What does "human in the loop" actually mean?
A human-in-the-loop AI agent does useful work (reading messages, looking things up, drafting replies, routing tickets), but a person makes the decisions that carry risk. The agent speeds up the work around a decision without taking it.
In our shorthand: the agent drafts, routes and escalates. A person decides.
Why are fully autonomous agents risky for customer-facing work?
Many AI products are sold on the promise of working "without human intervention". Large language models are fluent, and they can be confidently wrong. In customer-facing work, the costs are lopsided:
- A wrong answer about a refund, a price or a delivery date can turn into a complaint, a chargeback or a lost customer.
- A wrong answer about health, money or legal matters can cause real harm.
- An agent that takes an action, such as cancelling an order, issuing a discount or updating a record, can create problems that are hard to undo.
Autonomy is fine for low-risk, reversible steps. For the rest you want a rule, not a hope.
What is a stop rule?
A stop rule is a one-page document written before an agent is built. It has four parts:
- Alone: actions the agent may take without approval, for example answering opening hours, sharing an order status or creating a lead record.
- With approval: actions the agent prepares but a person confirms, for example replying to a refund request or sending a quote.
- Hard stops: situations where the agent must stop and hand over immediately, for example a complaint, a legal threat, a medical question, an abusive message, low confidence, or a customer asking for a human.
- Handover package: what the person receives, such as the conversation, the data collected, the agent's suggested next step and the reason it stopped.
How do you decide what goes in each bucket?
Ask three questions about every action:
- Is it reversible? Sharing information is easy to correct; issuing a refund is not.
- What happens if it's wrong? Inconvenience is acceptable; financial, legal or health harm is not.
- Is there a clear rule? If staff would need judgement to decide, the agent should draft, not decide.
Does human-in-the-loop make automation slower?
It moves people to where they add value. Instead of typing every reply, staff review drafts, approve with one click and handle the genuinely hard cases. Most routine questions can sit in the "alone" bucket, and the review queue holds only what needs judgement.
How should you review a stop rule after launch?
- Review escalations weekly for the first month. Were they the right cases?
- Look for answers staff keep editing heavily. The prompt or the source content needs work.
- Move actions between buckets deliberately and in writing, never by quietly loosening the rules.
- Keep logs of what the agent did and why, which also helps with data protection obligations.
How do we apply this in client work?
Everything on this page feeds into how we run pilots. We start with a consultation to understand the process and the people involved, then write a one-page scope: the trigger, inputs, steps, outputs, a success measure and a stop rule that says what the agent may do alone, what needs approval and when it must hand over. We build inside the tools you already use, test on real or realistic cases including awkward ones, and review the results with your team before anyone talks about expanding.
We keep a few habits on every project: we never publish invented results, we label examples as illustrations, we list third-party costs separately, and we say so plainly when something isn't worth automating yet. If that sounds like the kind of partner you want, read the pilot process or message us on WhatsApp.
What should you prepare before talking to us?
You don't need a technical brief. A useful first conversation needs only a few things, and gathering them is valuable even if you never work with us:
- The process in plain words: what triggers it, who does what, and where it usually goes wrong.
- Rough volumes: how many messages, orders, tickets or invoices per week. A one-week tally is enough.
- Your tools: WhatsApp app or API, CRM, sheets, accounting or helpdesk software, including versions or plans where you know them.
- Three real examples: typical messages or cases, plus one awkward one, with personal details removed.
- What must never happen: the mistakes that would hurt most. These become the first lines of your stop rule.
- An owner: the person who will review the workflow once it runs.
Send these on WhatsApp or by email, or run the pilot scorecard first.
How does a pilot with us work?
Every engagement follows the same four steps. Consult: you walk us through the process, the tools and where it breaks. Pick the pilot: we choose one workflow and write down its inputs, outputs, tools, stop rule and success measure. Ship: we build it into the tools you already use and test it on real cases. Review: your team checks what the agent did and what it escalated, then decides whether to tune it, expand it or stop. Read the full pilot process.