Legal & Trust
Responsible AI
How Klugliv uses OCR and automation — with humans in control.
Responsible AI
Last updated: 2026-07-20
This page explains how Klugliv uses automated systems today, how we keep humans in control, and what we plan for the future. It is an engineering transparency document — not legal advice.
Summary
Klugliv’s early “AI” is receipt OCR and structured parsing — not a general-purpose chat assistant. We treat machine output as suggestions until you review and save it.
We do not claim that Klugliv is “fully autonomous,” “always accurate,” or certified under any AI governance framework we have not formally achieved.
What Klugliv uses today
| Capability | What it does | Provider / location |
|---|---|---|
| Receipt OCR | Extracts text from receipt images | Google Cloud Vision (via AWS Lambda in our migration slice) |
| Receipt parsing | Converts OCR text into fields (merchant, date, line items, totals) | Klugliv-owned parsers in the app and backend |
| Shopping insights | Aggregates and summarizes your saved purchase history | Derived from data in your account — not generic benchmarks |
There is no large language model (LLM) chat in the current product. There is no always-on voice assistant shipped today.
Human review before trust
Before a receipt affects your saved household history:
- You upload or capture a receipt.
- OCR and parsing produce a draft for review.
- You can correct merchant, dates, line items, and totals.
- Only after you confirm does the data become part of your history.
This “review gate” is intentional. OCR can misread faded paper, unusual layouts, or handwritten notes.
What we do not do with your data (today)
- We do not sell personal receipt data.
- We do not use your receipts to train a proprietary foundation model in the current architecture.
- We do not permanently store voice recordings — voice features are not implemented yet.
- We do not present OCR output as financial, legal, or medical advice.
Third-party OCR (Google Cloud Vision) processes receipt images according to their terms and your project configuration. See the Vendor Register.
Analytics and product telemetry
When PostHog is configured for a mobile build, we may collect usage events (for example, which screens are used) — only after you opt in. This helps us improve reliability and is separate from OCR model training.
Manage your choice in Settings → Trust & privacy, or when prompted on first launch. See GDPR Compliance Review for jurisdiction-specific notes.
Accuracy, bias, and limitations
Receipt OCR works best on clear photos of standard grocery receipts. Known limitations include:
- damaged, crumpled, or low-light images
- non-standard receipt formats or languages not tuned in our parsers
- ambiguous product names on discount lines
Insights and suggestions are based on your household’s history, not population-wide profiling. They can still be wrong if underlying receipt data was wrong — another reason review matters.
Planned capabilities (not shipped)
These are documented in product and engineering docs as future work:
| Planned feature | Status | Design intent |
|---|---|---|
| Hands-Free Shopping (voice during active trips) | Planned | Optional, trip-scoped; no permanent audio storage; vocabulary learning — not voiceprints |
| Predictive list refinement | Planned | Suggestions you can override |
| Richer async enrichment | Planned | Background jobs after OCR — still tied to your account data |
For voice-specific privacy design, see our internal docs/hands-free-shopping.md (engineering). Public summary: audio is processed for the utterance, then discarded; only structured trip updates remain.
Governance and updates
We update this page when:
- we add a new automated capability that materially affects users
- we change how third-party AI/OCR services are used
- we introduce voice or generative features
Related documents:
Contact
Questions about how Klugliv uses automated processing: hello@klugliv.com