DIREKT World-Class and AI-Native Product Direction Decision

Date: 2026-07-21
Status: Accepted for planning and VC1–VC8 implementation sequencing; production AI providers/use cases remain separately gated
Owner direction: Modernize the completed functional product toward world-class marketplace quality and make AI a critical cross-feature capability without rebuilding the backend/integration foundation.

Context

DIREKT has a comparatively mature backend, trust, privacy, integration and cross-client functional foundation, but its visible customer/provider/operations experiences remain materially less polished than leading global local-service marketplaces.

The product also needs an explicit AI architecture so AI is designed into discovery, provider enablement, support and operations rather than added later as an isolated chatbot.

Decision

Adopt a composite benchmark and an eight-stage modernization programme.

Benchmark stack

  • Urban Company for marketplace/service-professional maturity;
  • Checkatrade for recurring trust checks and customer-facing trust proposition;
  • Taskrabbit for short task-to-provider/action flow;
  • Thumbtack for AI-guided natural-language/multimodal service discovery;
  • DIREKT's own differentiation in check-specific proof, privacy-by-precision, payment-independent trust, Zambia resilience and human-accountable operations.

Programme

  • VC1: design-system reconciliation;
  • VC2: high-fidelity benchmark directions and explicit owner approval;
  • VC3: approved Design DNA/component foundation;
  • VC4: customer world-class experience;
  • VC5: provider professional workspace;
  • VC6: operations mission control;
  • VC7: AI intelligence layer;
  • VC8: world-class visual/product/AI quality gate.

AI authority model

AI may interpret, classify, suggest, summarize, extract, draft and rank within approved bounds.

AI may not autonomously:

  • authorize access;
  • approve/reject/revoke/suspend verification;
  • publish trust claims;
  • decide serious complaints/appeals;
  • alter financial ledger/reconciliation truth;
  • make legal conclusions;
  • expose restricted evidence;
  • execute unrestricted tools or database/storage/admin actions.

Alternatives considered

Copy one global marketplace

Rejected. No single product matches DIREKT's trust, privacy, informal/mobile-provider and Zambia-specific operating model.

Visual redesign only

Rejected. It would improve appearance without resolving AI-native discovery/operations opportunities or defining safe model authority.

AI chatbot as a standalone feature

Rejected. A generic chatbot does not address the highest-value AI opportunities in search, onboarding, evidence assistance, operations and support, and can create excessive-agency/privacy risk.

AI-first autonomous trust operations

Rejected. Probabilistic models cannot become the final authority for high-impact trust, complaint, payment or legal decisions.

Consequences

  • Design work must use the approved benchmark/Design DNA rather than ad-hoc aesthetic changes.
  • AI requires a backend orchestration boundary and source-controlled use-case governance.
  • AI model/provider activation becomes an integration-status concern and cannot be inferred from planning or credentials.
  • Manual/deterministic fallbacks remain mandatory for core journeys.
  • Restricted evidence is not sent to external AI providers by default.
  • VC8 adds AI evaluation/security gates to visual/accessibility/regression gates.
  • Phase 11 real pilot and Phase 12 production gates remain unchanged and cannot be cleared by VC work alone.

References

  • docs/product/WORLD_CLASS_PRODUCT_AND_AI_PLAN.md
  • docs/design/VISUAL_COMPLETION_PLAN.md
  • docs/architecture/AI_PRODUCT_ARCHITECTURE.md
  • docs/product/PRODUCT_REQUIREMENTS.md
  • DEFINITION_OF_DONE.md
  • Issue #259