THE SHORT ANSWER
In a sound human-AI workflow, people define the objective and constraints; AI may research, draft, analyze or execute bounded steps; people verify evidence, handle exceptions, decide and remain accountable. The allocation changes with consequence and system reliability.
Use explicit handoffs
| Owner | Responsibility |
|---|---|
| Human | Define outcome, context, boundaries and quality bar |
| AI | Generate alternatives, retrieve material, draft or analyze |
| Human | Check sources, assumptions, omissions and consequences |
| AI | Revise or execute an approved bounded step |
| Human | Approve, communicate, act and own the outcome |
Assistance, recommendation and execution need different controls
A brainstorming assistant can tolerate weak suggestions because a person filters them. A recommendation affecting a customer needs traceable inputs and challenge. An agent with system access also needs permission limits, monitoring and exception handling.
Capability is jagged and changes by task
A field experiment with 758 consultants found that AI effects differed across selected tasks depending on whether they were inside or outside the model's capability frontier. The study warns against treating strong performance on one task as evidence for adjacent tasks.
Microsoft's 2026 framework identifies several collaboration modes, but its evidence includes self-reported survey responses and product telemetry. Use it as a design prompt, not a universal maturity model.
Evidence & context: Harvard Business School · Microsoft Work Trend Index
Write a collaboration contract
- What may AI do?
- What evidence must accompany its output?
- What must a person check?
- Which action requires approval?
- Who handles exceptions?
- Who owns the final outcome?
Sources & further reading
- AI Risk Management and Human-AI Interaction
NIST AI Resource Center. Official guidance on different human and AI decision roles and oversight. Appropriate reliance depends on context, consequence and system evidence.
- Navigating the Jagged Technological Frontier
Harvard Business School. A field experiment with 758 BCG consultants performing selected knowledge-work tasks. AI effects differed depending on whether a task was inside or outside the model's capability frontier; the sample and tasks limit broader inference.
- Agents, Human Agency, and the Opportunity for Every Organization
Microsoft Work Trend Index. A vendor study drawing on a 2026 self-reported survey of 20,000 workers across ten markets, including India, plus Microsoft product telemetry. Its collaboration framework is useful, but self-reporting and commercial context limit causal claims.
Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.
A correction, a counterexample or an experience worth sharing?
Join the conversation ↗