Customer Support Resolution Agent
Answer the question, take the action, and know when to ask a human first.
Most support bots can recite your returns policy but can't check an order or issue a refund, so customers still end up waiting in a queue. With BoldAgent, you build an agent that resolves the request end to end. It answers from your own help content, acts through your own systems, and pauses for a reviewer before any action you've decided needs a person's sign-off.
At a glance
Everything this agent covers
- Built for
- Support, customer experience and e-commerce operations teams
- Handles
- Policy and product questions, order status, refunds and escalations
- Runs on
- Your help centre, your app, and an internal page for your support team
- Core features
- Agent Orchestration · Knowledge Libraries · API Connections · Skills · Decision Models · Human approval · Responsible AI
The challenge
The steps are simple. Together they become a queue.
A large share of support tickets follow the same pattern: find the order, read the policy, decide whether the request qualifies, then act. Each step is simple, but together they create long queues and repetitive work for your team.
Giving an AI access to your order and payment systems could solve this, but it raises fair questions:
BoldAgent answers each of these with a specific feature, described below.
-
Control.
What stops it from refunding something it shouldn't?
-
Accuracy.
What happens when it doesn't know the answer?
-
Reliability.
How do you know an instruction change hasn't broken something that worked yesterday?
-
Security.
How do you keep customer data and credentials safe?
Before and after
From reciting policy to resolving the request
| Task | Typical support bot | BoldAgent resolution agent |
|---|---|---|
| Policy questions | Answers, sometimes with invented details | Answers strictly from your knowledge, or returns your fallback |
| Order status | "Please contact our team" | Looks up the order in your system and replies |
| Refunds | Hands off to a person every time | Issues eligible refunds and routes the rest for approval |
| When it's unsure | Guesses | Escalates with a structured summary |
| Improving it | Trial and error | Evaluation datasets, Knowledge Gaps and trace logs |
How it works
A manager routes, specialists resolve
You build a small team of agents rather than one large one. A Manager Agent talks to the customer and routes each request to a specialist Worker Agent, based on usage instructions you write for each worker.
-
Help Centre Answers
Policy, product and shipping questions, answered strictly from your knowledge library
-
Order Desk
Order lookups and refunds through your own order system
-
Escalation
A structured hand-off to your team, with a summary of the issue, what has been tried and what the customer wants
Each worker is an ordinary agent, so it can be tested, versioned and improved on its own, and reused elsewhere.
See it in action
One refund request, from question to resolution
Conversation
CustomerMy order #48213 arrived with a cracked screen protector. Can I get my money back?
AgentSorry about that. I've issued a full refund of $34.00 to your original payment method. You should see it within 5–10 business days, and you don't need to return the item.
Run trace
- Triage. Decision model · Ticket Triage
The manager calls the Ticket Triage decision model, which classifies the request as a refund request with low urgency and no need for a person.
intent: refund requesturgency: lowneeds human: no - Route. Route · Order Desk
The request goes to Order Desk.
Worker Agent - Look up. API action · Get order
Order Desk retrieves the order from your system: one $34.00 item, delivered nine days ago.
$34.00 · 1 itemdelivered 9 days ago - Check. Skill · Refund Procedure
The Refund Procedure skill confirms that damaged items within 30 days qualify for a full refund.
damaged · within 30 daysfull refund - Act. API action · Issue refund
Order Desk issues the refund through your order system.
$34.00 refunded
What if the refund is $240?
The skill flags that the amount is above your approval threshold. The conversation pauses, and a reviewer can approve the refund, edit it, reject it, or respond to the customer directly. The conversation then continues from where it stopped.
What if the answer isn't in your knowledge?
The agent returns your fallback message instead of guessing, and the question is recorded as a Knowledge Gap so you know what content to add next.
What you build
Six pieces, each tested on its own
- 01
Knowledge library
Crawl your help centre, upload policy documents, add exact Q&A answers for questions where wording matters, and import your KB articles. Open any file to check its chunks exactly as the agent sees them.
- 02
API Connection
Register your order system with its base URL and authentication, then add actions such as Get order and Issue refund. Order numbers and amounts are passed as variables, so one action covers every order. Test each action before publishing.
- 03
Refund Procedure skill
Write your eligibility rules and approval thresholds once as a versioned skill and link it to every agent that needs it. Update it in one place when the policy changes. You can also start from a ready-made skill in Plugins.
- 04
Ticket Triage decision model
Define three questions: intent (choice), urgency (score) and needs a human (bool). Every request then gets a consistent, structured classification. Tune it in the built-in test panel before publishing.
- 05
Worker and manager agents
Write clear usage instructions for each worker so routing stays reliable. Pass customer details, the current page URL and the date into each run as context.
- 06
Guardrails
Add human approval to Order Desk, turn on grounded responses for Help Centre Answers, and apply a Responsible AI policy to all agents.
Built-in safeguards
A specific feature for every concern
| Concern | How BoldAgent handles it |
|---|---|
| Unauthorised actions | Human approval puts a reviewer in front of high-value or out-of-policy refunds |
| Made-up answers | Grounded responses return your fallback message when nothing relevant is found |
| Manipulation | Prompt injection detection blocks attempts to talk the agent past its rules |
| Sensitive data | Leakage and secrets detection catch credentials, API keys and sensitive data in requests and responses |
| Harmful content | Content moderation screens for toxic, biased, hateful and adult content |
| Untested changes | Versioning keeps drafts away from customers until you publish |
| Accountability | Audit logs record who changed what and when, and Trace Logging shows exactly what happened in each run |
Measure and improve
Know what works, and fix what doesn't
-
Evaluation
Test the agent against questions with known answers, and re-run the dataset after every change to catch regressions before customers do.
-
Knowledge Gaps
See which questions your content can't answer, grouped into recurring themes, and track how many you've addressed.
-
Conversations
Review every conversation along with the agent version that answered it, so you can tell whether an issue comes from today's configuration or last week's.
-
Analytics
Track credits, tokens and latency by agent and model, so you can see which worker is driving cost or slowing responses.
Deploy anywhere
Where your customers and your team already are
-
Chat Widget
Embedded on your help centre and order pages, styled to match your brand
-
Developer API
Built into your web or mobile app, with streamed responses and image attachments, such as a photo of a damaged item
-
Agent Hub
A hosted page where your support team can use the agents to look up policies while handling tickets
-
Webhooks
Push events from BoldAgent into your own systems
Who benefits
Better for everyone in the conversation
-
Customers get answers and resolutions in one conversation, at any hour.
-
Support agents spend less time on repetitive lookups and more on complex cases, and receive escalations with the context already summarised.
-
Support leads control exactly which actions need approval, and see where content is missing.
-
Developers connect existing systems through API Connections or MCP servers, without rebuilding them.
Take it further
Extend the same agents as you grow
-
Proactive updates
Schedule Order Desk to run each morning and summarise delayed shipments for your team.
-
Exact calculations
Turn on sandbox code execution so partial refunds and pro-rated amounts are computed rather than estimated.
-
More specialists
A manager can delegate to up to ten workers. Add an Account worker for billing questions, or a Product Advisor for pre-sales queries.
-
Generated documents
Enable file generation so the agent can return a refund confirmation or return label.
Frequently asked questions
Questions support teams ask first
Can the agent issue refunds without anyone checking?
Only within the rules you set. You decide which conditions require human approval, such as amount, timing or refund history. Everything that meets those conditions waits for a reviewer.
What if our order system doesn't have an MCP server?
Use an API Connection. It lets the agent call your REST endpoints directly, with no MCP server required.
How do we update the refund policy?
Upload a new version of the Refund Procedure skill. Every linked agent picks it up on publish, with no instruction rewrites needed.
Can we test changes before customers see them?
Yes. Edit a draft version, test it in the Playground and against your evaluation dataset, then publish. The live version stays untouched until you do.
Get started
Five steps to your first resolution agent
- Create a library and add your help content.
- Connect your order system .
- Build your first agent and test it in the Playground.
- Require human approval before sensitive actions.
- Embed the chat widget on your help centre.