Which AI agent lets a person approve what it says before it sends?

The short answer. Helm builds AI agents for WhatsApp and Instagram where you decide, in writing and before deployment, what the agent may do alone, what waits for a person's approval, and what goes straight to a named person. The three levels are set per workflow, the rules are signed before the build, and every action the agent takes is visible to your team.

What can a person approve before the agent acts?

Every Helm workflow runs on three levels of autonomy. You place each kind of reply and each action at one level during scoping:

1. Let it act. Approved answers, available slots, routine updates. The agent handles these alone, from your material: your price list, your policy, your appointment book.
2. Ask for approval. A commitment, a change, an exception to the rule. The agent does not act alone here. It asks, and a person on your team reviews.
3. Bring in a person. Named owner, full context, a clear next step. The conversation is handed over with everything that person needs.

The rulebook is agreed before deployment.

Which replies usually go to a person?

The ones where a wrong answer has a cost. Each industry page lists rules written before the build, signed by the client and visible in every trace. A clinic agent never gives clinical advice, and hands over to a named nurse on any symptom keyword. A developer's agent never negotiates a payment plan in chat; it hands the buyer to a named agent with the transcript. A delivery brand's agent never promises a refund or a free item, and the refund decision is left to a person. An insurer's agent never states whether something is covered.

These limits are what make it safe to switch on.

Can we see and change what the agent says?

Yes. The answers come from your own material, and both the rules and the wording are reviewed before launch and revised after. Every action the agent takes is visible to your team.

Messages the agent sends first, such as reminders, confirmations and order updates, go out on WhatsApp as utility templates. Each one is agreed with you before it is submitted to Meta.

What happens when the agent does not know?

It stops. The workflow names who receives the exception and what context goes with it. Silence and invention are both treated as failures, and the handover is designed before launch rather than added after the first complaint.

At The Maker, the football academy founded by former Egypt captain Ahmed Hossam (Mido), Helm handled 5,511 conversations between 23 July and 22 August 2026 and passed 245 to the team because they needed a person.

Five thousand conversations some months. In thirty days my team sent ninety-nine messages, and Helm handled the rest. I only see the ones that actually need us.

Ahmed Hossam (Mido), Founder, The Maker. The full figures are in The Maker case study.

Who decides where the line sits?

You do, per workflow, before the build. Helm's pilot agrees authority in writing: what the agent may do, what needs approval, and what it must hand off. It also names the person responsible for the workflow and for each exception. The output is a Scoping & Deployment Brief. After launch, a named engineer stays on the workflow, with ongoing review.

Is Helm the right choice if every reply needs approval?

Probably not. If a person must read and approve every single reply before it goes, the agent is not taking work off that person. A support desk with an AI layer that drafts or suggests replies fits that better. Helm fits when most of the work can be agreed in advance and the exceptions can be named. The comparison of support desks, builders, agencies and Helm sets the options side by side, and the FAQ covers the control questions in full.