AI agents that can answer, act and hand work over.
Build around a real job: handle an enquiry, progress an order or turn a staff question into an assigned task. Canvo connects the information and tools needed to complete that job within your authority rules.
A system that fits the work.
A business AI agent can interpret a request, retrieve relevant information and use approved tools to take action. Its scope must define what it may change, how success is checked and when a person takes over. A fluent answer alone does not complete a business process.
What changes when the work gets complicated?
A customer wants to change a booking
- What arrives
- The customer requests a different time in a conversation.
- What is checked
- Customer identity, current booking, availability and approved change rules.
- What happens next
- Make an eligible change through the booking system, then check its confirmation; route exceptions to staff.
- What finishes the work
- A verified change or a clear handover. A suggested time is not a confirmed booking.
A staff question reveals an unresolved fault
- What arrives
- A staff member asks what to do about equipment that will not start.
- What is checked
- Equipment identity, approved troubleshooting steps and existing open issues.
- What happens next
- Provide relevant guidance and create or update the issue for the responsible team.
- What finishes the work
- A traceable task and next step. Recording the issue does not prove a physical repair.
An account enquiry needs a system action
- What arrives
- A customer asks for the status of an incomplete order.
- What is checked
- Access to the order, current fulfilment events and permitted follow-up actions.
- What happens next
- Explain the verified status and raise the authorised warehouse enquiry if evidence is missing.
- What finishes the work
- An informed reply and an owned follow-up, without inventing a delivery promise.
Business knowledge and access
Approved information sources, freshness rules and access appropriate to the user and task. Unsupported answers have an explicit fallback.
Actions with result checks
Tools for the agreed work, with validated inputs, permission checks, duplicate protection and confirmation from the receiving system.
A usable handover
Retain the request, evidence and next decision when staff take over. Test realistic conversations, exceptions and recovery paths.
How the work moves through the system
Understand the request
Identify the task and required details without guessing missing facts.
Read relevant information
Retrieve current records and approved guidance for that request.
Check permission and rules
Validate eligibility, limits and whether the action is authorised.
Act through the tool
Perform the permitted action or send an explained exception to staff.
Confirm the result
Check the resulting record before reporting completion to the user.
Define the job before choosing how much freedom to give it.
The business decides the action permissions. The model does not get to expand them because a message asks it to.
| Capability | What it does | Completion check |
|---|---|---|
| Knowledge assistant | Finds and explains approved information | Answer matches the available source |
| Task agent | Uses permitted tools to progress a request | Receiving system confirms the action |
| Human handover | Transfers an exception with context | The responsible team can continue the work |
Keep exact rules outside the conversation
Prices, eligibility, permissions and transaction limits should use validated records and explicit checks.
Use a fixed workflow when it fits
A predictable sequence may need conventional automation. An agent becomes useful when interpreting the request changes the next valid step.
Test useful outcomes, including a sensible refusal to act.
Use representative requests, missing information and conflicting instructions. Review the result and the actions taken, not just the wording.
Measure service and staff effort
Track correctly completed requests, failed actions, repeat contacts and human handling time. A confident answer can still be wrong.
G3’s implemented customer layer
In our founder-owned business, Peter uses product and venue information and supports retained tickets and staff replies. Peter itself does not confirm bookings or complete refunds; the booking scenario here illustrates a separately scoped build.
Questions worth asking
Is an AI agent the same as a chatbot?
A chatbot describes a conversational interface. An agent may also choose and execute permitted tool actions. Some businesses need better answers; others need actions and cross-system follow-through. The scope should say which result is required.
Does someone have to approve every action?
No. Eligible actions can run automatically inside tested permissions. Missing evidence, failed checks and requests outside those boundaries require the agreed exception path. More authority is a business decision, not a model setting.
What does a first agent project cost?
Start with a free triage. The A$950 ex GST diagnostic is credited to implementation under the agreed scope; a bounded first workflow starts at A$4,900 ex GST. Channels, connected tools, access boundaries and evaluation needs determine the final scope.
Do we need an ongoing Canvo subscription?
Implementation includes 12 months from acceptance for agreed workflow defects and minor compatibility changes within existing integrations. No Canvo subscription is required. Optional proactive care starts from A$490/month ex GST with separately agreed scope; covered repairs do not require that plan. Software, API, model and hosting costs are separate. The proposal records coverage and responsibilities.
Can one agent work across our website, email and messaging channels?
A build can share approved knowledge and handover records across several channels where their interfaces and permissions support it. Each channel still needs its own access, identity, attachment and reply checks. We confirm the required channels and supported actions during scoping; connecting a website chat does not automatically establish support for every messaging platform.
How does the agent know which customer it is allowed to help?
Account-specific information and actions need a verified session or an agreed identity check matched to the sensitivity of the task. A typed name or email address alone is not proof of access. The system should retrieve only records that identity may use; if verification fails, it can answer public questions or hand over without exposing account details.
Bring one process. We’ll map the next step.
Start with the input, the result your team needs and an exception that keeps coming back. Estimate the value, then check the assumptions with George.
12 months of support from acceptance covers defects in the agreed implementation and minor compatibility changes to existing integrations. Optional proactive care from A$490/month ex GST and further development are scoped separately. Third-party software, hosting and API costs are separate.