If an AI tool takes text in and gives text back, with no connection to your other systems, a well-written prompt and the AI access you already have will usually do the same job. Pay for a product when it does something a prompt cannot: work inside the systems you use, handle data you could not paste into a chat, control who can do what, or take the testing and upkeep of prompts off your hands. Before you buy, test it against your own prompt on ten real inputs.
What is inside most AI tools
Strip the interface off a typical AI writing or support tool and you find four parts: a system prompt written by the vendor, perhaps a few examples, a call to a model from Anthropic, OpenAI, Google or another provider, and a screen to type into. Founders and investors argue about whether products like that can survive being copied. That is their problem. Yours is narrower: is the part you cannot see worth the price, compared with writing it yourself?
The hidden prompt is often good, written by people who have thought about the job longer than you have. That is real value. It is also value you can mostly reproduce, because a prompt is just text, and yours is text you can read, test and improve.
The arithmetic
Take a hypothetical team of eight using an AI writing tool at $40 a seat per month to draft customer emails. That is $320 a month, or $3,840 a year. Now price the same work through an API with a good prompt, at list prices checked October 20, 2026:
8 people × 15 drafts a day × 22 working days = 2,640 drafts a month
Each draft: 1,500 input tokens (prompt, notes, thread), 500 output tokens
GPT-6.1 Sol (1,500 × $2 + 500 × $10) ÷ 1,000,000 = $0.008
× 2,640 = $21.12 a month
Claude Haiku 5.5 (1,500 × $0.10 + 500 × $0.50) ÷ 1,000,000 = $0.0004
× 2,640 = $1.06 a month
2,640 drafts of 1,500 tokens in and 500 out. The API figures are the model bill only. Source: thinkhard.si model catalog, list prices checked October 20, 2026.
The model bill is a rounding error next to the seat price. So the seat price pays for everything except the model: the interface, the hidden prompt, the integrations, the support and someone keeping it all working when models change. The real question is whether your team uses those things.
The API route has costs of its own. Someone has to set it up, and people need somewhere to use it. Many teams can skip that entirely: most chat assistants a company already pays for let you save standing instructions for a project or a custom assistant, and a good prompt saved there covers a lot of small-team work. If you do use an API key, remember that every call is billed to your account, so set a spending limit.
A free tier costs nothing, but read its terms. Google’s Gemini API pricing page says content sent on the free tier is used to improve Google’s products, which rules it out for customer data.
The ten-input test
Before you buy or renew, put the tool up against your own prompt.
- Collect ten real inputs from the job the tool is for: ten customer emails, ten sets of product notes, ten call transcripts. Include two awkward ones.
- Run them through the tool (most offer a trial) and save the outputs.
- Write your own prompt for the same job. Start from one of our prompt templates if one fits, and check it in the Prompt Grader until it passes most of the ten checks.
- Run your prompt on the same ten inputs in Prompt Tester, with checks for what every output must contain. Try a fast model and a mid-tier one.
- Shuffle the two sets of outputs, hide which is which, and ask the person who will use them to pick.
If your prompt wins or ties, the tool has to justify its price with something other than writing quality. If the tool wins clearly, read its outputs closely and work out what it does that your prompt does not. Sometimes that is a lesson you can copy into your prompt. Sometimes it is exactly what you are paying for, and then you know.
When a product earns its price
A product is worth paying for when the value sits outside the text box.
- It works where the work is. It reads the ticket history from your helpdesk, writes the summary back to your CRM and files the result, with no copying and pasting. Plumbing is real work, and it is the best reason to pay for a tool.
- It handles data you should not paste into a chat. Admin controls, retention settings, the data processing terms your own customers require, and a clear answer about whether your data trains anyone’s models.
- Many people need the same process done the same way. One shared, versioned prompt with permissions beats twenty people each keeping their own copy in a notes app.
- Someone else maintains it. When a model is retired or changes behavior, the vendor re-tests and fixes the prompt. Ask how they do that. A vendor that cannot show you test results is asking you to trust a feeling.
- It does what a model cannot do alone, such as searching your documents well, acting in your systems with the right permissions, or keeping a record an auditor will accept.
Questions to ask a vendor
- Which models does it use, and can we choose?
- What happens to our data, how long is it kept, and is it used for training?
- How do you test changes when the underlying model changes, and can we see the results?
- Can we see or edit the instructions it uses for our account?
- Will you run it on ten of our inputs next to a plain chat assistant with a good prompt, and show us the difference?
- If we leave, what can we export?
When a better prompt is enough
A better prompt is enough when a few people do the job, the input and output are text, the volume is modest and nothing has to flow into another system. That covers a lot of what small teams buy: drafting emails, rewriting copy, summarizing calls, turning notes into a brief.
- Drafting emails: Has to flow into other systems 1/10, People who need it done the same way 2/10
- Summarizing calls: Has to flow into other systems 3/10, People who need it done the same way 3/10
- Helpdesk to CRM: Has to flow into other systems 9/10, People who need it done the same way 6/10
- Twenty people, one process: Has to flow into other systems 3/10, People who need it done the same way 8/10
Illustrative: placed by judgment. Text in and text out for a few people sits bottom left; value outside the text box moves a job up or to the right.
For those jobs, write the prompt properly, save it where the team will find it, and test it on real inputs before everyone relies on it. Our My Prompts keeps each version of your own prompts in your browser while you work on them. Then spend the seat money on something your prompt cannot do.
Key takeaways
- Many AI tools are a hidden prompt, a model call and an interface. The model is the cheapest part.
- For text in, text out work with no integrations, a well-written prompt usually does the same job.
- Pay for integrations, data controls, shared and versioned processes, and maintenance you would otherwise do yourself.
- Before buying, compare the tool with your own prompt on ten real inputs, judged blind by the person who will use the output.
- Ask vendors which models they use, what happens to your data, and how they test when models change.