Howzer v2.4.0: from reading messages to understanding customers
Our biggest release. Every analysis that used to run on hand-written rules now runs on models trained for German customer feedback, and around them sits a full picture of the customer: where they are in their journey, where they are heading, and what to do about it.
By Howzer Team, Product · product ·
Up to v2.3, Howzer answered one question very well: what is in this message. v2.4.0 answers the harder one: what does this message mean for this customer, right now. Everything below exists to serve that shift.
Judgement instead of keyword lists
Every analysis that matters is now handled by a model trained specifically on German customer feedback rather than by a list of words and rules. The difference shows up on exactly the messages that used to be hardest: sarcasm, indirect complaints, politeness covering real anger.
- Urgency: how quickly this needs a human, from routine to drop everything.
- Risk: how likely this customer is to leave, escalate, or involve a lawyer.
- Intent: what the person actually wants. A refund, a cancellation, information, or simply to say thank you.
- Topic: which part of your business this belongs to, so it reaches the right desk.
- Emotion: several feelings at once, because messages rarely carry just one.
- Root cause: why they are writing, not just what they wrote.
The journey, not the snapshot
A message means something different depending on who sent it and when. Howzer now tracks where a customer stands, notices when they move, and suggests the next sensible step in German.
On top of that sits prediction: which customers are drifting towards the exit, which patterns look unusual for this time of year, and which groups behave alike. Early enough to still do something about it.
Answers without a query language
Ask in plain German. Show me all critical complaints from this week. You get structured results back. There is nothing to learn and nothing to configure, and it works on your own data, on your own machines.
Replies that quote your own rules
Your internal policy documents are indexed. When a question arrives that a policy already answers, the suggested reply points at the relevant passage, so two agents on two different days give the same answer.
Weekly summaries, written for management
German executive summaries are generated automatically each week. All of it runs on your own infrastructure, no data leaves the building, and if you would rather use your own language model you can plug it in.
Changing models without holding your breath
A new model can run quietly alongside the live one on real traffic without affecting a single result. Compare, watch for drift, then promote when the numbers convince you.
The unglamorous half
- Data export, deletion and portability with a full audit trail.
- Personal data removed before analysis and never stored in the results.
- Separate access levels for administrators, analysts and viewers.
- Containers that run without elevated privileges, and upgrades without downtime.
- Live dashboards, spike detection and automated escalation routing for the operations team.