The litmus test: do you actually need a custom AI solution?

Use three filters. If you cannot answer these with confidence, push pause on the build.

1) Business problem framework

State the primary objective in one line and pick one bucket:

• Cost down: reduce manual effort, shrink cycle time, cut unit cost
  Examples: ticket deflection, claims triage, intelligent document processing

• Revenue up: improve conversion, lift retention, raise average order value
  Examples: next best action, dynamic pricing, churn risk scoring

• Risk down: reduce chargebacks, spot compliance breaches, catch bad data early
  Examples: anomaly detection, fraud signals, KYC automation

Add two numbers: the current baseline and the minimum change that makes the project worthwhile. If you cannot quantify both, the scope is not ready.

2) Data readiness audit

Data funds the model. Audit it before any code.

• Inventory: where the data lives, who owns it, schemas, volumes, refresh rates

• Quality: missing fields, label coverage, bias pockets, duplicates, drift history

• Access: legal basis, consent, retention, residency, PII handling

• Pipelines: how raw data turns into features, lineage, and reproducibility

• Governance: who approves usage, who audits, who can shut it down

Score each item red, amber, or green. Reds block the project. Ambers require a plan that fits inside the budget and timeline. Only greens move straight ahead.

3) Talent and resource equation

Be honest about build versus buy. Consider:

• Speed to first outcome: in-house hiring adds months

• Scarce skills: applied ML, MLOps, secure data engineering, evaluation design

• Run costs: monitoring, retraining cadence, incident response

• Change management: training, SOP updates, stakeholder communication

If speed matters and the skill mix is rare, partner. If the problem is core IP and you already have the team, build.