The Agentic Revenue CompanyRevenue as a Service
Blog

Human-in-the-loop in sales: where a person approves and where not

Human-in-the-loop in sales, in practice: which steps an AI agent carries out itself, which it proposes and where a person has to approve.

Overnight, an agent has prepared a long list: dormant customers due for reactivation. In the morning, your sales manager writes in the chat: “Create the cases.” Sounds clear. But for which customers exactly? For all active contacts, or only for those where the use case is enabled? And do emails go out straight afterwards? Moments like this decide whether human-in-the-loop in sales works – or is just a label.

Why the approval question decides success

Agents that are allowed to do too little bring no benefit: every step waits for a click, and in the end the person does the work twice. Agents that are allowed to do too much put customer relationships at risk that have grown over years. Either can make AI projects fail. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 – due to escalating costs, unclear business value or inadequate risk controls.

What an AI agent fundamentally is and why autonomy and human-in-the-loop belong together, we described in Agents: AI agents as the new workforce in sales. This article is about the practice of agentic sales – AI agents in sales: how do you decide, for a specific step, who carries it out?

Five questions for every step

For every type of case in sales, we define an action type together with your functional leads: prepare, recommend, execute with approval, or act autonomously within defined limits. The decision follows five criteria:

  • Impact: What happens if the step is wrong? An internal note weighs less than an email to a customer.
  • Reversibility: Can the step be undone? A status can be reset; a sent email cannot.
  • Uncertainty: How clear is the situation? A clear rule carries further than an assumption.
  • Verifiability: Can you trace why the agent acted the way it did?
  • Accountability: Who answers for the result if it goes wrong?

What this looks like in everyday sales

These criteria produce a clear picture. In our system, the action types are distributed like this, for example:

Prepare: Before a call, an agent summarises the situation for the customer and contacts at the push of a button. It drafts individual replies to customer emails. The decision about what happens with them stays with a person.

Recommend: The agent proposes the next step from the playbook, the next best action. It recommends new contacts, which a person then decides on in a review task. It drafts a rule set for customer classes that only applies once a person has approved it.

Execute with approval: Individual emails to customers only go out once a person has approved them. Bulk tasks – such as creating many cases at once – are only carried out from a saved preview, after an explicit human response, and can be cancelled.

Act autonomously within defined limits: The editorially approved first outreach runs without individual approval once rules, channel and address have been checked. If a customer responds, the system withdraws the planned call task itself. If an email clearly bounces as undeliverable, it stops sending to that address and plans a follow-up by phone.

One asymmetry stands out: what stops outreach may run on its own more readily than what triggers outreach. A stopped send can be made up for; an unsuitable email cannot.

External impact and bulk tasks: always a person

Two kinds of steps always need a human decision with us: steps with external impact beyond approved standard paths, and bulk tasks.

For external impact, the reason is obvious. An email speaks on behalf of your company. It cannot be recalled, and a person answers for its content.

For bulk tasks, the reason is less obvious but just as important. A single mistake affects one customer; a mistake in a bulk task affects a great many at once. Back to the opening scene: “Create the cases” leaves open for whom. That is why our rule is: the target group of a bulk task is never silently chosen by the model. The instruction has to name it explicitly, and the preview shows what will be created before anything is created.

Approvals as a rule in the system

An approval that only exists in the prompt is a request to the model. We design it differently: the approval requirement is enforced on the server side. The agent cannot carry out a step that needs approval without it, however convincingly it phrases things. This is a design principle of Governed by Design, not a security certificate. But it ensures that the limits you set apply in the system, not just in an instruction. More on this under scopes of action, approvals and states.

The value shows particularly clearly in customer reactivation: many contacts, long-standing relationships and a first outreach that has to be right. What that first outreach looks like is described in the article on the reactivation email.

Why this kind of control in operation matters more than the next model is explained in The real AI shift is not new models – it is control.

Your next step

Take a sales process you know well and go through it step by step: what could an agent prepare there, what could it recommend, what could it do with approval and what on its own? Where you are unsure is where the conversation about your mandate begins. If you want to have it with us, check your REVAGO potential.


← All articles