An AI agent that talks to your customers speaks for your business. That is a good reason to design its limits deliberately, rather than hoping it behaves. Guardrails are the rules and checks that keep an agent useful and safe. "Human-in-the-loop" means a person stays involved where it matters.
What can go wrong
- Wrong answers stated confidently.
- Promises you cannot keep, such as discounts, delivery dates or refunds.
- Sharing information it should not.
- Handling a sensitive situation badly.
- Repetitive or annoying behaviour that damages trust.
- Taking actions in your systems that should have been approved.
None of these are reasons to avoid agents. They are reasons to design the boundaries.
Seven guardrails worth having
1. Define the scope
Write down exactly what the agent handles. For example: "Answers questions about opening hours, prices and delivery, and books appointments." Everything else goes to a person.
2. Ground it in your information
Give the agent an approved source of truth: your price list, policies and FAQs. Instruct it to say it does not know rather than guess. Keep that source up to date.
3. Limit what it can promise
Set hard limits. It cannot offer discounts above a threshold, confirm refunds or commit to delivery times unless a system confirms them. Anything outside the limit needs approval.
4. Set clear escalation triggers
Hand over to a person when:
- The customer asks for one.
- The customer is upset or the topic is sensitive.
- The agent is unsure or has failed to help twice.
- The request involves money, legal matters or safety.
- A high-value deal is involved.
When it hands over, the person should see the full conversation and a short summary. See how we handle the same principle in chatbot hand-off design.
5. Require approval for consequential actions
Distinguish between reading and doing. An agent looking up an order is low risk. An agent issuing a refund or changing a record deserves an approval step, at least at first.
6. Protect personal data
Give the agent access only to the data it needs. Follow privacy laws that apply to you, be clear with customers about how their data is used and keep records of consent. Our note on opt-in practices covers messaging consent.
7. Log everything and review it
Keep transcripts and actions. Review a sample every week, looking for wrong answers, awkward phrasing and missed escalations. Fix the instructions or source material, then check again.
Be transparent
Customers should not be misled into thinking they are talking to a person. A short line such as "I am an automated assistant. I can connect you to a colleague at any time" builds trust and reduces frustration.
A staged rollout
- Shadow mode. The agent drafts replies; a person approves each one.
- Assisted mode. The agent answers simple, low-risk questions alone; a person handles the rest.
- Expanded mode. Widen the scope as results prove reliable.
Moving through these stages builds confidence for both your team and your customers.
Who is responsible?
Assign a named owner for each agent. They approve changes, review results and decide when to expand or pause. Accountability should never be "the AI did it".
A guardrail checklist
- Scope written down.
- Approved knowledge source.
- Limits on promises and actions.
- Escalation triggers and a hand-off summary.
- Approval for consequential steps.
- Data access limited.
- Logging and weekly review.
- Clear disclosure to customers.
- Named owner.
ForGrow AI agents are designed to work within rules you set and hand over to your team when needed. You can see the agents, and read our overview of what AI agents are.
Frequently asked questions
What is human-in-the-loop?
It means a person reviews, approves or takes over at important points, instead of the agent acting entirely alone.
Can guardrails make an agent less useful?
Well-designed guardrails make it more useful, because customers and staff can trust it. Start strict and loosen as results improve.
How often should I review agent conversations?
Weekly at the start, then less often once results are stable. Always review after changing instructions.
Meet ForGrow AI
ForGrow AI is our line of AI agents for voice, finance, sales, support, hiring and marketing, designed to work within your rules and hand over to your team when needed.