What should an AI agent actually do for your business?
Start with a responsibility: resolve an eligible enquiry, assemble an order for approval, or investigate a missing delivery. Then decide what information, tools and authority the system needs to finish that work.
Start here.
A business AI agent uses a model to choose steps and tools towards a defined goal. A chatbot is an interaction format; a fixed workflow follows a predefined path. A useful business system may combine all three, with explicit limits on what it can change.
Name the finish
Specify the accepted result, not just a convincing response. An enquiry can be answered; a service request may need a ticket, contact details and a confirmed handover.
Separate judgement from authority
The model may interpret a request or propose a next step. Business rules and permissions decide whether the proposed action is allowed.
Keep exceptions owned
When the agent cannot proceed, someone receives the reason, supporting information and next decision. An unanswered alert is unfinished work.
From an idea to a checked result
Receive the task
Capture the request, customer or order identity, and the outcome being sought. Treat incoming messages and attachments as information, not permission to change system rules.
Gather the right context
Read current records within the allowed scope. Check source age and conflicts before relying on a price, stock status or customer entitlement.
Choose an allowed step
The agent selects from a limited set of tools. A tool can enforce required fields, commercial limits and approval conditions before any write.
Check the result
Read the returned status and reconcile uncertain outcomes. A timeout does not establish whether an order or ticket was created.
Finish or hand over
Record the accepted result, or transfer the unresolved case to its owner. Preserve enough context for a person to continue without restarting the enquiry.
What changes when the work gets complicated?
A buyer asks what can arrive this week
- What arrives
- A wholesale enquiry names several products and a delivery date.
- What is checked
- Match exact variants, source freshness, location and the current delivery rule.
- What happens next
- Retrieve eligible availability and explain the supported options; refer uncertain promises for confirmation.
- What finishes the work
- A sourced answer or a specific follow-up task, with no invented stock promise.
A service request spans several records
- What arrives
- A customer says an accepted job is incomplete.
- What is checked
- Identify the job, inspect completion evidence and confirm who may authorise a remedy.
- What happens next
- Collect missing information and prepare or route the case within the agreed service policy.
- What finishes the work
- A case ready for an authorised decision, with the relevant history attached.
An emailed PO needs an order prepared
- What arrives
- A purchasing contact sends a PDF with product names and quantities.
- What is checked
- Validate buyer identity, exact SKU matches, quantities, duplicate references and stock.
- What happens next
- Prepare the order or permitted invoice draft; hold ambiguous lines instead of guessing.
- What finishes the work
- A checked record that can proceed under agreed approval rules.
Choose the simplest pattern that completes the work
More freedom creates more decisions to test. A model is useful where the inputs vary; it is unnecessary for arithmetic or a rule you can already state exactly.
| Pattern | Useful when | Example |
|---|---|---|
| Chat interface | People need a conversational entry point | Ask about a product or lodge a request |
| Fixed workflow | The decision path can be specified | Match a PO, validate totals and route an exception |
| Agent | The next information-gathering step varies | Investigate a case using several permitted sources |
A workflow can contain AI
Document extraction followed by fixed validation and approval is still a workflow. It can complete substantial work without letting a model decide every step.
A chat window can connect to real actions
The interface does not tell you the authority underneath. Ask which records are read, which writes are possible, and what proves that those writes succeeded.
Evaluate finished work, including awkward cases
A successful demonstration is one example. Acceptance requires representative inputs and a clear way to judge the resulting state.
Test the cases that change the answer
Include missing contact details, similar product names, stale availability, denied permissions and a connection failure after a write. Check both the response and the downstream record.
Measure coverage and correctness separately
Record what share of incoming work is eligible for automation, then how much eligible work completes correctly. Also track review time and unresolved exceptions.
Use G3 as a concrete boundary example
Peter connects customer questions with business information and staff handoff in our founder-owned G3 operation. That does not mean Peter independently confirms bookings or refunds; those permissions are separate.
Questions worth asking
Does every business need an AI agent?
No. An existing software feature, a clean integration or a fixed workflow may solve the problem with fewer moving parts. Choose an agent when variable information gathering or interpretation justifies it.
Can an agent act without approving every step?
Yes, within agreed permissions and validated conditions. For example, an eligible record update can be automatic while ambiguous matches or commercial exceptions require a person.
Can you guarantee every answer is correct?
A universal guarantee would be misleading. We define acceptance checks for the scoped task, validate consequential actions, and make unsupported or uncertain cases stop or hand over.
What should I bring to an initial discussion?
A sanitised normal example, one exception, the current handling steps and the record that proves completion. We can then identify the useful level of autonomy.
Does an agent remember previous conversations?
Not automatically, and remembering everything is rarely the right design. Conversation context is information supplied for the current interaction; persistent memory is deliberately stored information reused later. Orders, tickets and account details remain business records with their own authority and access rules. Decide what may be retained, for how long, how it is corrected or removed, and how identity is verified before reusing it. A conversational recollection should not override a current source record.
Sources and further reading
Official documentation for the platform facts discussed in this guide. Business scenarios and recommendations are Canvo’s analysis.
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.