For a small business, an AI agent is software that answers a customer and also does something about it: books the slot, drafts the quote, looks up the order. It works best on repetitive tasks with clear rules and a system it can check. Anything involving money, complaints or judgement should still reach a person.
| Job | Needs access to | Hand to a person when |
|---|---|---|
| Booking appointments | Your calendar or booking system | No slot fits, or the request is unusual |
| Answering order status | Read-only order lookup | The parcel is lost or the customer is upset |
| Drafting quotes | Price list and rules | Always, before the quote is sent |
| Qualifying leads | A short list of questions | The lead is a good fit and wants to talk |
Updated
Most owners do not need a general assistant; they need one annoying, repeated task taken off their plate. Pick a job below to see what the agent would do, which tools it needs, how risky it is and the moment it should step aside. Then read the honest limits and the cost section before you sign up for anything.
Each job shows what the agent does, the access it needs, its risk level and the point where a person takes over.
What it does: Asks for the order number and email, looks the order up and reports where it is.
Access it needs: Read-only order lookup. No ability to change orders.
Hand to a person when: The tracking shows no movement for days, or the customer asks for a refund.
Risk: Low
What it does: Answers from a single document of your hours, delivery areas and return rules.
Access it needs: None beyond the document.
Hand to a person when: The question is not covered by the document.
Risk: Low
What it does: Offers open slots, books one and sends the confirmation.
Access it needs: Calendar or booking system with create and move rights, no delete.
Hand to a person when: A group booking, a special request, or a cancellation with a fee.
Risk: Medium
What it does: Asks three to five questions about budget, timing and need, then saves the answers.
Access it needs: Write access to a leads list or CRM.
Hand to a person when: The lead fits and wants a call: book it with you directly.
Risk: Medium
What it does: Builds a draft from your price list and the customer's answers.
Access it needs: Price list read access; drafts saved, never sent automatically.
Hand to a person when: Every time: you check and send the quote yourself.
Risk: High: draft only
What it does: Collects the details and a summary of the problem.
Access it needs: Read-only order lookup. No payment access.
Hand to a person when: Always, after collecting details. Money and apologies come from a person.
Risk: High: draft only
Most agent builders and frameworks let you express rules like this, either in settings or in code. Adjust the words and amounts to your business.
Plain, checkable conditions. The agent passes the conversation to a person when any one matches.
{
"handoff_when": [
{ "customer_says_any": ["refund", "complaint", "manager", "lawyer"] },
{ "order_value_over": 200 },
{ "agent_unsure_twice": true },
{ "action_in": ["refund", "cancel_paid_order", "delete_customer"] }
],
"handoff_message": "I'll pass this to a member of our team now. You'll hear back within one business day."
}A customer forgives an assistant that says "let me get someone for you". They do not forgive one that confidently refunds the wrong order. Security guidance from OWASP recommends requiring a human to approve high-impact actions, and for a small business that list is short and obvious: money out, anything deleted, anything promised.
Write the handoff rule as plainly as a note for a new employee. If you cannot write it, the task is not ready to automate.
Compare the agent against what the task costs you now: minutes per request times requests per week. Then add the setup time, the review time and the monthly bill. If the result is close, pick a smaller task first.
Typical jobs are booking and rescheduling, answering where an order is, drafting quotes from a price list, and asking new leads a few qualifying questions. Each one combines a conversation with a lookup or action in a system you already use.
They can state wrong facts with confidence, follow instructions hidden in text they read, misjudge an angry customer, and take an action they should have asked about. Narrow tools and a clear handoff rule limit the damage.
There is usually a platform fee plus usage, often billed per conversation, per message or per credit. Long conversations and agents that loop through many steps cost more. Ask any vendor how a single typical conversation is billed, and test with your own volume.
NIST's AI Risk Management Framework is a free, voluntary framework many organisations use as a guide. It is not a law, but it gives a structured way to think about risks before and after launch.
Related: AI agent test checklist, AI agent script template, AI agent vs agentic AI, No-code AI agent builder, How to build an AI agent.
Start for free or see all free tools. We.Inc is free to start, with no card needed. AI building is on paid plans from $20 a month.
We.Inc is an AI-powered website builder you can resell under your own brand. Launch a branded client dashboard, bill on Stripe Connect, and deliver AI-generated websites in minutes. White-label plans start at $99 a month for 25 client sites, with a 7-day free trial and no per-site fees.