Most AI demos end at the moment the model writes something. Real products start there. The question that decides whether a business can trust an AI system is simple: what happens between the draft and the send?
In three products we built this year, the answer is the same. The AI does the legwork. A person makes the call.
The pattern: draft, review, send
Human-in-the-loop means the model produces a draft and a person approves, edits or rejects it before it reaches a customer, a feed or an inbox. It sounds slower than full automation. In practice it is the version that gets switched on and stays on, because the team can see what the AI is doing and stop it when it is wrong.
The pattern has four parts:
- Context: the AI gets the facts it needs, not only the message in front of it.
- Draft: the model writes a first version in the right voice.
- Decision: a person approves, edits or rejects it, in one place.
- Record: every decision is logged, so the team can see what was sent and why.
Customer support: CRAS
CRAS is the approval dashboard for an AI customer-support agent. Every inbound email is sorted into a green, amber or red tier with a confidence score. For each ticket the agent pulls the customer's order from Shopify, including payment status, fulfilment and carrier, so the draft answers the real question, such as where a parcel is, instead of guessing.
A team member then approves, edits or rejects the draft. Nothing goes out on its own, and every decision lands in an audit trail.
The most important rule is what the AI is not allowed to do. Red-tier tickets, such as complaints, refunds and anything involving money, never get an AI draft. They go straight to a person. Deciding where the AI stops is as much a part of the design as the prompt.
Social content: Operation Sparta
Operation Sparta turns product data into finished Instagram posts. A brand project holds the product name, colour, voice, description and image. One click in the Content Studio generates the post image, a caption and hashtags in that voice.
The output does not go live until someone approves it. Approved posts land in a weekly planner or go straight to Instagram, and the app tracks each post's status and engagement afterwards. Generation runs through n8n workflows on the brand's own OpenAI or Gemini key, so the brand keeps control of its models and its costs.
Sales outreach: CVR-MATE
CVR-MATE finds new leads in the Danish Central Business Register every morning. You describe your ideal customer in plain language, save it as a trigger, and review fresh matches each day with the reason each company fits.
For each company, Claude writes a briefing and a ready-to-send outreach draft. Ready to send is not the same as sent: the user saves the lead or passes, and decides what goes out. The deal then moves through a built-in pipeline from quote to order.
What the approval step needs to work
An approval step only helps if it is fast. If reviewing a draft takes as long as writing it, people stop reviewing. Across these products, a few rules held:
- Show the evidence next to the draft. In CRAS the Shopify order sits beside the reply, so checking it takes seconds.
- Make the decision one action: approve, edit or reject.
- Sort before you draft. Tiers and confidence scores tell the reviewer where to look first.
- Keep people on the high-stakes cases by rule, not by hoping the model notices.
- Measure it. CRAS tracks approval outcomes and the median time to a decision, which shows whether the AI is saving time.
When full automation makes sense
Some tasks can run without review: tagging, routing, summaries for internal use. The line we draw is simple. If a mistake reaches a customer, costs money or speaks for the brand in public, a person approves it. As the approval history builds up, you can see which categories are always approved unchanged and decide, with data, whether to automate them.
Building an AI agent?
We design and build AI integrations and chatbots and automation with the approval step built in from day one. Tell us what you want to automate.