AI work gets messy fast. Simon Willison highlights a simple fix: assign a Directly Responsible Individual (DRI) for every AI-critical piece of work. One name. One owner. Better outcomes (source).
What is a DRI?
A DRI is a single person accountable for a specific outcome. The idea is popularized by Apple’s product culture and embraced across tech because it prevents fuzzy ownership and slow decisions.
Why AI teams need DRIs now
- Speed: One owner can cut standstill debates and ship experiments faster.
- Safety: Clear accountability for red-teaming, guardrails, and rollback plans.
- Cost: Someone actively manages token burn, caching, and model/provider choices.
- Quality: Owners maintain evals and regression tests as prompts and data shift.
Where to assign DRIs in AI work
- Model and provider selection (e.g., GPT-4o vs local model)
- Prompts and system instructions for each feature
- Evaluation suite and acceptance criteria
- Data pipelines and retrieval/grounding docs
- Safety guardrails: PII handling, jailbreak resistance, policy filters
- Monitoring: latency, cost, quality drift, incident response
- Docs and runbooks for maintenance and handoffs
Implement DRIs this week (lightweight playbook)
- Create a visible roster: feature → DRI → goal metric (e.g., deflection rate, CSAT, cost/call).
- Attach ownership to artifacts: prompt files, evals, dashboards, and on-call schedules.
- Define success and safety gates: pass/fail thresholds, rollback triggers, PII rules.
- Set a decision cadence: weekly 20-minute reviews per DRI with a short decision log.
- Record changes: every prompt or model tweak gets a changelog entry linked to eval results.
- Backups: name a secondary DRI for vacations and incidents.
DRI vs RACI: when to use which
RACI clarifies roles; DRI accelerates decisions. Use RACI for multi-team programs. Use a DRI to ensure one person can decide and be accountable in day-to-day AI work (RACI background).
Example: Shipping an AI support bot
- Retrieval pipeline DRI: owns embeddings, chunking, and freshness SLAs.
- Prompt DRI: owns system prompt, tools, and eval-based acceptance criteria.
- Safety DRI: owns abuse/jailbreak tests and escalation runbook.
- Cost/latency DRI: owns caching, batching, model switches, and SLOs.
- Monitoring DRI: owns dashboards, alerts, and rollback button.
Common pitfalls to avoid
- “Group DRI” is not a thing—ownership must resolve to one person.
- Hidden ownership: if the roster isn’t visible, it doesn’t exist.
- No evals: if quality isn’t measured, DRIs can’t protect against drift.
- Permanent DRIs: rotate periodically to avoid burnout and share knowledge.
Takeaway
Attach a named DRI to every model, prompt, and eval. You’ll ship faster, lower risk, and keep quality from decaying as your AI surface grows.
Enjoyed this nugget? Get our best AI tactics in your inbox—subscribe to The AI Nuggets newsletter: theainuggets.com/newsletter.
