# Local vs international AI contractor — weighted decision matrix

Score every option from 1 (weak fit) to 5 (strong fit). Multiply the score by the weight and divide by 5. Change weights before scoring, not after seeing the result.

| Criterion | Default weight | Evidence to request |
|---|---:|---|
| Workflow and market context | 15 | Process map, user interviews, language and regulatory assumptions |
| Specialist depth | 15 | Named delivery roles, relevant architecture examples, escalation path |
| Time to first reviewable result | 10 | Discovery plan, milestone definitions, dependency list |
| Total cost of ownership | 20 | Build, integration, evaluation, operations, change and exit costs |
| Delivery and continuity risk | 15 | Bus-factor plan, backups, documentation, source and credential ownership |
| Control and communication | 10 | Decision log, demo cadence, issue ownership, approval boundaries |
| Support and handover | 15 | Monitoring, incident response, runbooks, training and transition plan |
| **Total** | **100** | Weighted result is a discussion aid, not an automatic verdict |

## Scenario weights

| Scenario | Context | Depth | Speed | TCO | Risk | Control | Support |
|---|---:|---:|---:|---:|---:|---:|---:|
| Armenia-first operational automation | 25 | 10 | 10 | 15 | 15 | 10 | 15 |
| Narrow frontier-model research spike | 5 | 30 | 15 | 10 | 20 | 10 | 10 |
| Multi-country production platform | 10 | 20 | 10 | 20 | 15 | 10 | 15 |

## Decision guardrails

- Reject any option that cannot name the owner of production incidents, data permissions and source-code handover.
- Treat a score gap below 10 points as inconclusive; run a paid discovery or controlled prototype with the same acceptance criteria.
- Do not use day rates as the final comparison. Compare the cost of reaching and operating an accepted outcome.
- Record assumptions and confidence beside every score so procurement can revisit the decision when facts change.