Agentic AI: Autonomous Planning and Where It Breaks
What happens when AI stops merely answering and starts deciding what to do next? The promise of agentic AI is seductive: give it a goal, connect it to tools, and let it plan the path - but the same autonomy that makes it powerful also creates the places where it breaks.
- Agentic AI is AI that can plan actions, use tools, observe outcomes and adjust its next step toward a goal.
- The core loop is: goal - plan - act - observe - replan.
- It is useful when work has repeatable goals, digital tools, clear constraints and measurable outcomes.
- It breaks when goals are vague, tools are unsafe, data is incomplete, feedback is delayed or accountability is unclear.
- Managers should control agents through permissions, checkpoints, audit logs, test environments and human escalation rules.
- The interview-safe answer is not “AI will automate everything”; it is “autonomy must be matched with governance.”
Big Picture
Think of agentic AI as a junior analyst with software access: it does not only generate a recommendation, it can break a goal into tasks, call tools, check results and decide the next move. The management question is not “Can it act?” but “Where should we allow it to act without creating unacceptable risk?”
Core Explanation: How Autonomous Planning Works
An ordinary chatbot gives a response. An agentic system chooses a sequence of actions. That difference sounds small, but it changes the managerial risk profile completely.
For example, a procurement chatbot may answer, “Supplier A has the lowest quoted price.” A procurement agent may read past contracts, compare suppliers, draft an RFQ, send it for approval and update the sourcing tracker. That shift from advice to action is where agentic AI becomes valuable - and dangerous.
The simplest architecture has five moving parts:
A clean MBA answer must separate capability from authority. The model may be capable of drafting a purchase order, but the business may authorize it only to prepare a draft, not release the order.
Where Agentic AI Works Best - and Where It Breaks
Agentic AI performs best in bounded, digital workflows where the goal is clear and mistakes can be detected quickly. It struggles in ambiguous, high-stakes environments where the agent has to infer intent, handle exceptions or make commitments on behalf of the firm.
The useful zone is not “everything AI can technically do.” It is the zone where the agent has enough structure to act and enough guardrails to avoid costly surprises.
In operations, agentic AI becomes especially powerful when connected to demand signals, inventory records and reorder logic. If this area is unfamiliar, revise Using AI for Inventory Optimisation and Replenishment before applying agentic workflows to stock decisions.
The Manager’s Control Framework
The right way to deploy agentic AI is to decide the autonomy level first. Do not start with “What can the model do?” Start with “What is the maximum safe action it should be allowed to take?”
A practical governance design has four controls:
This is also where procurement examples become concrete. A sourcing agent may analyze spend, cluster suppliers and draft RFQs, but contract release should still follow the approval logic covered in Using AI in Spend Analysis, Sourcing & Contract Review.
Definitions
Agentic AI is AI that plans, uses tools, observes outcomes and adapts actions to achieve a goal with limited human prompting.
- Autonomous planning: breaking a goal into ordered tasks and choosing the next action based on context.
- Tool use: allowing the AI system to call external software such as search, ERP, CRM or email.
- Human-in-the-loop: a design where people approve, correct or stop important AI actions.
- Guardrail: a rule, constraint or control that limits unsafe AI behavior.
- Escalation: routing a case to a human when risk, ambiguity or uncertainty crosses a threshold.
How to Measure Whether an Agent Is Working
Do not evaluate an agent only by how impressive its output looks. Evaluate it like an operating system: completion, quality, cost, risk and control.
A simple worked example: suppose a claims-processing agent handles 1,000 requests in a pilot. It completes 720 correctly, escalates 220 to humans and makes 60 incorrect completions. Task success rate is 720 / 1,000 = 72%. Human intervention rate is 220 / 1,000 = 22%. The key question is not whether 72% sounds good in isolation; it is whether it improves cost, speed and quality versus the current process without increasing risk.
Case Study: Air Canada and the Cost of Uncontrolled AI Commitments
Air Canada’s chatbot dispute shows why AI systems that interact with customers need clear authority limits, policy grounding and escalation rules.
The situation was simple and powerful. A customer asked Air Canada’s chatbot about bereavement fares. The chatbot gave information that was inconsistent with the airline’s actual policy, and the customer later relied on that information. In 2024, the British Columbia Civil Resolution Tribunal held Air Canada responsible for the chatbot’s misleading response in Moffatt v. Air Canada, 2024 BCCRT 149.
This is not a classic “agent autonomously placed an order” case. It is more important: it shows the boundary problem. Once an AI system becomes the customer-facing representative of a company, even a generated answer can behave like a commitment.
The primary failure was authority without sufficient control: the system could communicate policy-like guidance without ensuring the response matched official policy. Supporting drivers included weak grounding in authoritative policy, insufficient escalation for sensitive fare rules, and unclear customer-facing disclaimers. The lesson for agentic AI is direct: if an agent can affect customer expectations, money, rights, service commitments or compliance, it needs stronger guardrails than a simple conversational bot.
An Indian parallel is easy to imagine in banking, insurance, travel or quick-commerce support. If an AI service agent tells a customer that a refund, claim, cancellation or warranty is approved, the business may face financial, regulatory and trust consequences. In India, that makes human approval, consent, auditability and escalation especially important in workflows touching payments, KYC, insurance claims or grievance handling.
How AI Changes Agentic AI
Agentic AI itself is changing quickly because the underlying models, tools and governance layers are becoming more capable. In 2026, three shifts matter for managers.
A practical student workflow: load a company’s annual report, operating process notes and this lesson into NotebookLM, then ask it to generate “10 workflows where an agent could help, the required tools, the approval gates and the failure risks.” Use the output to prepare a sharper interview answer, not as a substitute for your own judgment.
Interview Relevance
“Our company wants to use agentic AI in customer support or operations. How would you decide what to automate and what to keep human-controlled?”
Use the phrase “graded autonomy”. It signals that you understand AI should move from insight to recommendation to action only as controls mature.
Common Mistake
The biggest mistake is treating agentic AI as “ChatGPT plus automation.” That answer ignores permissions, liability, auditability and exception handling. The one-line fix: always discuss the agent’s goal, tools, autonomy level, guardrails and failure mode together.