Artificial Intelligence Usage Policy

1. Purpose

This policy sets out how HomeSewa designs, deploys and governs artificial intelligence (AI) across its hyperlocal home services platform. It applies to all AI-driven and AI-assisted features used by customers, service providers (workers/vendors) and internal staff, and is intended to guide product, engineering and operations decisions as the platform scales across Nepal.

2. Scope

This policy covers:

• Voice-to-message conversion in Nepali within the booking form
• AI-based service request matching with service providers
• Personalised service suggestions based on booking history
• Custom offers generated from those suggestions
• Seasonal service recommendations
• Conversational chatbot for booking
• Smart worker matching and dispatch
• Basic DIY helpdesk AI automation
• Multilingual communication and content-to-voice conversion

3. Guiding Principles

• Human oversight – AI recommends and automates; a human (customer, worker or ops team) retains the ability to review, override or escalate any AI decision affecting a booking, price or dispatch.
• Transparency– Customers and providers are told when they're interacting with an AI system (chatbot, voice assistant, auto-matching) rather than a human.
• Data minimisation– Only data required for the specific AI function is collected and processed; voice and behavioural data aren't reused outside their stated purpose without consent.
• Fairness – Matching, dispatch and offer models are periodically checked for bias across geography, gender, provider tenure and customer segment.
• Security & privacy – Voice recordings, location data and personal information are encrypted in transit and at rest, with access restricted on a need-to-know basis.
• Compliance– AI features handling personal data of Nepali users are designed to align with Nepal's Individual Privacy Act, 2075 (2018), and related regulations (see Section 8).

4. Acceptable & Prohibited Use

AI may be used to:

• Transcribe, translate, match, recommend, price (within approved ranges), dispatch and answer routine queries, as described in Section 6.
• Support human decisions with recommendations and rankings that a human can review or override.

AI must not be used to:

• Set a final price or discount outside finance-approved ranges without human sign-off.
• Make or influence hiring, deactivation, or disciplinary decisions about a worker without human review.
• Deny, deprioritise or delay service to a customer or worker based on caste, ethnicity, religion, gender, disability, or any other protected characteristic.
• Impersonate a human agent without disclosing that the customer/worker is speaking with an AI system.
• Take an irreversible action (final payment capture, permanent suspension, account deletion) without a human-reviewable checkpoint (see Section 7).
• Be repurposed for surveillance, profiling, or data use beyond the stated purpose of the feature it was built for.
• Give medical, legal, or safety-critical advice beyond platform-approved, factual helpdesk content.

5. Ethical Guidelines

• Non-discrimination – Matching, dispatch, offer and suggestion models are built and tested to avoid disadvantaging customers or workers on the basis of caste, ethnicity, religion, gender, or account tenure alone.
• Accountability – Each AI feature has a named internal owner accountable for its outcomes; HomeSewa, not the AI system, is responsible for decisions made using it.
• Explainability – Customers and workers can request a plain-language reason for a significant AI-driven outcome (e.g. why a particular offer, match, or dispatch was given), routed through support.
• No manipulative design – Offers, suggestions and chatbot prompts must not use false urgency, hidden terms, or exploit customer vulnerability to drive bookings.
• Worker wellbeing – Dispatch and matching logic must not be used to systematically overload, underpay, or penalise workers; workload distribution is monitored as part of fairness review (Section 3).

6. Feature-Specific AI Policies

6.1 Voice-to-Message (Nepali) in Booking Form

• Purpose: Let users record a voice note transcribed into text in the booking form, lowering the barrier for non-typing users.
• AI role: Nepali speech-to-text converts voice input into editable text; does not auto-submit without confirmation.
• Human oversight: Transcribed text is always shown before submission for edit or re-record.
• Data handling: Voice clips processed for transcription only, not retained longer than necessary; consent is captured before recording begins.
• Risk & mitigation: Mis-transcription risk for dialects/accents is mitigated by manual edit and re-record options.

6.2 Service Request Matching with Service Providers

• Purpose: Match incoming requests to the most suitable available provider based on category, location and profile.
• AI role: Ranking model scores eligible providers using proximity, skill, ratings and availability.
• Human oversight: Customers see top matches rather than a single locked assignment; ops can manually reassign.
• Fairness: A visibility mechanism ensures newer/lower-volume providers still get a share of requests.
• Risk & mitigation: Over-concentration to top providers is monitored and corrected via rotation.

6.3 Service Suggestions Based on Earlier Bookings

• Purpose: Recommend relevant follow-on or recurring services using booking history.
• AI role: Recommendation model uses past bookings, frequency norms and category affinity.
• Human oversight: Suggestions are optional prompts, never pre-selected or auto-booked.
• Data handling:Only the customer's own history is used; not shared across accounts.
• Risk & mitigation: Suggestion frequency is capped to avoid feeling intrusive.

6.4 Custom Offers Based on Suggestions

• Purpose: Convert a relevant suggestion into a time-bound personalised discount or bundle.
• AI role: Pricing engine determines eligibility using recency, frequency and margins.
• Human oversight:Finance/ops sets the allowable discount range; AI can't exceed pre-approved limits without sign-off.
• Fairness: Reviewed so similar customers get comparable offers, avoiding arbitrary price discrimination.
• Risk & mitigation: Offer-abuse (e.g. account farming) checked with fraud rules alongside the model.

6.5 Seasonal Services

• Purpose: Surface services relevant to the current season or festival period in Nepal (e.g. pre-monsoon roof/waterproofing checks, deep cleaning before Dashain and Tihar).
• AI role: Hybrid of seasonal calendar and demand-forecasting signals.
• Human oversight: Calendar and promoted categories reviewed/approved by ops/marketing each season.
• Risk & mitigation: Forecasts validated against actual booking data each season.

6.6 Chatbot for Booking

• Purpose: Let customers describe a need in natural language (text/voice) and complete a booking conversationally.
• AI role: LLM-based chatbot interprets intent, asks clarifying questions, pre-fills the booking form.
• Human oversight: Bot identifies itself as AI; live-agent handoff exists for complex or failed interactions.
• Data handling: Logs used for improvement are anonymised where feasible.
• Risk & mitigation:Restricted to booking-related tasks; no medical/legal/safety advice; can't confirm a booking without explicit customer confirmation.

6.7 Smart Worker Matching & Dispatch

• Purpose: Assign and route the nearest, most qualified available worker to a confirmed job.
• AI role: Dispatch model combines real-time location, skill match, workload and ETA.
• Human oversight: Workers can accept/decline; ops has manual override for emergencies, complaints or no-shows.
• Fairness: Assignment volume and earnings opportunity monitored across the worker pool.
• Risk & mitigation: Dispatch decisions logged for auditability in case of disputes.

6.8 Basic DIY Helpdesk AI Automation

• Purpose: Resolve common, low-complexity queries (rescheduling, payment status, basic troubleshooting) without a human agent.
• AI role: Retrieval-based assistant answers from an approved knowledge base; can perform simple actions like rescheduling within policy limits.
• Human oversight: Anything outside the knowledge base, or involving refunds/safety complaints/disputes, escalates automatically to a human.
• Risk & mitigation: Restricted from inventing answers outside its knowledge base (no unverified pricing, refund or legal claims).

6.9 Language Communication and Content-to-Voice

• Purpose:Deliver notifications, reminders and support content in the user's preferred language (Nepali, English, and relevant local languages) and as voice where reading is inconvenient.
• AI role: Machine translation and text-to-speech convert platform content into the target language/format.
• Human oversight: Critical content (payment amounts, cancellation terms, legal notices) reviewed for translation accuracy; language preference is user-selected, never assumed.
• Risk & mitigation: Numeric values, dates and prices stay in a standard, non-translated format regardless of language to avoid critical translation errors.

7. Human Oversight & Escalation

No AI feature may take an irreversible action (final payment capture, permanent account suspension, worker deactivation) without a human-reviewable checkpoint. Every automated feature has a defined escalation path to a human for edge cases, errors or disputes.

8. Data Privacy Requirements (Nepal)

• Legal basis– HomeSewa's data handling is designed to align with the Constitution of Nepal's Article 28 (right to privacy), the Individual Privacy Act, 2075 (2018) and Individual Privacy Regulation, 2077 (2020), the Electronic Transactions Act, 2063 (2008), and the Data Act, 2079 (2022).
• Consent– Personal information (name, phone, address, voice recordings, location, biometric or behavioural data) is collected, recorded, disclosed or processed only with the individual's informed consent, except where collection is required by law or court order.
• Sensitive data – Caste, ethnicity, religious affiliation, health information, and similar sensitive personal data are given heightened protection and are not processed unless required for service delivery, health/emergency purposes, or the individual has made it public themselves.
• Rectification – Customers and workers can request correction of inaccurate personal data held by HomeSewa, consistent with Section 28 of the Individual Privacy Act. (Note: Nepali law does not currently provide a right to erasure/deletion or data portability in the way GDPR-style frameworks do; HomeSewa nonetheless deletes data it no longer needs as a matter of internal policy.)
• Retention – Voice clips are retained for a maximum of 30 days after transcription; chat logs used for QA are retained for 90 days; booking/behavioural data used for personalisation is retained only while the account is active plus a defined post-closure window.
• Third-party/cross-border processing – If any third-party AI vendor (speech-to-text, translation, LLM API) processes data outside Nepal, this is disclosed to users and limited to vendors with adequate data-protection safeguards.
• Breach response– Any suspected data breach involving AI-processed personal data is reported internally within 24 hours; affected individuals are notified where required, and complaints can be escalated to the District Court under the Individual Privacy Act's three-month complaint window.
• No resale – AI training or improvement data is never sold or shared with unrelated third parties.

9. Quality Assurance & Testing

• Pre-deployment validation – Every AI feature is tested against a defined accuracy/performance threshold (e.g. minimum transcription accuracy, minimum matching precision, minimum translation accuracy for critical fields) before going live.
• Bias audits – Matching, dispatch and offer models undergo a documented fairness check across geography, gender, caste/ethnicity, provider tenure and customer segment before launch and at each quarterly review.
• Ongoing monitoring – Each feature owner reviews accuracy, fairness and complaint rates quarterly, or sooner if a pattern of errors or complaints is identified.
• Incident logging – Errors, overrides and escalations are logged per feature to identify recurring failure patterns and inform model retraining.
• Versioning & rollback – Model updates are versioned; any update that degrades performance or fairness beyond threshold is rolled back to the last approved version.
• Vendor vetting – Third-party AI tools (speech-to-text, translation, LLM APIs) are evaluated for accuracy, data-handling practices and compliance before integration.

10. Implementation Roadmap

• Phase 1 – Pilot: Launch voice-to-message, chatbot and basic helpdesk automation to a limited user/provider group in a single Nepali city; validate accuracy thresholds and escalation flow.
• Phase 2 – Limited rollout: Extend to service matching, suggestions and seasonal recommendations across a wider metro area; run first fairness audit.
• Phase 3 – Full scale: Enable smart dispatch and custom offers platform-wide once QA thresholds are consistently met.
• Staff training: Ops and support staff are trained on escalation handling, bias-review basics, and how to explain AI-driven outcomes to users.
• Review ownership: Each feature has a named internal owner responsible for its quarterly QA and fairness review (to be assigned as the team grows).

11. Ownership & Responsibility

Applies to all team members, contractors and third-party vendors building or integrating AI features into the platform. Any new AI feature is reviewed against this policy before launch.