PRODUCT
Use the decisions your team already made. StayCharted can train and answer, your company’s AI can answer with your past rows as examples, or Amazon Bedrock can train in your own AWS account. All three are tested against held-out rows and support the same review workflow.
THREE OPTIONS. ONE WORKSPACE.
Scroll across to compare all three options.
| What changes | Trained and answered by StayCharted | Your AI Model | Dedicated AI Model in your own Amazon Bedrock |
|---|---|---|---|
| What it is | Classifier, AI Classifier or Dedicated AI Model (and the picture models), trained on your examples and answered by StayCharted | Your company's own AI, given your categories, your labeling guide and the most similar rows your team already categorized with every question | A Dedicated AI Model trained by Bedrock's own fine-tuning in your AWS account |
| Is anything trained? | Yes | No | Yes, by Bedrock |
| Plans | Every plan (Dedicated AI Models on Business) | Essentials and Business: set up and test included; publishing needs the Your AI Models add-on | Business |
| What you need | Your labeled examples | Your labeled examples and an AI platform: Amazon Bedrock, Microsoft Foundry, Google Cloud, OpenAI, Anthropic, or any OpenAI-compatible API | Your labeled examples and an AWS account with Amazon Bedrock in US East (N. Virginia) |
| Where the model lives | In your StayCharted workspace | Nothing is trained, so there is no model to keep: your AI stays on your platform | In your AWS account, as a Bedrock custom model |
| What answers each row | StayCharted's servers | Your AI, on your platform. Each row is asked three times by default, and the answers are compared | Your trained model, in your Bedrock |
| Pictures | Yes (Vision add-on) | Text only | Text only |
| Who pays for computing | Included in your plan | Your AI provider bills its own usage, on your account with them | Training and answers are on your AWS bill |
| External calls per answer | Answered on StayCharted’s servers | One or more calls to your AI platform | A call to your Bedrock for each row |
| Can you take it with you? | Dedicated, AI Classifier and AI Image Classifier Models can be downloaded from Manage | There's nothing to take: the AI is already yours | It's already in your account; your own systems can call it directly |
| Tested before anyone uses it | Yes | Yes | Yes |
| Review Queue | Essentials and Business | Essentials and Business | Business |
| Counts toward your plan as | Its own kind | An AI Classifier Model; rows it fills count like any fill | A Dedicated AI Model |
Which does better on your data? Set up or train more than one on the same rows. Each is tested on the same held-out rows, so the reports compare like for like. Publish the one that makes the fewest mistakes that matter.
WHEN TO CHOOSE WHICH
The default. Nothing to set up beyond your examples, with computing included in the plan. Start with a Classifier or AI Classifier Model; try a Dedicated AI Model on Business when categories depend on context or subtle wording.
For a company with an approved AI platform, or when you want to test an option before you have many examples. Nothing is trained: your AI receives your categories, labeling guide and similar past rows with each question.
Quality depends on your AI and those examples. Our assistant benchmark provided category descriptions, not these past rows. Test on your own data before publishing.
For a trained model that stays in your AWS account and can also be called directly by your own systems. On Business, Bedrock trains Amazon Nova Micro, Lite or Pro in US East (N. Virginia).
Training on Google Cloud and Microsoft Foundry is not available yet.
THE SAME REVIEW WORKFLOW
No model gets every row right. Every answer comes with a confidence; answers below your cutoff go to review. What your team decides helps the next version, whether it is trained or uses your AI with examples.
Each model gives one with every answer, between 0 and 1. It means slightly different things for each kind (for Your AI Model it reflects how consistently your AI answered and whether the most similar past rows agree), so each model's cutoff is chosen from its own report.
When a model is tested, its report shows what each cutoff would do: how many rows would go to review, how many mistakes that catches, and projected accuracy after review, assuming flagged errors are corrected. Each kind starts with a sensible default, and you can change it.
Answers below the cutoff come into one Review Queue, whichever way they were asked: filled files, the Try box, the Prediction API (including Zapier, Make and Power Automate), Google Sheets, and Claude or ChatGPT. The same question asked twice is reviewed once.
A Reviewer sees the answer and similar rows your team already categorized (for a trained model, its other likely answers too), and confirms it, corrects it, or marks it as not enough information. One key per category, Enter to confirm.
Reviewed answers, corrections first, join the rows the next version learns from: up to a quarter of a retrain, with the rest kept for the one after. For Your AI Model nothing is trained; reviewed answers join the past rows it is shown. For a model in your Bedrock, the next fine-tune runs there.
A Reviewer can start the retrain from the queue. The new version is tested like the first, and a Builder decides whether to publish it, so the model never changes underneath you.
A filled file can be downloaded with the reviewed answers in place of the model's.
The Review Queue is on Essentials and Business. Up to 2,000 answers wait per model; unreviewed answers are deleted after 30 days by default. Reviewers and Builders can get a daily summary.
The goal isn’t a model that never needs a person. It’s a model that needs one less often, version after version.
Yes. Your AI Model's uncertain answers go to the same queue, and what reviewers decide joins the past rows it is shown.
Yes. Set up or train another kind on the same model and compare the reports; publishing chooses which one answers. Your files, Try box, API calls and Sheets keep working, because they call the model, not the kind.
Yes. AI Model Trainer still holds your training files and filled files, and sends rows to your AI platform for answers. What changes is who answers, not where your files are kept.
EXPLORE NEXT