Your AI Needs a Job Description

Effective human–AI teamwork requires us to define what the AI should do, what it may decide and where human responsibility begins.

We spend considerable time defining missions, responsibilities, governance structures and RACIs. We would not bring a new colleague into a team and simply tell them to “help wherever you can.” We explain the objective, their role, the limits of their authority, how their work connects with others and when a decision must be escalated.

Yet AI is often introduced in exactly the opposite way. Employees receive access to a system and a broad instruction to use it responsibly. The technology changes, but the need for clarity does not.

Designing human–AI teamwork

It is tempting to assume that combining human judgement with artificial intelligence must produce better results. The research suggests otherwise.

A meta-analysis of 106 experiments found that human–AI combinations performed better than humans working alone, but on average worse than the better of the human or AI working independently. The results were particularly weak for decision-making tasks.

Placing a human “in the loop” does not automatically improve the outcome. If the AI produces an answer and the human merely accepts or rejects it, people may follow an incorrect recommendation or overrule a correct one. They are also not always good at recognising whether they or the AI are better suited to a particular task.

Leaders therefore need to design a division of work that takes advantage of their different capabilities.

From prompting to delegation

Much of the current discussion focuses on prompting: how should we ask an AI to perform a task? Delegation goes further. It requires an operating model.

For human–AI work, that means answering several questions:

  • Mission: What outcome are we trying to achieve, and what does good look like?
  • Role: Is the AI researching, generating alternatives, drafting, checking, recommending, deciding or executing?
  • Authority: Which information and tools may it access? May it take action, or only propose one?
  • Boundaries: What must it never do without approval?
  • Handover: What evidence must it provide, and who reviews the result?
  • Escalation: When should it stop because information is missing, confidence is low or consequences are significant?
  • Accountability: Who owns the final outcome and has the competence and authority to intervene?

Consider an AI used in an engineering review. Its job might be to search a defined set of requirements and calculations for inconsistencies, missing information and questionable assumptions. The responsible engineer assesses the findings and decides how to address them. Any resulting design change follows the established approval process. The AI contributes breadth and speed, while engineering judgement and accountability remain where they belong.

That is much clearer than asking an AI to “review the design.”

Beyond a conventional RACI

A RACI can clarify who performs the work, who is accountable, who must be consulted and who should be informed. But AI introduces questions that a conventional RACI was not designed to answer.

An AI may be recorded as the operational executor of a bounded task, but it cannot carry organisational or moral accountability. At the same time, accountability should not simply be transferred to the nearest employee. Researcher Madeleine Elish describes the risk of creating a “moral crumple zone”: a person absorbs the blame for a system failure despite having had little meaningful control over the system.

Human oversight is credible only when the person has the information, time, competence and authority to challenge the AI. This principle is reflected in both the NIST AI Risk Management Framework, which calls for clearly differentiated human–AI roles and executive accountability, and the EU AI Act, which requires oversight of high-risk systems to be assigned to people with the necessary competence, training and authority.

In practice, responsibilities need to be combined with decision rights, operating boundaries, verification requirements and escalation rules.

Governance should be proportionate to the possible consequences. An AI helping to generate workshop ideas can operate within broad boundaries. An AI communicating externally, modifying controlled data or influencing an engineering decision needs much tighter ones.

Designing the team is a leadership responsibility

The job description cannot remain static. AI capabilities will continue to change, as will our understanding of their strengths and limitations. Leaders will need to review the division of work, examine failures and near misses, and adjust responsibilities as the human–AI team learns.

AI makes the design of work itself one of leadership’s most important jobs.

References

AI helped to research and write this article.