Strategy and roadmap
A roadmap listing every AI capability the team finds interesting is not a strategy. It is a wish list. Bizzi runs AI strategy across three nested horizons and a single build-vs-buy decision rule. Every quarter we defend why we are building one thing and not another.
Context
Section titled “Context”The pressure on B2B finance vendors to ship AI is constant. The risk is the team chases shiny capabilities and never finishes the boring ones moving customer KPIs. Three horizons keep the team honest about the difference between now, next, and later. The build-vs-buy rule prevents us from locking into a single LLM vendor.
Three horizons
Section titled “Three horizons”- Now (1 year). Run existing AI features at stable quality. OCR, AI assistant, 3-way matching. Target: STP rate above 85% across all Enterprise tenants.
- Next (3 years). Extend the multi-agent architecture to new use cases. Cashflow forecasting, vendor risk scoring, contract-negotiation assist. Target: AI handles more than 50% of processing volume autonomously for Enterprise customers.
- Later (5 years). Bizzi becomes the default AI partner for Vietnamese CFOs. BAGF becomes the industry reference for AI governance in ASEAN fintech.
The AI Governance Board reviews each horizon quarterly and rebaselines when market or regulatory conditions shift.
Build-vs-buy criteria
Section titled “Build-vs-buy criteria”Every AI feature is decided against four criteria:
- Strategic depth. If the feature creates a proprietary advantage, build. If it is commodity, buy.
- Data sensitivity. If training requires sensitive Bizzi customer data, build with private hosting. If the data is public, buy or use OSS.
- Time-to-market. If a customer commitment depends on speed, buy and fine-tune. If we absorb the timeline, build.
- Lifecycle cost. Cost per Transaction is the deciding KPI. Build has a higher up-front cost dropping with scale. Buy is the opposite.
Most large language models today are bought through our AI Gateway with fallback routing. Specialized OCR models and rerank models are hosted internally.
Vendor posture: no lock-in
Section titled “Vendor posture: no lock-in”Bizzi runs an AI Gateway with automatic fallback to a backup model when the primary fails. The posture has three concrete rules:
- No single LLM vendor exceeds 70% of total inference volume.
- Every prompt is written portably. No vendor-specific syntax.
- Every production model has at least one backup which passed our three-criterion LLM-as-a-Judge evaluation.
The payoff. When a vendor changes pricing, closes an API, or goes down, we shift workload in hours, not weeks.