What the agent does
A support queue is a pipeline. A message arrives, somebody decides what it is, somebody answers it, somebody decides whether anything in a system needs to change, and somebody owns the outcome. Most automation covers the second step and leaves the rest on the team.
An agent covers the whole path. It classifies the request, pulls the answer from your live data, takes the bounded action the policy allows, records what it decided and why, and escalates what it should not touch. Resolution is the job. Answering is one step of it.
An agent vs a chatbot
The market sells both under the same label. Six dimensions separate them, and the last two decide what it costs when you guess wrong.
| Dimension | A chatbot | An agent |
|---|---|---|
| What it does with a request | Answers it | Resolves it |
| Where the answer comes from | The documents it was pointed at | Your data plus your systems, checked live |
| What it can change | Nothing. It talks | Bounded actions: refunds, credits, plan changes inside limits |
| What a mistake looks like | A wrong answer the customer ignores | An action taken you may not notice until it lands |
| What it hands to a human | A transcript | A resolved case, or a full-context escalation |
| What the record shows | What was said | What was decided, why, and under which policy |
What the agent handles
Triage and routing
Reads the incoming message, classifies it, and sends it to the right queue or resolves it outright. Misroutes stop being the first failure.
Tier-one resolution
Order status, policy questions, account how-tos. Answered from your live data with sources, in seconds, at any hour.
Bounded account actions
A refund under a limit, a plan change inside policy, a credit within a cap. The agent acts where the rules are written down.
Escalation with context
Outside the boundary, the agent hands over the full record: what was asked, what it checked, what it would have done. The human starts informed.
Where the boundary sits
A support agent that can act can also act wrong. The difference between a deployment and an incident is whether the limits were written down before launch. Every agent we ship runs inside an authority boundary expressed in actions: refund amounts, plan tiers, which systems it may write to.
Each action leaves a decision record carrying the verdict, the reason, the policy version, and the message it acted on. Outside the boundary, the case goes to a person with the full record attached, the human in the loop that makes the whole system auditable. This is the governed version of a support agent, and it is the only version worth deploying where money moves.
Where Flytebit sits
We run this in production already. Our banking support agent resolves 78% of conversations without a human, handles around half a million conversations a month, and operates inside PCI-DSS. The same stack sits behind our conversational builds and the agentic systems we ship for engineering pipelines.
Our own products are agents built the same way. PASSR acts on every pull request inside a bounded review scope, posts findings, and tracks them across commits. TESTR generates and runs tests against code it did not write. DOCKR regenerates documentation on every merge rather than letting it drift. Each acts inside limits rather than stopping at an answer. The support agent is that pattern aimed at a queue instead of a repository.
An engagement starts with a feasibility study that maps your queue: which share is bounded work, which systems the agent would touch, and where the boundary should sit. Where regulated workflows are involved, the governance and risk work sets the controls before anyone builds.
What to ask a supplier
Five questions that separate a real support agent from a chatbot with a new label. They work on any supplier, including us.
Does it act, or does it only answer?
Most "agents" on the market are chatbots with new labels. If it cannot change anything in your systems, it is a chatbot. Price it accordingly.
Where is the boundary written down?
In actions, not topics. "Refunds under $200" is a boundary. "Help with billing" is not.
What happens when it cannot decide?
Escalating with full context is a design choice. Guessing is a defect. Ask which one the system was built to do.
What does the record show?
The verdict, the reason, the policy version, the customer message it acted on. A chat transcript is not a decision record.
Can a wrong action be undone?
Refunds can be reversed. Sent messages cannot. Ask which actions the agent may take and which it may only prepare.
Common questions
Are AI agents for customer service overhyped? ▾
Mostly yes, because most products sold as agents are chatbots with new packaging. A chatbot answers questions. An agent resolves requests, which means it acts inside your systems under limits you set. The overhyping is real and the capability is also real. The test is simple: ask what the system changes when it responds.
What is the difference between a chatbot and an AI agent? ▾
A chatbot talks. An agent acts. The chatbot can tell a customer your refund policy. The agent can check the order, apply the policy, issue the refund inside its authority boundary, and record why. Resolution is the difference.
When should a team use AI agents in support? ▾
When two things are true at once: the volume is past what a team can read carefully, and a share of the queue is bounded work. Order status, policy answers, account changes inside limits. If every ticket needs judgment, an agent is the wrong buy.
What can an AI agent safely do in a support queue? ▾
Whatever the written policy already allows. Refunds under an amount, plan changes inside contract terms, credits within a cap, status updates. Outside the boundary it escalates to a human with the full record attached. The boundary is a design decision, not a hope.
How much does an AI agent for customer service cost? ▾
Scoping starts with a feasibility study from $2K, which maps the queue, the systems the agent would touch, and the boundary it needs. The build is priced from those findings, so the number reflects your volume and your integrations rather than a rate card.
Will AI agents replace support teams? ▾
They take the bounded share of the queue. Judgment calls, angry customers, edge cases, and anything outside the written policy stay with people. Teams that deploy well end up smaller in tier one and busier in everything above it.