---
title: "Build the business system. Choose the AI later."
seoTitle: "Build an AI business system you own"
author: "Enzo Sison"
description: "What running Prism’s TikTok with different AI agents taught me about building a business system you own: clear rules, useful memory, and human control."
date: "2026-09-27T12:00:00-07:00"
category: "Marketing"
gradientClass: "from-neutral-900 via-stone-800 to-neutral-900"
---
Since June, I’ve used AI agents to help run Prism’s TikTok. Over the summer, I changed the models and tools behind the workflow several times. Eventually, I moved the model’s thinking onto a computer on my desk.
The posts kept going out.
That was the most useful result of the experiment. The workflow no longer depended on one model, one app, or me remembering every step.
For a small business owner, that changes the question worth asking. Before comparing AI subscriptions, ask: **What part of my business should keep working when the tools change?**
That is the system to build.
## The part you own
I think about an AI business system in three layers.
**The brain is the model.** It weighs options, chooses a promising clip, drafts a caption, or helps interpret an error.
**The body is the software that lets it work.** It connects the model to files and approved tools. You may hear this called an agent or a harness.
**The business layer is everything you have taught the operation.** Your source material. Your quality standards. The scripts that carry out routine steps. A record of what happened. A playbook of what you have learned.
This third layer deserves more of your attention.
Think about a restaurant. A new chef needs recipes, supplier relationships, prep routines, and service standards. If all of that exists only in the previous chef’s head, replacing the chef means rebuilding the restaurant.
The same problem appears when important instructions live only inside a chat. A new tool arrives, and you start explaining your business from scratch.
Keep that knowledge in files and records your business controls, in formats you can inspect and move. Your next AI should inherit a working operation.
The model and tools can change. Your operating knowledge should stay with the business.
## Give AI judgment. Give scripts the routine.
The clearest way I can explain my setup is this:
> The AI is the manager, and the scripts are the assembly line.
The model helps decide which moments are worth sharing and what an unexpected error means. Routine execution follows defined steps.
In the workflow I describe in the video, the agent reviews our interview library, proposes clips, checks for duplicates, and prepares captions. I approve the batch. The publishing process follows a fixed schedule and checks that each post is actually live.
That last distinction matters. “Scheduled” and “published” are different states. A successful request to a publishing service is still something to verify.
The operating rules also limit the account and pace. The tools handle credentials without putting passwords in the model’s conversation. If something fails, the workflow inspects the failure before trying again.
Those boundaries reduce the consequences of a bad decision. They do not make the system infallible. A model can still choose a weak clip or draft something I would not publish. That is why approval and review remain part of the work.
For your first workflow, write the allowed actions before giving an agent access to them.
Judgment, approval, execution, and verification each have a clear job.
## Write down what happened, then what it taught you
My system keeps a posting ledger and a playbook. They solve different problems.
The ledger answers factual questions: What clip did we use? Was it approved? When was it scheduled? Where is the live post? Did the last attempt fail?
The playbook captures what we think the results mean: Which openings hold attention? Which subjects attract useful conversations? What should we try next?
Without the ledger, the agent can repeat work or mistake an intention for an outcome. Without the playbook, each run can repeat the same mistakes.
In the video, I describe reviewing performance at two days and seven days instead of judging everything in its first few hours. The exact windows can change by workflow. The useful habit is to review comparable results at a consistent age.
Keep lessons specific and revisable. “This opening performed well across these posts; test it again” is more useful than “Always use this opening.” A playbook should preserve evidence and uncertainty alongside the recommendation.
When I changed the model, that accumulated knowledge stayed available. The new model did not need to rediscover every lesson.
Keep the record of what happened and the lessons for the next run outside the model.
## Measure the operation and the outcome separately
As I explored in [Content Is a Power Law Game Now](/blog/content-is-a-power-law-game-now), a few posts can account for much of an account’s growth. That makes measurement as important as consistency.
A content system has at least two jobs: carry out the work reliably and help produce content people value.
It can succeed at one while struggling with the other.
The workflow described in the video showed continuity across the configurations we tried. It did not establish that every model makes equally good creative decisions. In the first local-model period I discussed, the workflow kept operating, but we had not seen a breakout post. I also personally posted the biggest hit of the summer.
There is another complication: a queue scheduled weeks ahead can contain decisions made before a model change. Publishing a clip after a switch does not tell you which model selected it.
A useful comparison would track who selected each clip, when it was made, human changes, and performance at the same age. It would also count time saved, failures, and cost.
For a business owner, this prevents a common mistake: celebrating automation activity while losing sight of the business result. Ten drafts are activity. A useful draft approved with less effort is progress. A post that earns a qualified inquiry is another outcome to measure.
## Local AI changes the bill, not the need for a system
Moving inference onto my own hardware gave me another way to run the workflow. It did not make the entire operation free.
There was an upfront hardware cost. Electricity, maintenance, and setup still matter. I also pay for the service that publishes the content.
The precise benefit is avoiding a per-use cloud inference bill for work the local model can handle. Whether that saves money depends on utilization and the time required to keep it running.
Local inference also does not mean the whole workflow stays offline. Content still leaves the machine when it is published, and connected services have their own data flows.
The practical choice is to match the tool to the job. You might use a hosted model for early experiments, move predictable work locally later, or combine them. There is no need to buy hardware before you have a useful routine worth running. For a deeper look at where local processing helps, see [ten ways a business can use DGX Spark](/blog/10-ways-dgx-spark-improve-online-business).
## Start with one loop you can inspect
For a small business, I would begin with one recurring task where the result is easy to review. A weekly content shortlist is a good example because the first version can stop at a draft.
Write a short operating brief:
- **Input:** one recording you own and the previous posting history.
- **Decision:** propose three moments that answer real customer questions.
- **Rules:** preserve the speaker’s meaning, avoid duplicates, and explain each selection.
- **Approval:** a named person reviews every clip and caption before publication.
- **Record:** save the selection, changes, approval, and eventual live link.
- **Review:** compare results at fixed intervals and record the next experiment.
Run it manually first. Then automate the repetitive steps. An AI can help write the scripts, but inspect and test them with drafts before connecting public actions.
As the process becomes dependable, you can grant narrow standing permission for a specific action under specific conditions. Keep an obvious way to pause it. A workflow that stops when it cannot verify the next step is doing its job.
## Keep the material only you can provide
An agent needs something worth working with.
Your customer questions, experience, demonstrations, and point of view give it useful material. Record the explanation you repeat every week. Show how you solve a real problem. Explain the tradeoff a customer usually misses.
Those inputs help a content system express your expertise instead of filling a calendar with generic advice.
That connects to how we think about [content at Prism](/content): build a repeatable way to turn useful knowledge into work that helps customers understand and trust the business.
Start with one loop. Keep its rules, records, and lessons. Improve it until another person, or another model, can pick it up without guessing.
Then you have something more durable than a favorite AI tool. You have an operation your business owns.
*Prism is currently at capacity. [Join the waitlist](/waitlist?focus=content) if you want help building your website, content, and ads systems.*
## Watch the experiment
In the video, I walk through the TikTok workflow, the move to local AI, the results I observed, and the questions I am still testing.
[Watch the full video on YouTube](https://www.youtube.com/watch?v=S_Ol7aMNOeU).
