Where is my order, can I return it, does it come in blue. Answered with live stock and order data, refunds issued inside policy.
- resolved without an agent
- 78%
- cost per contact
- −41%
Put an AI agenton every conversation
Synth builds, runs and grades AI agents across chat, voice and email. Every decision is traced, every answer can be replayed, and every rule stays yours to change.
Resolved without a hand-off4,912 conversations today
SynthOnline, replies instantly
- Agent: Hi Maya, I can see your latest order. How can I help?09:41
- Customer: My order arrived damaged. Can I get a refund?09:41
- Agent: Sorry about that. Order #48213 was delivered on Tuesday and is inside the returns window, so I have refunded it.09:42
- Refund issued€64.00 to Visa ending 482109:42
- Customer: That was fast. Thank you!09:42
Running in production at 400+ support teams
- of conversations resolved end to end, with no hand-off
- 71%
- median time to first word on a voice call
- 0.8s
- languages, detected and answered in the same turn
- 38
- from signed contract to live traffic, on average
- 5 wks
One system to build, run and grade every agent
Most teams stitch together a prompt tool, a bot vendor and a spreadsheet of spot checks. Synth is the whole loop, so a fix found in grading is live in minutes.
Write the rules once. The agent follows them every time.
Language models handle the conversation. Flows handle the money. Refunds, cancellations and identity checks run as fixed steps you can read, so the model never improvises policy.
- Persona, tone and escalation rules in one editor
- Knowledge synced from help centres, PDFs and product feeds
- Flows that call your APIs with typed inputs and outputs
- Brand voiceWarm, brief, no jargon
- Knowledge2,418 articles synced 4 minutes ago
- Refund flow4 steps, published as v14
- Escalation rulesDraft with 2 changes
One agent, every channel your customers already use.
The same agent answers the chat widget, the phone line and the support inbox, and hands over to a person with the full context when it should.
- Voice with barge-in and sub-second replies
- Chat, WhatsApp, SMS and email from one configuration
- Warm hand-off to your helpdesk queue, summary attached
Live call
- My train tomorrow was cancelled. Can I move to the 9:40?
- Yes. The 9:40 has seats in coach C at the same fare. Shall I switch you?
- Please, and a window seat if there is one.
- Done. Seat C14, by the window. Your new ticket is on its way by text.
- Perfect, thank you.
- Voice
- 12
- Web chat
- 86
- 30
- 9
Every conversation scored, not a sample of fifty.
A reviewer model scores each conversation against your rubric. Scores roll up by intent, channel and flow, so you know which step to fix before a customer tells you.
- Rubrics for accuracy, policy and tone, weighted your way
- Regression tests replayed on every change
- Alerts when a score drifts past its threshold
Overall score
97.8+3.0 across 4,912 conversations
- Accuracy
- 97.8
- Policy
- 99.2
- Tone
- 92.1
Every answer arrives with its working
Open any conversation and replay it step by step: what the customer meant, which tool was called, which rule allowed the action, and what the model was given to write the reply.
- of steps logged, tool calls and rules included
- 100%
- of full traces kept, longer on Enterprise
- 90 days
- to replay a conversation on a new flow or model
- 1 click
Synthtrace · 446 ms
My order arrived damaged. Can I get a refund?
Sorry about that. Order #48213 was delivered on Tuesday and is inside the returns window, so I have refunded it.
How this answer was made
- Intentrefund_request, confidence 0.97
- Order lookup#48213, delivered, €64.00
- KnowledgeReturns policy §2, damaged on arrival
- RuleRefund under €200, allowed
- ActionRefund RF-90412 issued
- ReplyDrafted in brand voice
- ChecksPII and policy scan clean, sent
Bring any model. Keep the stack you have.
Pick a model per step, not per contract. Route simple intents to a fast model and hard ones to a frontier model, and switch either without rebuilding the agent.
Model routing
- Classify intentfastOpen-weight 8BYour fine-tune v3
- Extract fieldsfastYour fine-tune v3Open-weight 8B
- Compose replystandardFrontier model AFrontier model B
- Transcribe voicespeechSpeech model S2Speech model W4
- Grade conversationcarefulFrontier model BFrontier model A
Regression suite passes on every swap
Connected systems
- HelpdeskTickets and warm hand-offs
- CRMCustomer profiles and history
- CommerceOrders, stock and returns
- PaymentsRefunds and disputes
- Data warehouseTraces and scores, exported
- IdentityVerification and SSO
Built around the work, not the chat window
Each industry ships with flows, rubrics and integrations for the conversations that fill its queues.
Card freezes, payment status and disputes, with step-up verification before anything touches an account.
- faster dispute intake
- 3.2×
- of actions logged
- 100%
First notice of loss taken by phone at 2am, photos requested by SMS, and the claim filed before the adjuster logs on.
- saved per claim
- 11 min
- intake coverage
- 24/7
Outage lookups by postcode, plan changes and device troubleshooting that knows which router the customer owns.
- fewer repeat calls
- 64%
- voice response
- 0.7s
When a storm cancels two hundred flights, every passenger is offered a rebooking in minutes instead of a hold queue.
- peak capacity
- 9×
- points of CSAT
- +22
Appointment booking, reminders and pre-visit questions, with clinical topics routed to staff by rule, never by guess.
- no-show rate
- −35%
- aligned deployment
- HIPAA
Teams that stopped sampling and started reading every conversation
parcelryPriya RamanVP Customer Operations, Parcelry−38%handle time in the first quarter
“We used to review fifty tickets a week and hope they were typical. Now every conversation has a score, and the first thing we fixed was a flow we thought was fine.”
Harbor BankJonas LindqvistHead of Digital Service, Harbor0policy breaches in eleven months
“Our compliance team signed off because they could read the refund rules themselves. The model writes the reply; it never decides the money.”
vela airSofia AlmeidaDirector of Customer Care, Vela9,000rebookings on one morning
“Storm day used to mean a four-hour hold queue. Last January the agent rebooked nine thousand passengers before our morning shift arrived.”
northwindMarcus BellPrincipal Engineer, Northwind1 dayto change model providers
“Switching the model behind our triage step took an afternoon and a regression run. With the last vendor it would have been a new contract.”
1 / 4
Ready for the teams who get audited
Synth runs in your region or your cloud, keeps customer data out of model training, and records every action with who approved the rule behind it.
Data residency
EU, US or UK hosting, or a private deployment in your own cloud account.
Single sign-on
SAML and OIDC with SCIM provisioning, and roles down to a single flow.
PII redaction
Card numbers, IDs and health data masked before any model sees them.
Guardrails
Policy checks on every reply, with blocked answers routed to a person.
Audit log
Every rule change, deployment and action, exportable to your SIEM.
No training on your data
Conversations are never used to train shared models. In writing.
- ISO 27001
- SOC 2 Type II
- HIPAA
- GDPR
Synth meets the requirements for ISO 27001, SOC 2 Type II, HIPAA and GDPR.
Questions teams ask before they sign
Which language models does Synth use?
Whichever you choose. Synth works with the major frontier providers, open-weight models you host, and your own fine-tunes. You can route each step of a flow to a different model and change it without touching the agent.
How long does a typical launch take?
Five weeks on average from signature to live traffic. The first week connects knowledge and systems, the next three build and grade flows against real transcripts, and the last runs a staged roll-out.
What happens when the agent cannot help?
It hands the conversation to your existing helpdesk queue with a summary, the customer's details and everything it already tried, so nobody has to ask the same question twice.
Is our data used to train models?
No. Conversations stay in your tenant and are never used to train shared models. Enterprise plans can keep all processing inside your own cloud account.
What counts as a resolved conversation?
One where the customer's issue was closed without a hand-off and they did not return about the same issue within 72 hours. Hand-offs are never billed.
Notes from the team
All posts
Why our refund flow never asks the model for permission
Language models are good at conversation and bad at policy. Here is how we split the two, and what it did to our error rate.
Grading every conversation: what changed when we stopped sampling
Reviewing fifty tickets a week felt rigorous. Scoring all of them showed us how much the sample was hiding.
Sub-second voice: where the 800 milliseconds go
People notice a pause on the phone at around a second. This is the latency budget we work to, stage by stage.
Put Synth on your busiest queue
Bring a week of real transcripts. We will show you how many the agent resolves, and exactly how it got there.
- to live traffic
- 5 weeks
- resolved end to end
- 71%
- policy breaches
- 0