The layers at a glance
No single mechanism prevents a wrong answer. Typewise puts several between the model and the customer, and you decide how many stay on.
| Layer | What it does | When |
|---|---|---|
| Grounding | Answers come from the knowledge you connect and live lookups in your systems, not from the model’s memory. | Every request |
| Verified lookups | Customer data is read only when the customer input is a verified value; otherwise the lookup does not run. | Every lookup |
| Review agent | A separate agent checks each outgoing response against your guidelines before it is sent. | Every reply |
| Hand off conditions | Situations you define per specialist where a person takes over instead of the AI guessing. | When they apply |
| Human review | Every AI reply on email and WhatsApp waits for a colleague to approve, edit or reject it. | While you choose |
| Tests before go live | Policies, simulation and evaluation validate a change before it ships. | Every change |
| Monitoring | Quality, flagged replies and guardrail events reported in the Insights Hub; wrong answers are not charged. | In production |
Grounded in your sources, not in the model’s memory
The AI answers from the knowledge you add and the lookups you connect: your website, help center articles, uploaded documents, and live data from systems such as your order, booking or billing system. Customer specific facts come through lookups rather than from the model, so an order status is read from the order system at the moment the customer asks.
Lookups can be limited to verified values, such as a customer signed in to your portal, an email address confirmed with a one time code, or the phone number on WhatsApp. Without a verified value the lookup does not run, and in an anonymous chat the AI answers from public knowledge only.
Each knowledge source carries an access level. Material marked for internal use helps your team without being shared with customers. More on answering only from your knowledge base.
The review agent
Typewise is built as a team of agents rather than a single model. The AI supervisor analyses each request and decides whether to route it to a specialist agent or to a human. Specialist agents handle one topic each, with knowledge agents retrieving information and action agents executing actions within guardrails.
A separate review agent checks every outgoing response against your protocols and company guidelines before it is sent. It is a second pair of eyes on the reply, independent of the agent that wrote it.
Hand over instead of guessing
You give each specialist hand off conditions: the situations where a person takes over. A typical example is a request where the data the AI needs is missing or outdated, such as a parcel with no tracking update, so it hands the case over rather than guessing a date. Other common conditions are highly complex situations, upset customers, and decisions your policy keeps with people, such as claims, goodwill or commercial remedies.
A hand off carries the conversation history, the customer and order data and a summary, in the Typewise Inbox or the helpdesk your team already uses. Once a colleague has answered, the AI continues the conversation. See ticket management.
Human review while you build trust
On email and WhatsApp you can turn on human review, so every AI reply waits until a colleague approves, edits or rejects it. Any action in a connected system is configured separately, and approval by a person is on by default.
You can keep human review on for every topic and switch it off specialist by specialist as the replies prove themselves. How the approval settings work together is on human approval and automation levels.
Tested before it goes live
Built in policies, simulation and evaluation validate every change before it goes live. You can run test conversations that check which specialist answers, which lookups run and whether the case is handed over, and Simulator tests check the expected behaviour again after every change to knowledge, instructions or models.
Nova, the AI Operator, tests each change before it ships and drafts fixes for you to approve. Its changes stay drafts until someone on your team publishes them.Watched in production
The Insights Hub monitors response quality and policy adherence across every interaction. A content safety view lists flagged replies, blocked topics and every case where a guardrail held the AI back, and an outcome view shows where conversations escalate to human support.
Agents can rate conversations, and an Unrated filter finds the ones still waiting for a quality check. A message or action your team flags as incorrect is not charged, and the flag feeds the evaluation loop so the same mistake is less likely next time.
Models and data handling
AI model inference runs through Microsoft Azure OpenAI Service, OpenAI and Google Vertex AI, in the European Union, with zero data retention: the providers do not store prompts at rest. Your customer data is not used to train AI models for other customers, and PII masking can automatically mask sensitive data such as credit card numbers before the AI processes it. Details, subprocessors and documents are in the trust center.
Where to read next
The platform end to end is on how it works. Security, hosting and compliance are in the trust center. Plans and the per resolution rate are on pricing, and the free trial runs on your real tickets for 14 days with no credit card.