ReplayAutonomous app testing for teams who ship at AI-speed
Promoted
Typewise is an AI agent platform that gives every customer a first-class concierge across service, sales, delivery, billing and retention. Its agents remember context, communicate across channels and resolve requests end to end by taking action in the systems you already use. Typewise Nova runs, tests, monitors and improves the AI operation, while your team controls what it can do, what needs approval and when people take over.
This is the 7th launch from Typewise. View more
Typewise Nova
Launched this week
Nova is the AI Operator that builds and improves your AI customer experience team, no developer required.
Describe in plain language how customers should be handled and connect your tools. Nova builds the agents, tests them on past tickets and shows you what failed before anything goes live. Once running, it monitors quality and proposes fixes 24/7 for approval.
Agents resolve whole requests, from orders and refunds to plan changes, across email, chat and WhatsApp. Unresolved requests are free.
Free Options
Launch Team / Built With
For me full automation feels a bit risky, tell me how well does it actually handle weird edge cases without making mistakes?
Typewise
@manjesh_yadav1 totally, that's why our idea is blending the best of human skills and AI skills. It's AI-first, so you have the AI supervisor taking on every single customer request (= fast response, 24/7, any language) but it then hands off (fully, or partially) to a human when needed or desired (e.g. as defined in hand-off conditions). all managed in a shared inbox, so it's super easy to collaborate for humans + AI.
I am curious how it handles duplicate entries if a customer emails and chats at the same time.
Typewise
@sandy_leventh it recognizes that and merges the ticket; it can of course only recognize this if: customer is logged in (e.g. shop / portal) or shares their name/email/ID on the chat, and you can match that with the email address, or something similar depending on your setup. how are your customers identified?
How smoothly does it maintain that context when conversations involve multiple agents?
Typewise
@jason_scott8 the AI Supervisor always manages that, and involves various AI Specialists that step in and out. That also works when a conversation switches channels, e.g. you start in WhatsApp and switch to email.
Congrats @davideberle Can Nova help improve an existing AI support agent, or is it mainly for new agents?
Typewise
@barnaby_lloyd what existing AI support agent are we talking about? Nova wouldn't configure your existing ticketing system. But, Nova could read out your existing Zendesk/Intercom setup as a starting point, and create Typewise AI Specialists out of that. Those specialists then get continuously improved by Nova, or new ones might get added, etc.
This could save teams a lot of time and consultant costs. How long does it usually take to set up Nova?
Typewise
@easton_carter easy case is 15min :-) say for a start-up with 2-3 systems to connect (CRM, Payment, Meeting Booker, Core System) via MCP you probably need ~1h to hook them in, and spend another ~1h with Nova to set up all the skills and first test cases. Channels are then set up in another 15min, or ~1h in case you need proper authentication (login, magic link, etc.) for sensitive customer issues. So say for an SMB/Start-up you're live and tested same day.
In what manner does Nova evaluate the agents with respect to old tickets? Is it by playing back real conversations or simulating new conversations?
Btw, Congrats @davideberle & team on the Excellent launch of Typewise Nova. 🚀✌️
Typewise
@aymi_malik you could task it to back play, but usually you'd use past tickets to determine what AI Specialists shall be set up, and then the automated testing will run using simulated conversations (to ensure variability and avoid over-optimization)... where do you handle tickets currently?
how do you keep tests on past tickets from overfitting to the old way the team handled them
Typewise
@mattkilmer the tests don't run on past tickets but on simulated conversations (multiple variants). based on past tickets you (or Nova) define the test cases, which then contain a scenario and an expected outcome (with optional assertions of what tools must/must not be called). but then simulations create variants of this. also the results don't automatically adapt your AI specialists, but if tests fail Nova will analyze why, and will suggest adapting the instructions if that's the best course of action.
@davideberle that separation makes sense, do you version the expected outcome and tool assertions when the policy behind a test changes
Typewise
@mattkilmer good question, right now you'd update the test and the versioning gets logged, but a visible versioning should come in the future