- 16 de October, 2026
- Posted by: Filipa Ferreira
- Category: Artificial Intelligence
The main challenge in adopting AI agents is not simply technological. It is organisational. Companies will capture limited value if they introduce agents without redefining roles, decision rights, workflows and accountability.
AI agents should be managed as participants in a system of work. Their responsibilities must be explicit, their authority limited and their performance visible.
Technology creates capability. Work design turns that capability into value.
1 Define the role before activating the agent
Every agent needs a clear job description. Leaders should specify what it is expected to do, what remains outside its remit and where human judgement is mandatory.
The role should also define decision rights, approval thresholds, escalation points and handovers between people and systems. Broad instructions create ambiguity; precise mandates create control.
An agent without boundaries is an unmanaged risk.
2 Design agents around real human friction
Adoption becomes easier when agents solve recognisable workplace problems. The priority should be work that is repetitive, predictable or unnecessarily time-consuming, while preserving human involvement where context, empathy or judgement matter.
The objective is not simply to automate activity. It is to improve the quality of human work.
Remove friction before adding complexity.
3 Measure outcomes not technological novelty
AI agents require measurable performance standards linked to process outcomes. Accuracy matters, but so do timeliness, reliability, consistency and the quality of escalation.
Regular review allows managers to identify weaknesses, refine instructions and distinguish acceptable variation from genuine failure. Feedback should improve both the agent and the process in which it operates.
What cannot be evaluated cannot be responsibly scaled.
4 Keep a human accountable
Every agent should have a named human supervisor. That person must oversee its mandate, permissions, training, exceptions and interaction with other human and digital participants.
Human oversight is not a temporary safeguard. Organisations remain responsible for the decisions and consequences produced through the systems they use.
Delegation does not transfer accountability.
5 Introduce gradually and expand on evidence
New agents should begin with a limited scope, close supervision and clear performance thresholds. Access and autonomy should increase only after consistent results have been demonstrated in the organisation’s own operating context.
Capability claims are not evidence of readiness. Trust should be earned through observed performance.
Pilot with discipline. Scale with proof.
6 Give each agent a clear identity
A distinct name makes an agent’s role easier to discuss, trace and evaluate. The purpose is not to humanise the technology, but to prevent responsibility from disappearing behind the vague statement that a decision came from AI.
A clear identity should always be connected to a defined role, an accountable owner and a known scope of authority.
Clear identity creates visible responsibility.
The management challenge
Embedding AI agents requires more than a compelling business case. It requires a management system that connects roles, workflows, controls, performance and accountability.
Organisations should begin with a small number of well-defined agents, learn from their performance and expand only when the operating model is ready.
AI agents should earn trust through performance not inherit it from technology.
The future of agentic AI will be shaped as much by management discipline as by technical progress.
Adapted for PTCG from: Joseph Fuller, Create an Onboarding Plan for AI Agents, Harvard Business Review, 25 March 2026.