AI agents
AI Agent
What is an AI agent?
- Given a goal, plans and uses tools on its own
- Goes beyond answering to actually acting
- Carries out multi-step tasks
When AI agents come in, the first thing to change is the shape of the work order. Traditional automation only moved when a person spelled out every step, such as 'open screen 1, copy the value, paste it into sheet 3.' With an agent, you give instructions in terms of the deliverable, such as 'summarize this week's outstanding receivables and prepare a draft report for the team lead.' A manager's job shifts from supervising procedures to reviewing results and designing permission boundaries.
In practice, it helps to sort agents by level of autonomy. There are assisted agents, where a person clicks approve at every step; semi-autonomous agents, which run on their own within a set scope and only report results; and fully autonomous agents, which even judge exceptions themselves. Most deployments in Korea fall between assisted and semi-autonomous. A chatbot ends after a single question and answer, and RPA follows fixed screens and coordinates, whereas an agent reads the situation and chooses which tool to use. That is the decisive difference.
That said, it doesn’t fit every task. The target system must be accessible through an API or at least a view-only screen, logs must be kept of which tools were called and how many times, and wrong actions must be reversible. Rather than attaching ten tools from the start, it is safer to begin with 3 to 5 tools and about 5 to 10 steps, see where it fails, and then expand.
The most common misconception is that 'since it works on its own, we can cut back on review.' Even at 95% accuracy per step, chaining 10 steps drops the final success rate below 60%. And if you leave write permissions open without approval, such as sending email or editing inventory, you get incidents like a quote with a wrong unit price going out automatically to twenty client companies. Rollouts that don’t first decide at which step a human steps in mostly stall within three months.
John KimInstructor John Kim, author of B2B Marketing Design and CEO of Nanum Consulting
Seolyeong ParkGenerative AI and AX training specialist with 17 years of hands-on experience and 200+ lectures, helping non-technical staff put AI to work right away| Category | Agent | Copilot |
|---|---|---|
| Role | Runs multiple steps on its own | Assists alongside people |
| Leads | AI-led | Human-led |
| Yes | Researches and processes automatically | Helps write documents and code |
It emerged from the push (since 2022) to combine 'tool use' and 'step-by-step reasoning' in LLMs. Techniques such as ReAct, which alternates between reasoning and acting, were a key catalyst.
Let’s look at a hypothetical case: Company A, a home appliance distributor with 180 employees and annual revenue of about KRW 90 billion (≈ US$67,335,000). Its 6-person CS team handles 3,200 delivery and return inquiries a month. Because they keep switching between the order management system and the courier tracking screen, average handling time per case exceeded 7 minutes. Ahead of the Q3 peak season, the CS lead proposed an agent: 'Let’s try this once before hiring more people.'
- Park, a deputy manager on the customer support team, sorted 9,600 support records from the last three weeks by type and put a hard number on it: delivery tracking, return requests, and exchange guidance make up 68% of all inquiries.
- The IT team connected the order management system and the courier API with read-only permissions only, and locked refunds and customer emails so nothing goes out until a support rep clicks the approval button.
- When an inquiry like 'When will the order I placed yesterday arrive?' came in, the agent ran a five-step sequence on its own: find the order by customer number, check the tracking status, and put a draft reply on the support rep's screen.
- When the error rate came in at 11% in the first two weeks, Deputy Manager Park added an exception rule: 'always hand scheduled-delivery and partial-cancellation cases to a person.' By week three, the error rate had dropped to about 3%.
- Three months later, average handling time per case fell from 7 minutes 12 seconds to 2 minutes 40 seconds, and the team handled 3,200 cases a month without the three contract hires it had planned for peak season. One support rep moved into a dedicated reply-review role.