GetirFinansAssistant streaming shipped · related work under evaluation
Designing answers and next actions in a fintech app
Suggested prompts account for 31% of assistant messages. The share of closes before a question fell from 46.6% in June to 34.0% in September.
- Role
- Product Design, Prototyping
- Timeline
- 2026 — Present
- Company
- GetirFinans
- Team
- Product and iOS engineering
- Platforms
- iOS
This work focused on the assistant, its composer, and what happens when a question arrives in search rather than a chat.
Opening context
Continue from the task at hand
The assistant opens from the deposit calculator, so it continues a task rather than becoming a separate destination. The composer holds typing and dictation together in one place.
The assistant opens from the deposit calculator, so it continues a task rather than becoming a separate destination. The composer holds typing and dictation together in one place.
Decision one
When should an answer offer a next action?
A whole-string search match returned no result for some questions; matching individual words could suggest the wrong destination. “How do I pay my credit-card debt?” matched “kredi” into “kart” and suggested linking or unlinking an account to a card instead of paying card debt. The prototype moved to curated destination matching, with no route when an explanatory answer has no useful next step.
A question that search cannot match to a product can still receive an answer in place. The route into the assistant remains available when the answer does not finish the task.
Typing a topic surfaces related questions inline among search results, rather than moving people into a separate mode.
Interaction study
Thinking and answering
The assistant has distinct thinking and answering states. This recording captures the transition into a streamed answer and the campaign cards that follow it.
The recording shows the assistant thinking, then presenting a streamed answer with campaign cards. The treatment is still being evaluated as the wider conversation experience develops.
Decision three
What happens when composing gets difficult?
Dictation and typing share a composer, so a person does not have to learn a second place to compose. At the input boundary, the field preserves the draft and gives a local response.
Mic use doubled in the two months after launch, from 1.4% to 2.9% of chat users, and about 4% of chat opens tap it.
The field becomes the listening state, preserving where a future transcription will appear. This recording shows listening, not a completed dictation result.
The field grows to five lines, then stops at 200 grapheme clusters. The boundary keeps the existing text in place and gives a local response, rather than discarding what a person has composed.
Delivery and next evidence
What shipped, and what to measure next
Assistant streaming is shipped. Prompt suggestions carry 31% of every message sent, steady over three months. People who close the assistant before asking anything fell from 46.6% of closes in June to 34.0% in September.
The same data points at the next job. As the median wait grew from 3.7 to 6.4 seconds, closes during the wait rose from 0.5% to 3.6%. That is the case for a composer that stays live while the assistant works, with a Stop in the send slot. The deposit calculator entry is the other gap: 38% of those opens send a message, against 63% overall.
Dictation completion and inline search routing are not instrumented yet. They are the next evidence to collect.