Chapter 0.3 — AI as a sociotechnical system¶
🎯 Chapter objective¶
Establish that AI systems are not merely technical: they involve people, processes, politics and culture.
🧠 Core concept¶
A sociotechnical system operates under the interaction between:
- Technical: data, model, code, infrastructure.
- Social: users, operators, stakeholders, regulators.
- Organizational: processes, roles, internal politics, culture.
Operational failures, lack of documentation, absence of governance, or conflict between teams are often as damaging as wrong hyperparameters.
🏗️ How this shows up in production¶
- A good model is pulled because the compliance team was not consulted.
- A useful agent is shut down because nobody takes ownership of operating it.
- A correct pipeline produces wrong data because the upstream team changed the schema without notice.
⚖️ Trade-offs¶
| Option | Advantage | Risk |
|---|---|---|
| Centralize governance | Consistency | Slowness |
| Decentralize to squads | Speed | Inconsistency |
| Hybrid (federated) model | Balance | Coordination cost |
🚨 Failure modes¶
- Technical team without organizational mandate to change data, rules or process.
- No clear RACI: nobody owns it.
- Technical success without user acceptance (overreliance or rejection).
🛡️ Controls and mitigations¶
- Explicit RACI for each agent/model: technical owner, business owner, operations owner.
- Governance committees with multidisciplinary participation.
- User training on the system's capabilities and limits.
🧰 Related practical example (planned)¶
EX-FUND-03— a RACI template for an enterprise AI system.
📌 Checklist¶
- [ ] Are a technical owner and a business owner named?
- [ ] Is there a communication plan with users about what the system is and is not?
- [ ] Is there a multidisciplinary review process before go-live?
📚 References¶
- NIST AI RMF 1.0 — Sociotechnical perspective: https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf