I like AI because it can remove work. But there is a strange point where the system that was supposed to save you time starts eating money in the background.
You add another agent. Another long instruction file. Another research step. Another retry. The answer gets a little better. The bill gets much bigger.
The wrong question is, “How do I get more AI credits?”
The better question is, “Why did this job need so much AI in the first place?”
First, Understand What You Are Actually Paying For
FACT: AI API costs commonly depend on how much information you send in and how much the model generates back. OpenAI's current guidance says usage is priced by input, output, and cached tokens, with rates varying by model. It also recommends shortening or preprocessing large inputs when needed.
FACT: Repeated context can often be cached. OpenAI says prompt caching can lower the cost and latency of reused prompt prefixes. Anthropic also documents large savings from prompt caching for repeated Claude context.
INTERPRETATION: If your AI bill keeps climbing, the model price may not be the main problem. Your workflow design may be making the model read, reason about, and produce far more than the result requires.
MY BELIEF: A good one-person business should become simpler as its systems improve. If automation makes the company harder to understand and more expensive to operate, I want to question the automation.
The Seven Places I Would Look for AI Credit Waste
1. You keep sending the whole library
Your AI may not need your entire company brain to answer one support question. Give each job only the context it needs. Search or retrieve the relevant pieces first when the knowledge base is large.
2. Your stable instructions keep changing position
Repeated instructions can benefit from prompt caching when the reusable prefix stays consistent. OpenAI specifically recommends placing static instructions and examples before the variable information when you want cache matches.
3. You use the strongest model for everything
Classification, formatting, extraction, simple rewriting, and routing may not need your most capable reasoning model. Test the cheapest model that can meet the quality bar, then escalate only when it fails.
4. Your output is longer than the customer needs
Output tokens cost money too. If a customer needs a two-line answer, do not ask the system to write a page and then ask another model to summarize it.
5. Agents are talking to agents without a stopping rule
Every handoff can become another model call. Define the job, the maximum number of attempts, what counts as success, and when the workflow must stop or ask for human judgment.
6. You regenerate work that could be saved
If an answer, summary, classification, or asset is still valid, store it. Do not pay the model to rediscover the same stable answer every time.
7. You measure tokens instead of useful results
A cheap call that fails five times is not cheap. Track the full cost of producing one accepted result. That includes retries and the human time needed to fix weak output.
The Credit Audit I Would Run Today
- List every AI workflow that runs more than once.
- Write down which model each step uses.
- Check how much context is sent on every call.
- Remove instructions that are repeated or no longer needed.
- Put stable reusable instructions first so caching can work where supported.
- Test a smaller model on simple steps.
- Set a maximum retry count for agent workflows.
- Cap output length where long answers add no value.
- Save stable results instead of regenerating them.
- Measure cost per accepted result, not just cost per call.
A Simple Routing Rule
Before a workflow calls AI, I would make it answer four questions.
- Can a fixed rule do this? If yes, use the rule.
- Can a small model do this reliably? If yes, use the small model.
- Does this job need deeper judgment? If yes, send only that job to the stronger model.
- Would a mistake be expensive? If yes, keep a human approval point.
This is the same principle I use across a one-person AI business. Do not hire seven digital employees to perform three jobs. Do not make twelve AI calls when two reliable calls and one fixed rule can produce the same customer result.
What the Current Technical Guidance Says
OpenAI's current model guidance recommends leaner prompts, exposing only relevant tools, and measuring representative tasks instead of assuming that more calls improve quality. Its internal examples found large token and cost reductions from leaner instructions, though it correctly warns that results vary by workload.
Anthropic's published guidance likewise treats prompt caching as a way to reuse repeated context more efficiently. These are useful technical tools. But neither changes the deeper business question: should the workflow be doing all that work at all?
Read the primary sources: OpenAI model and prompt-efficiency guidance and Anthropic on contextual retrieval and prompt caching.
Start by making the work clearer.
If your AI systems feel messy, a better prompt is often the easiest place to simplify the job before adding another tool or agent.
Get My Claude27 Prompts Free →Keep Going
Bring It Back to Your Business
Look at the AI system you use most.
Do not ask whether it is impressive. Ask what work it removes, what result it creates, what it costs to create that result, and what would break if you simplified it.
The facts matter. So does the design you choose.
Your next move may not be adding more credits. It may be removing work that never needed to happen.