Agents, explained: from hold music to done.
One customer, one annoying problem, three ways to solve it. Watch what changes when software stops just talking about a problem and starts working on it, with a human stepping in exactly where it matters.
The old way: the phone tree
Maya calls the number on the back of her card.
38 minutes in, a person answers, asks for her card number again, and transfers her to disputes. 52 minutes, and it still isn't fixed.
“Thanks. Let me transfer you to our disputes team…”
Maya's path through the menu
- 1 English
- 3 Cards and payments
- 7 "For anything else"
- ↩ Back to the main menu (wrong branch)
- 1 3 2 Disputes
- # Enter your 16-digit card number
- Waiting for the next available agent…
Illustrative numbers for a made-up story.
The chatbot way: a brain without hands
Next time, Maya tries the chat bubble on the bank's website.
It can't. A chatbot reads and writes text, and has no connection to the bank's systems. So it sends Maya back to the phone line.
I was charged twice at Corner Café, $84.20 each. Can you fix it?
I'm sorry about that! Here's how to dispute a charge:
1. Sign in to online banking
2. Open the transaction
3. Choose "Report a problem"
4. Or call us at the number on your card.
Can't you just do it for me?
I'm sorry, I'm not able to issue refunds or open disputes. Please call Card Services at the number on the back of your card. Lines are open 8 a.m. to 8 p.m. Back to the phone queue
A chatbot is a language model in a glass box. It can understand the problem and explain the fix, but nothing in that box can open a ledger or file a dispute.
The agentic way: a brain with hands, and humans in the loop
This time the bank's assistant is an agent: a model that can plan, use tools, and ask people for decisions.
The agent files the refund, checks that the credit actually landed, tells Maya, and schedules a follow-up to confirm the merchant's reversal.
- Confirm it's really Maya (signed-in app)
- Pull this week's card transactions
- Find the duplicate
- Check my authority: refunds up to $50 alone
- Ask Maya to approve the dispute
- Over $50? Escalate to a specialist, then verify the credit lands
- MonGrocer's Market$46.10
- Tue 08:41Corner Café$84.20
- Tue 08:41Corner Café$84.20
- WedCity Transit$2.75
This refund: $84.20, over the limit, so a specialist must approve.
Then the humans step in, exactly where it matters
Northwind
Approve a dispute for $84.20 at Corner Café?
ApprovedCase #48213
- Identity verified in app
- 2 identical charges, same minute
- Duplicate: eligible for refund
- $84.20 is over the agent's $50 limit
- Customer approved
$84.20 credited
Duplicate charge at Corner Café refunded. We'll confirm the merchant's reversal on Friday.
Illustrative numbers for a made-up story.
The race
Same problem, three approaches, one clock.
Fifty-two minutes in, the phone call has only just reached the disputes team, and the refund still isn't done. The chatbot route is in the same queue, two minutes behind. The agent finished 49 minutes ago.
Illustrative times for a made-up story.
Under the hood: chatbot vs agent
Both start with the same language model. The difference is everything around it.
Finally, run it in a loop: plan, act with a tool, observe the result, and go again until the goal is met. That loop is what makes it an agent.
LLM chatbot
Text in, text out. One reply at a time.
Agent
A goal, tools, memory and limits, running in a loop until the job is done.
| LLM chatbot | Agent | |
|---|---|---|
| Does | Answers in text | Takes actions toward a goal |
| Reaches systems | No | Through tools and APIs, with scoped permissions |
| Works in | One reply at a time | A loop: plan, act, check, repeat |
| Remembers | The conversation | The case: who, what, what's done |
| "Done" means | It said so | It verified the result |
| Risky steps | Can't take them, so it sends you elsewhere | Prepares them, then asks a human first |
| When stuck | Hands you a phone number; you start over | Hands a person the full case; nobody starts over |
When should a human step in?
Good agents know their limits. Turn the dial to see who decides.
The customer decides
Anything that moves money, or can't be undone. The agent prepares it, and the customer approves: refunds, disputes, transfers, closing an account.
- Agent alone Look up transactions, explain a fee, check a policy
- Customer approves File a dispute, move money, change a limit
- Specialist approves Anything above the agent's limit (like refunds over $50), policy exceptions, anything unclear
Agents can be wrong too. That's exactly why these checkpoints exist.