Company / Security & trust
Security & trust

Where your data runs, who approves what, and what we don't have yet.

This page is written for the person who has to sign off on Settl. It says where conversations and order data live, which outside services see them, how money is gated, what is recorded, and — plainly — which certifications and controls Settl does not have today.

Reply policy · per channelyou set it
Autonomous
Replies directly, hands off when it should
Chat, WhatsApp, email, voice, app
Draft
Writes the reply; a person approves it
Public posts on X
Copilot
Drafts a private note for your team
Dealer portal, internal
Triage
Never replies — classifies, links, escalates
Legal notices, consumer forums
Proactive
Reaches out first on the customer's channel
Vehicle and device alerts
Self-hostedthe whole stack on your infrastructure, if you choose
1 outside callthe AI model provider you pick — Gemini, Anthropic or OpenAI
Every writeidentity-checked, policy-gated, on the record
0certifications held today — read on
Deployment

Your infrastructure, or ours. Either way, one network.

Settl is a set of services — the inbox, the agent, the database, the voice worker — that run together on a single host or cluster. We can run them for you, or deploy them inside your own cloud account or data centre, where conversations, order lookups and the knowledge base never leave your network. The inbox is the open-source Chatwoot helpdesk; the database is Postgres. There is no shared multi-tenant store of your customers' conversations on our side when you self-host.

Traffic to the web app and the inbox is over TLS; the app host sits behind a single signed login, and the only path outside it is the voice signalling socket, which is authenticated by a per-call token.

Case lifecycle10 steps · one record
01
Intake
Arrives on any channel, in that channel's shape
02
Identify
Phone, email, handle or VIN → the customer
03
Enrich
Live order, vehicle or account read automatically
04
Deduplicate
Same customer, same issue → the master case
05
Classify
Category, sentiment, priority, SLA
06
Decide
Next best action from your rules
07
Act
Reads auto-run; writes wait for approval
08
Communicate
Reply in the customer's own channel
09
Close
Checklist complete, closure record drafted
10
Grade
Scored by the judge, fed to analytics
What leaves

Exactly one kind of outside call: the model.

The agent sends the text it needs answered — the customer's message, the retrieved policy passages, the live record it read — to the AI model provider you choose. That provider is a setting: Gemini, Anthropic or OpenAI, under your own account and your own data-processing terms with them. Switching provider is a configuration change; the retrieval, the actions and the scoring are the same whichever is behind it. For voice, the speech provider is likewise your choice. Nothing else about a conversation is sent to any third party.

Caller stops → agent speaksmeasured
Settl · speech-to-speech0.9–1.2 s
STT → LLM → TTS cascade~1.65 s
Typical IVR bot3 s+
Both stacks ship. Speech-to-speech is the default; the cascade (VAD → Deepgram → LLM → Sarvam voices) is one setting away.
Money and actions

Nothing moves money without a person — and your backend has the last word.

Every action is either a read or a write. Reads — order status, rider position, vehicle telemetry, a payment record — run automatically. Writes — refunds, credits, cancellations, reschedules, remote vehicle commands — go through a policy with a cap you set, and an identity check first: the contact on the conversation must match the customer on the order. Below the cap the agent acts; above it the action waits in an approval queue for a named person. Your own system then enforces its rules on top, so a delivered order cannot be cancelled and a refund cannot exceed the order, whatever the agent proposed.

Handoff briefing · pinned to the conversationhuman assigned
low confidence · 0.41intent · damaged itemsentiment · angry
Why you're seeing this
Asking for a refund above ₹50 — that needs a person.
What they want
Full refund for a cracked blender jar, order ST-51840.
What's already known
Delivered yesterday 7:42 PM · photo attached · first complaint.
What I told them
That a colleague will confirm the refund within the hour.
Status → open · auto-assigned to the next available agent
Grounding

Answers cite their source, and an unsure agent says so.

Replies are built from passages retrieved from your own policy documents and carry those citations. Below a confidence threshold the agent does not guess; it hands the conversation to a person with a written briefing. Every conversation is then scored by an independent judge against your scorecard — identity verified, policy applied, escalation correct — and a regression in simulation blocks a release.

Closure checklist1 item open
✓Issue identifiedCharging / charger handshake
✓Root cause capturedCharger board fault, CHG-201
✓Action takenPickup booked · board replaced
✓Customer communication6 messages across 3 channels
!Customer confirmationNot yet captured
✓Warranty / partsBoard under warranty
✓Follow-upNothing outstanding
“Customer confirmation has not been captured. Please complete before closing the case.”
What is recorded

Who did what, when, on whose approval — the AI included.

Each case carries a timeline of every message, lookup, action and approval. CRM records keep field history, score history and dealer-assignment history with the user who made each change, and changes made by the AI are recorded the same way. Setup changes — rules, caps, policies — are logged with who changed them. Calls are filed with their transcript.

SLA watch · predicted, not reported3 open
AV-2077Battery · Punehigh · 78
82% of window used · followed up twice · Service backlog 9
AV-2063Charging · Ahmedabadmedium · 52
Similar cases take ~30 h; 18 h left
AV-2058App loginlow · 21
Assigned · inside typical time
High risk → reassign or escalate to the supervisor before the breach.
Being straight

What Settl does not have today.

A security review should know these before the first call.

  • No SOC 2, ISO 27001 or similar certification. Settl is a small company; a formal audit is planned, not done.
  • No automatic PII redaction before text goes to the model provider. The provider sees the customer's message as written; choose the provider and terms accordingly.
  • Encryption at rest is your infrastructure's. When self-hosted, disk and database encryption are whatever your cloud or data centre provides; we configure to it, we do not replace it.
  • No published penetration test yet. We will share the results of the first one with customers under NDA.
  • Single sign-on is not yet supported for the agent workspace; access is by individual login.
Ticket vs casesame job, different depth
TicketCase · how Settl works
What it isOne request and its threadOne customer problem, however many requests it takes
Linked toA requesterThe customer, their account, order or vehicle
ChannelsUsually oneAll of them, merged into one timeline
LifespanMinutes to hoursHours to weeks, with tasks and hand-offs
Who uses the wordZendesk, Freshdesk, HubSpot, IntercomSalesforce, Dynamics 365, ServiceNow
In short
Hosting

Self-hosted or managed

Docker services on your cloud account or ours; Postgres and Chatwoot underneath.

Access

One signed login per workspace

Agent, customer-demo and inbox surfaces behind sign-in; voice joins by per-call token.

Identity

Checked before every write

The conversation's contact must match the order's customer before anything changes.

Approvals

A named person above the cap

Refunds, credits, cancellations and vehicle commands wait in a queue with the evidence.

Audit

Timelines and field history

Every case step, every CRM change and every setup change, with who did it.

Exit

Your data in a standard database

Conversations in Chatwoot's Postgres and cases in Settl's, exportable with ordinary tools.

Questions

Frequently asked

Can Settl run entirely inside our network?

Yes. All services — inbox, agent, database, voice — can be deployed in your own cloud account or data centre. The only outside call is to the AI model provider you choose, under your own account with them.

Which AI providers see our customers' messages?

Exactly one, the one you configure: Gemini, Anthropic or OpenAI. The agent sends the message, the retrieved policy passages and the live record it read. There is no automatic redaction today, so the provider's data terms matter; we will help you pick.

Who can approve a refund?

Anyone you give the agent workspace login to; the approval is recorded with their name on the case timeline. Caps and the list of actions that need approval are edited in Setup, and those changes are logged too.

Is Settl SOC 2 certified?

No, and this page says so rather than implying otherwise. If a certification is a hard requirement for procurement, Settl does not meet it today.

How do we leave?

Your conversations are in Chatwoot's Postgres database and your cases and CRM records in Settl's; both export with standard database tools. Policy documents and adapters are yours.

GET STARTED

The queue that empties itself.

Give us your three worst intents and two weeks. We'll run them against your real tickets and show you what closes itself.

↗A working agent on your data, not slides
#Grounded on your policies — cited or handed off
✓Human approval on every action that moves money
TYPICAL RESPONSE · WITHIN ONE WORKING DAY
We use these details only to reply. Privacy policy