Introduction
No standard template for AI project acceptance? How should outcomes be evaluated? How should security be verified? This article provides a complete AI project acceptance criteria template to make acceptance evidence-based.
1. Functional Acceptance
1.1 Basic Functions
| Acceptance Item | Acceptance Criteria | Test Method |
|---|---|---|
| All functional points implemented | 100% of contracted functions implemented | Verify item by item against the functional test checklist |
| Permission control effective | Different roles see different content | Multi-role testing |
| Data flow normal | Data is correctly synchronized between systems | End-to-end process testing |
| Exception handling normal | Exceptions have prompts and fallback mechanisms | Exception scenario testing |
1.2 AI-Specific Functions
| Acceptance Item | Acceptance Criteria | Test Method |
|---|---|---|
| Intent recognition | Core intent recognition accuracy ≥90% | Verify with 200+ test cases |
| Knowledge retrieval | Recall rate (Recall@10) ≥85% | Evaluate using a standard test set |
| Answer generation | Answer accuracy ≥85% | Manually label 100+ real questions |
| Human handoff | Smooth handoff process with complete context | Simulate low-confidence scenarios |
2. Performance Acceptance
| Metric | Standard Value | Test Conditions |
|---|---|---|
| Average response time | ≤2 seconds | Normal load |
| P99 response time | ≤5 seconds | Normal load |
| Peak throughput | ≥ contracted value | Stress testing |
| System availability | ≥99.9% | Run for 7 days |
| GPU memory usage | ≤ contracted value | Continuous operation |
| Concurrency support | ≥ contracted number of concurrent users | Concurrency testing |
3. Security Acceptance
3.1 Data Security
| Acceptance Item | Standard | Test Method |
|---|---|---|
| Data transmission encryption | TLS 1.2+ | Packet capture verification |
| Data storage encryption | AES-256 | Configuration check |
| Sensitive data masking | ID card number/mobile number/bank card number | 100+ test cases |
| Access control | RBAC + document-level permissions | Unauthorized access testing |
3.2 AI Security
| Acceptance Item | Standard | Test Method |
|---|---|---|
| Prompt injection protection | Malicious instructions are not executed | 50+ injection attack tests |
| Hallucination control | Hallucination rate in core scenarios ≤5% | Manual labeling verification |
| Output filtering | Non-compliant content is not output | Sensitive word + non-compliant content testing |
| Operation audit | All key operations are recorded | Log integrity check |
3.3 Security Testing
4. Outcome Acceptance
4.1 Outcome Metrics
| Scenario | Accuracy Target | Hallucination Rate Target |
|---|---|---|
| Core scenarios | ≥95% | ≤3% |
| General scenarios | ≥85% | ≤10% |
| Edge scenarios | Allow "I don't know" | — |
4.2 Outcome Testing Methods
| Method | Sample Size | Executor |
|---|---|---|
| Automated evaluation | 500+ items | Technical team |
| Manual labeling evaluation | 100+ items | Business team |
| Real user testing | 50+ people | Target users |
| A/B comparison | Compare with the legacy system | Operations team |
4.3 Outcome Degradation Testing
Run continuously for 7 days, with accuracy fluctuation not exceeding ±3%.
5. Documentation Acceptance
| Document Type | Required Content |
|---|---|
| User manual | User operation steps, screenshots, FAQs |
| O&M manual | System architecture, deployment steps, monitoring metrics, emergency response plan |
| API documentation | API descriptions, request/response examples, error codes |
| Training materials | Training PPT, video tutorials, assessment questions |
| Knowledge base management | Document update process, templates, quality standards |
6. Acceptance Process
```
Pre-acceptance (internal) → Issue remediation → Formal acceptance (customer participation)
↓
Functional acceptance → Performance acceptance → Security acceptance → Outcome acceptance → Documentation acceptance
↓
Acceptance report → Open issue list → Rectification within deadline → Official launch
```
6.1 Acceptance Pass Criteria
Conclusion
AI project acceptance should not focus only on whether the "outcomes are good." Functionality, performance, security, and documentation are all indispensable. Establish systematic acceptance criteria so delivery is evidence-based and both parties share a common understanding of "completion."
Want to establish AI project acceptance criteria? Book a free acceptance consultation