Ask a business why they haven't let AI talk to their customers yet, and the answer is almost never "it can't." It's "what if it says the wrong thing?" What if it promises a discount it shouldn't, gives advice it isn't qualified to give, or mishandles an upset customer? That fear — losing control of the conversation — is the real thing standing between most companies and the value AI could deliver.
Guard rules, escalation, and human-in-the-loop are how Mihu answers that fear. Together they let you set exactly what an agent can and can't do, make sure the conversations that need a person reach one, and keep humans firmly in charge of the moments that matter. They're not an add-on for the cautious — they're what makes AI safe to deploy at all.
Why control is the real barrier
An AI agent that's brilliant 95% of the time but unpredictable the other 5% is, for most businesses, undeployable — because that 5% is where reputations and relationships get damaged. A single agent inventing a refund policy, making a promise you can't keep, or being tone-deaf with a grieving customer can undo a hundred smooth interactions.
So the question that actually decides whether AI ships isn't "how capable is it?" It's "how well can I control it?" Get that right and you can hand over enormous volume with confidence. Get it wrong and you can't hand over anything. Mihu is built so the answer is firmly the former.
Three layers of safety
Control in Mihu isn't one feature — it's three working together, each catching what the others don't.
Guard rules
The boundaries. What the agent must always do, must never do, and must never say — enforced on every conversation.
Escalation
The handoff. When a defined condition is met, the conversation moves to the right person — with full context.
Human in the loop
The oversight. People approve, monitor, and step in — so AI handles volume and humans handle judgment.
Guard rules: boundaries that hold
Guard rules are the boundaries you set for an agent — the non-negotiables that apply to every single conversation, no matter how it's phrased or how the customer pushes. You define what the agent must always do, and what it must never do or say.
Always
- Verify a caller's identity before discussing account details
- Disclose that it's an AI assistant when required
- Stay on approved topics and within your policies
- Keep the tone your brand expects
Never
- Promise discounts, refunds, or terms it isn't authorized to give
- Offer medical, legal, or financial advice
- Share information a customer shouldn't see
- Make commitments outside its remit
The result is an agent that's free to be genuinely helpful inside a space you've defined — and simply can't wander outside it. That's the difference between hoping an agent behaves and knowing it will.
Escalation: knowing when to hand off
No matter how capable an agent is, some conversations should go to a human — and the skill is recognizing them early. Mihu escalates whenever a condition you define is met, handing the conversation to the right person or team with the full history attached, so the customer never has to start over.
Notice what the customer experiences: not a cold "transferring you now," but a smooth handoff to someone who already knows the story. Good escalation isn't a failure of the AI — it's the system working exactly as designed.
Human in the loop: people stay in charge
Human-in-the-loop is the principle that people remain in control of what matters, even as AI handles the volume. In Mihu that takes a few practical forms:
- Approval before action. Require a human to sign off before the agent does something sensitive — issuing a credit, confirming a high-value booking, sending a particular message.
- Live monitoring and takeover. Watch conversations as they happen and step in instantly when you want to.
- Agent-requested input. The agent can pause and ask a human for a decision mid-conversation, rather than guessing.
- Review and coach. People review how agents handled things and refine the rules — so the system gets safer and sharper over time.
AI for volume, humans for judgment
The goal isn't to remove people — it's to spend their time well. The agent handles the hundreds of routine conversations so your team is free for the handful that genuinely need a human's judgment, empathy, or authority.
How easy it is in Mihu
Powerful control is worthless if it takes an engineer a week to configure. In Mihu, you set all of this the way you set everything else — by describing it. "Always verify identity before sharing account details." "Never quote a final price; escalate pricing to sales." "If a customer is frustrated or asks for a person, hand off to the duty manager with the full history." You say it; Mihu applies it.
Because guard rules, escalation, and human-in-the-loop are built into the platform rather than bolted on, they work consistently across voice, WhatsApp, and every other channel — and you can tune them anytime through the Mihu Assistant, or build deeper custom logic with the Builder Assistant when a rule gets truly bespoke.
Use cases by industry
Automotive
Guard rule: never commit to a final price or discount. Escalation: route any complaint or negotiation to the service or sales manager. The agent books services and answers questions all day, but the moment money or a frustrated customer enters, a human takes the wheel.
Healthcare & clinics
Guard rule: never give medical advice, and always verify identity. Escalation: any clinical question or distressed patient goes straight to qualified staff. The agent handles scheduling and reminders safely, and never strays into territory that requires a clinician.
Financial & insurance
Guard rule: no financial advice; verify before discussing any account. Human in the loop: require sign-off before anything is committed. Disputes and sensitive cases escalate immediately — so the agent informs and assists without ever crossing a regulatory line.
E-commerce & retail
Guard rule: follow the published returns and refund policy exactly — no improvising. Escalation: hand off anything outside policy or any unhappy VIP. Customers get fast, consistent answers, and exceptions still get a human's discretion.
Why it works
The payoff of getting control right is counterintuitive: strict boundaries let you automate more, not less. When you trust that an agent can't go off the rails and that the risky moments will always reach a person, you can confidently route huge volumes of conversation to AI — because the downside is contained.
It also makes for a better customer experience, not a more robotic one. Customers get fast, consistent, on-policy answers for the routine, and a well-briefed human exactly when their situation calls for one. And it keeps you compliant by default, since the rules you're required to follow are enforced on every conversation rather than left to chance.
Control is what unlocks scale
Guard rules, escalation, and human-in-the-loop aren't the brakes on your AI — they're the seatbelt that lets you drive fast. They're the reason you can hand a contact center to AI and sleep at night.
Frequently asked questions
What are guard rules in Mihu?
The boundaries you set for an AI agent — the things it must always do, must never do, and must never say. They keep agents on-script and within policy: required disclosures, topics to avoid, promises it can't make, identity checks before sensitive actions, and tone.
How does escalation work?
When an agent hits a condition you define — a frustrated customer, a complaint, a sensitive or out-of-scope request, or an explicit ask for a human — it hands the conversation to the right person or team with full context, so the customer doesn't have to repeat themselves.
What is human-in-the-loop?
It means people stay in control of the moments that matter. Humans can require approval before certain actions, monitor conversations live and take over, and be asked by the agent for input mid-conversation — so AI handles volume while people handle judgment.
Why are guard rules and escalation important?
They're what make AI safe to put in front of customers. They protect your brand, keep you compliant, and ensure the conversations that need a human reach one — which is what gives teams the confidence to deploy AI at scale.
Is it hard to set up?
No. You describe the rules and escalation conditions in plain language and Mihu applies them. Guard rules, escalation, and human-in-the-loop are built into the platform rather than bolted on afterward.
Hand over the volume. Keep the control.
Set the rules, define the handoffs, keep your people in charge of what matters — and let Mihu safely take the rest.
See it on your own rules Guard rules · Escalation · Human-in-the-loop · Set up in plain language