapier, and this guide is sponsored by Zapier. He builds there because his data, tools, apps, approvals, and business rules do not have to live inside any one AI company's ecosystem. If you want to follow along in the same tool, you can try Zapier free.

Apply the test to any tool you are weighing. The question is whether your data, tools, apps, approvals, and rules can stay in place while the model changes underneath them. That is the line step 1 draws.

A workflow that survives its provider changing adds up to more than one tuned to whoever is currently ahead. The tuned one may look better right now. The durable one holds up as the model news keeps moving.

One more thing to look for is the part step 6 covers. The platform should let you set permissions, approvals, and rules about what the agent is allowed to do.

How do I stop an AI agent from doing things without my approval?

Put guardrails in the workflow itself. Step 6, at the 163-second mark, is about permissions, approvals, and rules.

The video names four things to block. Do not let an agent randomly send emails, expose private information, edit client files, or make decisions without your approval. Those four cover most of the ways an agent can embarrass or harm a business.

In Zapier, Nick builds workflows with permissions, approvals, and rules that limit what the agent is allowed to do. Permissions decide which accounts it can touch. Approvals put a person in the path. Rules say what counts as allowed before the agent ever runs.

His own example shows the pattern. For his LinkedIn draft he still edits it himself before hitting post. The agent is allowed to draft. It is not allowed to publish. Writing that one line down is most of the safety work.

Nick calls this the line between playing with AI and using it inside a business you care about. A demo can act freely because nothing real is attached. Your business has clients, files, and an inbox, so the agent gets a shorter leash.

Is it bad to build my whole AI workflow on one model?

It is a bet more than a system. The key moment at 25 seconds says exactly that: one model is a bet, not a system.

Here is what you are betting on. You are betting one company will always have the best model at the right price with the same policies, performance, and access, and never change any of it on you. Each of those is a separate promise, and no vendor made any of them to you.

Models, prices, terms, and performance all change, and sometimes access gets restricted. That is not a prediction about any one company. It is just what the last few years have looked like for everyone.

The cost of the bet is not the money. It is the rebuild. Step 4 is clear that you should not have to rebuild the whole workflow from scratch. If everything lives inside one model, that is exactly what a change forces on you.

Build so the model can be swapped instead. Then a change at one AI company is a change to one part of your setup. The rest of the system stays where it is.

What are the best AI workflows for business to start with?

The ones where most steps are deterministic and only the genuinely ambiguous step calls a model. Nick's podcast-to-LinkedIn agent, at the 82-second mark, is a clean example of that shape.

Walk through it. Reading the transcript is deterministic. Passing it along is deterministic. Putting a draft together in his LinkedIn account is deterministic. Writing an article based on what was said in the interview is the ambiguous part, and that is the part the model handles.

Triggers, lookups, string matches, and formatting all have right answers you can write down, so they should never cost a token. Reserve the model for judgment. That keeps the run cheap and keeps the output steady, because the steps with right answers always give the right answer.

Good candidates in most businesses follow the same shape. Most of the path is fetching and filing, and one step in the middle needs judgment. If you can name that one step, you have found a good first build.

For a first build, keep a person in the path before anything ships. Step 6 is the reason. An agent should not send emails or make decisions without your approval until the permissions, approvals, and rules are in place.

How do I know a model swap actually made things better?

You test the agent again and compare the output, which is the second half of step 4 at the 126-second mark. Nothing in the video asks you to trust a benchmark.

The order matters. First you switch the model. Then you test the agent again. Then you compare the output. Only after that do you keep the system moving. Skipping the comparison turns a careful swap into a guess.

Judge on reliability, not hype. That is step 5, at the 137-second mark. In a real business, reliability matters much more than whatever is trending. A model that does well most of the time and odd the rest is harder to live with than a steady one.

Nick's own standard is about the task, not the headlines. He wants to be able to say Claude is better for this task today, GPT is better for that one, and change his mind when that changes. That is a judgment you make from the output in front of you.

If the new model does not clearly win, keep the one you had. The point of staying model-agnostic is not to change often. It is to be able to change when it pays.

What should I do this week to get started?

Run the 6 steps in order on one task you already do every week. The table on this page is the full list, and the whole video is 207 seconds if you want to watch it first.

Start with step 1 by deciding that the model is a replaceable part. Then do step 2 and build one real workflow end to end, even if it is rough. A rough workflow that runs beats a plan that never ships.

Next comes step 3: pick the model that fits that task today. Nick picked Claude for his article agent because of how it thinks through research and writing. Your pick should come from the task in front of you in the same way.

Before you let it touch anything real, do step 6. Write down what the agent may do and what it may not. Set permissions, approvals, and rules. For a first build, a human review before anything ships is a fair default.

Nick builds all of this in Zapier, the sponsor of this guide, and you can try Zapier free. Whatever tool you pick, the order of the steps is what carries over.

Method and sources: The source is a single video from The Nick Standlea Show, published May 15, 2026, running 207 seconds, in which Nick lays out 6 steps for keeping an AI workflow model-agnostic. The tables on this page copy those 6 steps and the 7 marked key moments exactly as the video presents them, with timestamps in seconds from the start. The page covers only what the video says; it does not rank models, quote prices, or report benchmark scores, because the video gives none of those. The video names Claude, GPT, and Gemini as examples but does not test them against each other, so no comparison data exists here. Steps are ordered as Nick presents them, not by importance, and this guide is sponsored by Zapier, the tool he builds in.