Human at the Helm
Human at the Helm
Paula Goldman's (Salesforce) reframing of Human in the Loop for the agent era, from How AI Is Changing Talent Not Just Tasks (HBR IdeaCast).
The distinction
- Human in the loop (as commonly practised): "AI is going to draft and people are going to approve." An approval gate on every output. Goldman notes the phrase comes from the Cold War military — who makes the consequential decision when a system detects an incoming missile? — and argues it "got kind of misunderstood" in this wave.
- Human at the helm: agents "reason through lots of complex tasks and do lots of things all at once at an incredible speed", so per-output approval doesn't scale. Instead, design the system so the right things reach human judgment at the right time. The human steers and owns consequential calls; they don't rubber-stamp everything.
She uses it two ways: as a design principle for AI systems and as a metaphor for how to manage in the AI era.
What "the right things" are (her heuristics)
- "Routine to AI, complex to humans" gets you ~70% of the way.
- The rest: customer emotions and preferences (angry → a person; embarrassed → AI), employee emotions and preferences, sensitive moments worth preserving even when AI is capable (a person books the first appointment after a cancer diagnosis), relationship-dependent work (complex B2B deals where trust unlocks context), and early-stage creative work (Front-Loading the Brief).
- Deliberate friction where the decision matters, so people "take a beat" rather than cognitively offloading it — accountability for your work product stays yours regardless of AI.
Where it sits in the vault
- It is the management-language version of Bounded vs Unbounded Tasks and Standardized vs Open Tasks: all three reject a blanket gate in favour of routing by task character. Goldman adds the emotional/relational axis those frameworks lack.
- It fits the "phase two" shift in Manager as Translator and Tasks to Responsibilities Shift: the manager's job becomes deciding and writing down which decisions stay human.
warning Tension with Human in the Loop practice elsewhere in the vault Several sources prescribe exactly the "draft-and-approve" pattern Goldman calls inadequate — e.g. the SDD course's read-every-diff loop (Full Course Spec-Driven Development with Coding Agents (DeepLearningAI)) and the "human verifies the email before the agent sends it" accountability case in Five AI Risks That Can Get You Fired (IBM Technology). Partial reconciliation: per-output approval survives where volume is low and stakes are high (code merges, regulated actions); at agent-fleet scale it collapses into routing + exception review. Goldman gives no mechanism for how to detect "the right time" — that's the unsolved engineering part (cf. Judgment Models (System 1 AI) as one candidate router).