In a recent conversation highlighted by Storyboard18, Perplexity CEO Aravind Srinivas drew a simple line: today’s AI can solve well‑posed problems at scale, but it doesn’t choose the problems. That choice—what actually matters—remains a human act rooted in curiosity and values. For organisations planning AI programmes, this distinction is more than philosophy; it’s an operating model.

The claim

• AI is a powerful optimiser: given a defined objective and constraints, it drafts, summarises, ranks, extracts, and reasons—often faster than humans.

• Humans set the objective: deciding what to ask, why it matters, and what trade‑offs are acceptable requires context, ethics, and stakeholder awareness.

• Curiosity is a moat: the uniquely human habit of asking better, bolder questions drives new products and policies—and remains outside present‑day AI capabilities.

Why this matters to leaders

Most stalled AI programmes fail not on model accuracy but on problem framing. Teams ship features that optimise the wrong metric or ignore regulatory and customer realities. Treat question design as a first‑class capability, not a nice‑to‑have.

Strategy implications (from idea to impact)

1. Build a Question Pipeline

   • Run quarterly “question audits” per function. Capture unanswered customer questions, compliance pain points, and decision bottlenecks.

   • Prioritise by impact × feasibility × risk. Convert top questions into machine‑answerable tasks.

2. Design a Human‑in‑the‑Loop Funnel

   • Draft → Review → Verify → Publish → Learn. Add gates where domain experts can veto or edit outputs.

   • Log prompts, sources, and reviewer decisions for auditability.

3. Instrument the Right Metrics

   • Discovery metrics: questions captured, questions validated.

   • Quality metrics: citation coverage, factual error rate, reviewer override rate.

   • Outcome metrics: time saved to decision, customer resolution rate, compliance exceptions avoided.

4. Treat sources as features

   • Prefer answer modes with citations. When a response matters (legal, safety, finance), require source‑grounding or retrieval over your controlled knowledge.

5. Own your prompts and patterns

   • Build a shared prompt library with versioning and examples per workflow (support, finance, legal, engineering).

   • Pair each prompt with a review rubric and known failure modes.

The on‑device AI angle

Srinivas also points to a shift: as models become more efficient, on‑device or edge inference may handle a growing share of tasks, reducing dependence on mega data centres. For buyers, that suggests:

• Latency & privacy gains for sensitive or mobile workflows.

• Hybrid architectures where low‑risk tasks run locally and high‑stakes tasks call secure cloud endpoints with logging and policy controls.

• Procurement updates to evaluate device‑side capabilities (NPUs, memory) and manage model deployment across fleets.

Governance: keep humans in charge

• DPIA/Privacy: document lawful basis, retention, and redaction for any PII.

• Model risk: register models, define intended use, test for drift and adversarial prompts.

• Explainability: store sources and prompt/output pairs; flag AI‑assisted content in records.

• Accountability: define when a human must sign off (e.g., credit, safety, legal notices).

• Ethics & bias: test outputs for disparate impact; publish mitigations.

Practical playbook: 60‑day rollout for a business unit

Weeks 1–2 – Question discovery: run workshops; harvest top 20 unanswered questions; rank by impact/feasibility/risk.
Weeks 3–4 – Thin slice builds: ship 3–5 prompts with retrieval over approved sources; require citations; set review rubrics.
Weeks 5–6 – Pilot: 50–100 users; track time‑to‑answer, override rate, and user satisfaction; add translation and summarisation tasks.
Weeks 7–8 – Governance hardening: add logging, RBAC, DPIA updates; publish known‑issue list and playbook.

Bottom line

AI will keep getting better at solving. Advantage will flow to organisations that ask better questions, design human‑in‑the‑loop systems, and adopt hybrid/on‑device patterns where they reduce risk and cost.

Next Steps: Want a question‑to‑impact sprint for your team? Generation Digital can facilitate discovery workshops, build thin‑slice assistants, and set up governance in weeks.

FAQ

Q1. What does “AI solves; humans decide” mean for my roadmap?
A. Treat question discovery and review as core capabilities. Build a pipeline from problem framing to governed deployment with clear metrics.

Q2. Do we need on‑device AI now?
A. Not everywhere. Start with latency‑sensitive or privacy‑sensitive tasks; pilot hybrid patterns and measure impact.

Q3. How do we avoid hallucinations?
A. Use retrieval over trusted sources, require citations for critical tasks, and enforce human review before publishing.

Q4. What metrics prove value?
A. Time‑to‑answer, reviewer override rate, citation coverage, resolution rates, and cost per task.

Q5. How do we keep auditors happy?
A. Log prompts/outputs, pin model versions, store sources, and maintain sign‑off points for high‑impact decisions.