How to use the OpenAI (ChatGPT) and Claude API without coding

You've probably pasted a list into ChatGPT or Claude and asked it to sort the items into categories. It works well for 30 rows. For 30,000 rows, or for every new ticket that arrives tomorrow, the chat window stops being the right tool. That work belongs on the API, the way software talks to the same AI.

The API usually needs a developer. It doesn't have to. This guide explains, without jargon:

• why the API is better than the chat app for repeated work;
• what you need to use it;
• how a non-technical person can put the OpenAI or Claude API to work on their own spreadsheet, with no code.

The chat app is ChatGPT or the Claude app, on the web, desktop or phone. A person types, the AI answers, and the conversation is the record.

The API is the same family of AI models, reached by software instead of by a person. OpenAI's is on the OpenAI Platform and Anthropic's on the Claude Console. Each has its own account, billing and API keys, separate from a ChatGPT or Claude subscription. You pay for what you use. API billing is managed separately from the chat app. Eligible Claude Max and Team plans now include monthly API credits that must be claimed in the Claude Console; check the current eligibility and credit terms.

Many teams make the same small decision over and over: which category a ticket belongs to, which GL account an invoice line goes to, what type a document is, which team should answer a message. If that's your work, the API wins in five ways.

1. It handles the volume, and it works where your data is

Nobody pastes 40,000 invoice lines into a chat window, and ServiceNow can't ask the chat app about each new ticket. Through the API, the AI can fill a whole Excel file, fill the empty rows in a Google Sheet, or answer every new record that arrives from a system or a workflow tool such as Zapier or Power Automate.

2. It uses your team’s setup every time

Separate chats can use different instructions and context. Two colleagues may get different setups and answers. The decision your team makes ("is this Billing or Account access?") is a convention, and the AI can't know your convention from a category name alone.

When we gave Claude only our category names and a one-line definition of each, it sorted 3,080 real banking support messages correctly 84.7% of the time. A model trained on the team's own examples got 93.7%, with fewer than half the mistakes. Claude's misses were about convention, not understanding: it never once chose one of the team's categories, because only the team's examples showed when that category applied. Read the benchmark →

Through the API, every question can carry the same setup: your categories, your written guide, and the most similar rows your team has already decided. Those past examples show the AI your convention.

3. You can measure how often it's right before you rely on it

A chat answer alone does not establish accuracy on your work. A repeatable evaluation makes mistakes easier to measure. With the API, you can test it first: ask it about rows your team has already decided, compare its answers with theirs, and see the accuracy category by category.

4. You can send the uncertain answers to a person

An AI can answer the same question differently when asked again. Our benchmark found that Claude sometimes changed its category when asked again. Through the API you can ask each row more than once and compare the answers. If the answers disagree, or don't match how similar past rows were decided, the answer is uncertain. Those go to a person to check; the rest go straight through.

5. The company is in control, not one person's account

API use runs under your company's own agreement with OpenAI or Anthropic, which security teams often review and approve separately from chat apps. Usage is billed to the company's account. The work isn't scattered across personal chat histories.

• Examples your team has already decided: a spreadsheet with the text (a ticket, an invoice line, a product title) and the answer your team gave. A few hundred rows is a good start. More helps with categories that look alike.
• An API key from the OpenAI Platform or the Claude Console. Creating one takes a few minutes in the browser and needs no code (step 1 below).
• Something to do the API part for you. This is where a developer usually comes in: writing the code that sends each row with the right setup, collects the answers, measures them, and puts them back in your file. StayCharted AI Model Trainer does that part, so you don't need a developer.

Step 1. Create an API key

For OpenAI (the models behind ChatGPT):

1. Sign in at platform.openai.com. This account is separate from ChatGPT.
2. Under Billing, add a payment method and some credit.
3. Under API keys, create a new key. Name it after what it's for, such as "StayCharted".
4. Copy the key straight away. It's shown only once.

For Claude:

1. Sign in to the Claude Console at platform.claude.com. This is separate from the Claude app.
2. Under Billing, add a payment method or credit.
3. Under API keys, create a key, name it, and copy it straight away. It's shown only once.

Also note the model ID of the model you want to use, from the provider's list of models. Use the exact ID it shows, for example the current GPT or Claude Sonnet model.

At work? Check whether your company already has an OpenAI or Anthropic account, or requires its own. Many IT teams prefer you use the company account, so the usage falls under the company's agreement.

Steps and screen names change from time to time; the providers' own help pages have the latest.

Step 2. Connect it to StayCharted

In StayCharted AI Model Trainer, the workspace Owner opens Workspace settings → AI platforms and chooses OpenAI or Anthropic. They paste the key, and type the model ID under AIs Builders may use. Press Test, and StayCharted asks each model a question to check the key works.

The key is sent once and stored encrypted, and it's never shown again. Requests at once (4 by default) sets how many questions are sent at the same time; lower it if your provider limits how fast your account can call.

Step 3. Upload the rows your team has already decided

Choose New model and upload your spreadsheet: CSV or Excel. Pick the column the AI should read and the column with your team's answer.

StayCharted checks the file before anything is sent:

• duplicates;
• categories with too few examples;
• categories that look like the same thing under two names;
• personal information such as emails and phone numbers, which you can mask in one click.

No answers yet? On Essentials and Business, StayCharted can group similar rows so you can name the groups and start from there.

Step 4. Choose "Use your own AI, without training"

At Confirm and train, choose Use your own AI, without training, then the model you connected. Nothing is trained. With every question, StayCharted gives your AI:

• your categories and your labeling guide;
• the 10 most similar rows your team has already decided.

Step 5. Read the test report

StayCharted keeps part of your rows aside and asks your AI about them as if they were new. Up to 500 held-out rows are asked, and each one is asked three times. The report shows:

• how often it agreed with your team, overall and category by category;
• which categories it confuses;
• how often it kept its answer when asked the same row again;
• for each confidence cutoff, how many rows would go to a person, and how many mistakes that would catch.

If one category is weak, it usually means it needs more examples or a clearer line in the labeling guide. Fix that, and test again.

Step 6. Put it to work

Once you're happy with the report, publish it. The same AI, with the same setup, then answers:

• in files: upload a CSV or Excel file and download it with the answers filled in, each with a confidence;
• in Google Sheets: fill the empty rows;
• through the REST API, for your systems, Zapier, Make, Power Automate or ServiceNow;
• from Claude and ChatGPT, through the AI assistant connector;
• in the Try box, for one question at a time.

Answers below your cutoff wait in the Review Queue. A person confirms or corrects each one. Reviewed answers become examples for the next version you test and publish.

There are two bills, and nothing is hidden in either:

• Your AI provider bills its usage on your account with them: per amount of text sent and received. Every row in a test or a fill is a request, with its examples included, and each row is asked three times by default. Testing on held-out rows is capped at 500 rows for that reason.
• StayCharted: on the Essentials and Business plans, connecting your AI, setting it up and testing it are included. To publish it, so it fills files and answers in Google Sheets, the API and the assistants, add Your AI Models: $49 a month on Essentials, $99 on Business. Compare plans →

If the volume grows large, a model StayCharted trains on your examples can answer the same work without a call to an AI provider for each row. Test both on the same rows and keep the better one.

The chat app is the right tool when:

• it's a one-off look at a few dozen rows;
• you're still working out what the categories should be;
• nobody will rely on the answers next week.

Use the API for work that repeats, at volume, where someone will act on the answer.

StayCharted AI Model Trainer turns your team's past decisions into an AI that makes the next ones, and shows how often it's right before you rely on it. You can:

• use your own AI with Your AI Model (OpenAI, Anthropic, Amazon Bedrock, Microsoft Foundry, Google Cloud or any OpenAI-compatible API);
• or let StayCharted train a model on your examples.

Either way, it's tested on rows it hasn't seen, and uncertain answers go to a person.

Start free → · Your own AI, or ours → · Compare plans →

• Compare Claude API and OpenAI API
• Compare five AI platforms for classification
• Your own AI or StayCharted models

• OpenAI API quickstart
• Claude API quickstart
• Claude subscription API credits