What is agentic ai?

Why it matters

Answering is cheap and doing is valuable. Most support cost sits in the steps after the answer: the lookup, the change, the confirmation. Agentic systems attack that cost directly, which is why the category moved from answer bots to action-taking agents in a single product cycle.

What to know

• Every action needs an explicit permission scope, a spending or blast-radius limit, and a durable audit log.
• Reversibility is the design principle. Prefer actions a human can undo over actions that cannot be walked back.
• Multi-step plans fail differently to single answers. Partial completion is the failure mode to test for.
• Deterministic guardrails around a probabilistic planner is the pattern that works in production.
• Confirmations to the customer at each consequential step reduce both risk and support follow-up.

An example

Example: a customer reports a duplicate charge. The agent locates both transactions, confirms they match the duplicate policy, issues a refund within a preset limit, writes the audit entry, and tells the customer the amount and the expected arrival date.

Common mistakes

• Granting broad write permissions before testing narrow ones
• Skipping the audit log because the volume seems small
• Testing only the happy path and not partial failure
• Automating irreversible actions first because they save the most time

Questions