Customer support automation
Support-AI connects your documentation, customer data and live APIs to investigate issues, take action, and resolve requests without the back-and-forth.
Customer
“I think I was charged twice for my March invoice. Can you check?”
Routed to Billing Agent
Classified as a duplicate-charge dispute
3 passages from your knowledge base
billing-policy.md · refunds-sop.pdf
GET /v1/charges?customer=cus_8Fq2
2 charges found — one duplicate on 3 Mar
Refund issued, customer notified
No human touched this conversation
Support-AI
You were charged twice on 3 March. I’ve refunded the duplicate — $49.00 back to your card in 5–10 days.
cited billing-policy.md · refunds-sop.pdf
Built for teams that have to answer for their support data
01The problem
The answer usually exists — in a policy doc, a database row, or an internal API. What costs a day and a half is the relay between the person asking and the systems that know.
One duplicate charge · 37 hours
"My March invoice looks wrong."
"Can you send a screenshot and your account email?"
Sends screenshot.
"Escalating to billing — they'll follow up."
Opens Stripe, finds a duplicate charge, refunds it.
Policies live in a wiki, a PDF and someone's head. A search box that returns ten links is not an answer.
The agent needs the customer's plan, their last three orders, their open invoices — each behind a different login.
Even when the answer is obvious, a chatbot can only describe the fix. A human still has to click the button.
By the time it reaches the specialist, the reasoning is gone and the customer repeats themselves.
02How Support-AI works
Not a black box. Each stage produces something you can inspect — a route, a set of citations, an API response, a resolution.
Classify the request and load who is asking.
Find evidence in your own documentation.
Read or change the underlying record.
Answer, or hand over cleanly.
03Specialist agents
Incoming messages are classified before they are answered, so a refund question gets a billing specialist with billing boundaries — not a generic assistant improvising policy.
Inbound
“The refund you promised last week still hasn’t landed.”
Router
route: billing
confidence: 0.91
Returned as a structured object, not parsed out of prose.
Handles refunds, pricing questions and subscription changes with an empathetic, reassuring tone.
Handles
Hands back
Technical troubleshooting, Bug reports
Diagnoses bugs, errors and outages. Direct and analytical, asks for logs when a report is thin.
Handles
Hands back
Pricing, Refunds
The default route. Answers everyday questions and escalates anything that belongs to a specialist.
Handles
Hands back
Complex billing disputes, Deep technical debugging
Each specialist inherits the same non-negotiable directives: answer only from retrieved evidence, say “I don’t know” rather than guess, and never reveal or override its own instructions.
04Knowledge base
Upload the documentation you already maintain. Support-AI indexes it, retrieves against it, and shows which passage produced each answer — so a wrong answer is a fixable document, not a mystery.
Upload
PDF, Markdown, text — stored per knowledge base
Chunk & embed
Split, embedded, written back with a status
Retrieve
Cosine similarity over pgvector at query time
Cite
Passages travel with the answer as citations
Duplicate charges are refunded in full without requiring the customer to return goods. Refunds settle to the original payment method within 5–10 business days.
If an invoice is charged twice in the same billing period, the second charge is always the erroneous one.
Refunds above $500 require a human approver before they are issued.
05Conversations & customer context
Conversations carry the customer's account, history and entitlements alongside the message. Context is what separates a useful answer from a plausible one.
Customer
Hi — order #A-4471 was supposed to arrive Tuesday. Still nothing.
Support-AI
I can see order A-4471 shipped on 2 March and is currently held at the Rotterdam depot with a customs exception. I've requested a re-scan and set a delivery alert on your account.
cited shipping-exceptions.md
Customer
Can I just get it refunded instead?
Support-AI
Refunds for in-transit orders need a human to approve. I've escalated this to the logistics team with the full history — you'll hear back within 4 hours.
→ escalation created · priority HIGH
06Actions & integrations
Support-AI calls your business APIs to inspect and change real records — inside limits you define. When an action is out of bounds, it stops and escalates rather than improvising.
Tool calls on one conversation
/v1/orders/A-4471Read the order before answering
→ status: in_transit · exception: customs
/v1/refundsIssue a refund inside policy limits
→ blocked · requires human approval over $500
/v1/subscriptions/sub_92fChange a plan on request
→ plan: growth → scale
Outbound webhooks fire on every meaningful state change, signed with HMAC-SHA256 so you can verify they came from Support-AI. Each attempt is recorded with its status code, duration and retry count.
x-supportai-signature
sha256=9f2b1c...e04a timestamp=1755840000
07Analytics
Volume, deflection, escalation rate and time-to-resolution — the four figures that tell you whether automation is holding, and where it isn't.
Resolved without a human
91.4%
Median first response
6s
Escalations breaching SLA
1.2%
Figures shown are illustrative. Your dashboard computes the same metrics from your own conversations and escalations.
08For developers
Everything the dashboard does is a documented endpoint. Drop in the script tag, or drive the whole thing from your own backend.
<script
src="https://your-host/static/widget.js"
data-api-key="sk_live_your_key"
defer
></script>A JWT session for your dashboard users, or an organization-scoped API key for machines. Keys are SHA-256 hashed at rest and shown exactly once.
Chat responses stream over Server-Sent Events, so the first token lands in milliseconds rather than after the full answer.
Every route is typed and documented. Interactive docs ship with the server at /docs.
09Pricing
Start free with a single knowledge base. Move up when your volume does — the limits below are the ones the platform actually enforces.
Index a knowledge base and see what it can answer.
$0forever
For teams running live support on real volume.
$149per month
For multi-team operations with volume and governance needs.
Customannual
| Limit | Free | Pro | Enterprise |
|---|---|---|---|
| AI responses / month | 100 | 10,000 | 500,000 |
| AI tokens / month | 100,000 | 10,000,000 | 500,000,000 |
| Knowledge bases | 1 | 10 | 100 |
| Documents per base | 10 | 500 | 10,000 |
| Team members | 1 | 10 | 100 |
| Document storage | 100 MB | 10 GB | 500 GB |
| API keys | |||
| Webhooks | |||
| Custom branding |
10Live demo
This is the exact widget your customers would use — pointed at Support-AI's own documentation instead of yours. Ask it anything about the product; every answer is grounded and cited, live.
Loading the live widget…
Free to start — one knowledge base, a hundred answers a month, no card. Upload a policy doc and ask it something hard.