LEADS · WEBSITE MESSAGES · SHORT EMAILS
Answer every new message the way your company always has.
Your team has answered thousands of leads, website messages and short emails. Which reply went to whom says a lot about your company: who gets a same-day call, who gets the self-serve link, which partner enquiries matter. StayCharted AI Model Trainer trains a private model on those decisions to suggest the reply your team would choose for each new message.
Trained on your reply choices · Approved reply types · Human review
Not generated. Chosen.
Generative AI can draft a new answer, but that also means reviewing its wording and any promises it makes. StayCharted doesn't write replies. It learns which of your own replies your team used for which kind of message and sender, and picks one for each new message, with a confidence.
- Your voice, every time: every reply is one your team already wrote and uses.
- Nothing made up: no invented prices, dates or promises.
- A person decides the rest: answers below your model’s review cutoff go to your team. Route unfamiliar or sensitive messages for review in your workflow too.
What it can learn
| A new… | It suggests |
|---|---|
| Inbound lead from your website | The reply (book a demo, self-serve trial, partner team, not a fit) and who sends it |
| Message from a key account | The reply your team gives those customers, and the owner |
| Short support email | The saved reply your agents would use, and the queue |
| Partner or press enquiry | The right reply and the person who handles it |
How it works
STEP 1. Export past messages with the reply your team used. From your help desk or CRM: each message, the saved reply, macro or template that was used, and who it was from (company, account tier, lead source). Use saved-reply or template identifiers from your help desk or CRM export; include only fields available to your workflow. Replies typed from scratch? On Essentials and Business, start from unlabeled examples and group them into reply types for your team to name and review first.
STEP 2. Check your examples. StayCharted points out duplicates, reply types with too few examples, and personal information such as email addresses and phone numbers. Mask them in one click.
STEP 3. Train and test. See how often the model picks the reply your team chose, on messages it never saw, and which replies it mixes up.
STEP 4. Use it on new messages.
Through the API (Zapier, Make, Power Automate or your own code), a file, or Google Sheets.
Each message gets the reply type’s name, the owner if you trained one, and a confidence. API answers also report whether the confidence is below your review cutoff (needsReview). Your help desk or automation maps the name to your approved template text and sends it. StayCharted does not write or send the reply. API access starts with Essentials; the Sheets add-on uses your plan’s allowances.
ALWAYS. A person checks the unsure ones. On Essentials and Business, answers below your model’s cutoff wait in the Review Queue. Your team picks the right reply, and the next version learns from it.
Why it has to be yours
Every company decides differently who is strategic, what's urgent and how to say no. A general AI assistant doesn't know that your largest customer gets a call from the founder, or that a student enquiry gets the education page. Your past replies do. That's what the model learns, and it stays private to your workspace.
Frequently asked questions
Does it write or send replies?
No. It picks one of the replies your team already uses and says how sure it is. Your help desk, CRM or automation sends it.
What if a message needs a new kind of reply?
The model can only choose from reply types it learned. Low-confidence answers go to a person, but unfamiliar messages can still receive high confidence. Route new or sensitive cases for review in your workflow. Add the new reply type to your labeled examples, and retrain so the next version can use it.
How does it know who's a key account?
Only if your examples say so. Include the sender's account tier, company or lead source as columns, and the model learns how your team treats each.
How many examples do I need?
A few hundred past messages, with about 20 or more for each reply type, is a good start.
Is my data used to train anything else?
No. Your examples train only the models in your workspace.
Explore next
Support ticket classification → · How it works → · Connect through the API → · Compare plans →