Module 5 · Page 3 of 4 · 20 minutes

Understand the real cost

A prototype can work perfectly and still be a bad business system. Before scaling an AI workflow, understand what it will cost to run, monitor, and maintain.

Working is not the same as economical

Vibecoding makes it easy to get something working quickly. But the cost of a prototype may look very different from the cost of running the same workflow every day across thousands of users or records.

Suppose you build a competitor-monitoring agent that searches the web and summarizes updates for five companies once per day.

That may be inexpensive. But if the same system monitors 2,000 companies every hour, the number of searches, model calls, and generated tokens increases dramatically.

The system still “works.” The business question is whether it's worth operating at that scale.

AI systems have several kinds of cost

Cost Examples
Model usage Tokens, API calls, or subscription usage
Tools and services Search APIs, databases, cloud storage, external software
Human oversight Reviewing outputs, approving actions, handling errors
Maintenance Updating prompts, workflows, integrations, and rules

The cheapest model isn't necessarily the cheapest system. If a lower-cost model requires much more human review, the total cost may be higher.

Model usage adds up

Many AI services charge based on usage. A common unit is the token: a small piece of text processed by the model.

A single task may involve much more than your final prompt and answer. An agent may also read project files, inspect previous tool results, search for information, revise its approach, and generate intermediate outputs.

That means cost often depends on:

Small inefficiencies become large at scale

Imagine an agent that analyzes one customer complaint by reading ten large files first. For one complaint, the extra work may barely matter.

If the system processes 100,000 complaints per month, repeatedly reading those same files can become slow and expensive.

This connects back to project organization and memory: give the agent the context it needs, but avoid making it rediscover unnecessary information on every run.

Human review is also a cost

Some workflows look cheap because the AI usage is inexpensive, but they require substantial human checking.

Suppose an agent drafts 1,000 customer emails each week. If every email requires two minutes of human review, that's more than 30 hours of review time.

Human oversight may be necessary and valuable. The point is to include it when judging whether the workflow saves time or money.

Maintenance is part of the system

AI workflows don't remain correct forever. Business rules change. Data formats change. APIs change. Models change. The organization may adopt new policies.

Someone needs to notice those changes and update the workflow.

A useful question is:

Who will own this after the prototype works?

If nobody is responsible for monitoring and maintaining the system, a useful prototype can quietly become unreliable.

Estimate before you scale

You don't need a perfect financial model. A rough estimate is often enough to reveal whether a design makes sense.

Ask Example
How often will it run? Once a month or every hour?
How much work happens each run? One document or 10,000 records?
What paid services are used? Model, search, database, or API fees?
How much human review is needed? Spot-checking or reviewing every result?
Who maintains it? Who updates the workflow when something changes?

A useful rule

A prototype that works is not necessarily a system that is economical to operate.

Before scaling, estimate the cost of the entire workflow — not just the AI model.

Exercise

Imagine you built an agent that reads incoming customer reviews, classifies the issue, summarizes the complaint, and suggests a response.

Now imagine the company wants to use it for 50,000 reviews per month. Identify:

  1. What model or tool costs might increase?
  2. How much human review would still be needed?
  3. What parts of the workflow would require maintenance?
  4. What could you change to make the system cheaper or more efficient?
Next · Page 4 of 4
Know when not to vibecode