Two stories landed on the same weekend, and they point at the same question.

On November 2, Abu Dhabi's state oil company ADNOC signed a multi-year deal with OpenAI and G42 to bring AI agents into its operations. It says it will be the first energy company to use OpenAI Frontier, a platform launched in February for building and managing business agents.

A day earlier, Axios reported that OpenAI plans to push for mandatory safeguards on the most powerful AI systems after the US midterms. OpenAI calls the idea "speed limits," and describes it as a way to let humans keep their hands on the wheel.

Read those together and the direction is clear. Agents are moving into real work, and the word everyone reaches for is control.

You do not need a frontier lab to take that lesson seriously. If you build chatbots, n8n workflows or any automation that can do things, you already face the same problem on a smaller scale: what happens when the AI is confidently wrong, and nobody is watching?

This post is a practical answer. It covers when to put a human in the loop, five patterns that work, a small code sketch, and the mistakes that make approval steps useless.

Quick answer: Add a human approval step wherever an AI action is hard to undo, costs money, or reaches a customer. Let everything else run automatically, but log it. Make "no answer" mean "do nothing," never "go ahead."

What the news actually says (and what it does not)

Let us keep the facts straight, because headlines on this topic travel fast.

Axios reported that OpenAI's global affairs team plans to unveil new ideas for mandatory safeguards on the most capable AI systems, timed to the lame-duck session of Congress that begins on November 9. As far as I could find today, the full proposal has not been published, so the details are still unknown.

People read this in different ways. Supporters see a sensible idea about human oversight. Sceptics see positioning ahead of regulation. You do not have to settle that argument to use the engineering idea underneath it, because a human checkpoint before a risky action is simply good system design.

The ADNOC deal is about scale, not safeguards. But it shows why this matters: once agents run inside large organisations, small mistakes repeat thousands of times.

Not every action needs a human

The biggest mistake is putting approval on everything. People stop reading the requests, and the safeguard dies. Start by sorting your actions by how much damage they can do.

Action typeExampleWhat to do
Read-only or draftSummarise a support ticketRun automatically
Low-risk, easy to undoTag a lead in your CRMRun automatically and log it
Reaches a customerSend a WhatsApp or email replyHuman approval at first, relax only with evidence
Money, legal or permanentRefund, delete data, change a contractAlways require a human

Ask three questions about any action. Can it be undone? Does it cost money? Does an outside person see it? One "yes" is enough to deserve a closer look.

Five patterns that work

1. Draft, don't send

Let the AI prepare the work and stop. The agent writes the reply, a person clicks send. This is the easiest pattern, and it still saves most of the time, because writing is the slow part and checking is fast.

2. Route by rules, not by feelings

Many people ask the model "how confident are you?" and trust the number. That is unreliable, because a model can sound sure and still be wrong.

Use checks you can measure instead. In a RAG system, send the answer to a human when the retrieval score is low or when no source document was found. For structured output, send it to a human when it fails validation. Rules like these are boring, and boring is what you want.

3. Show a preview of what will change

Do not ask "approve this action?" with no detail. Show exactly what will happen: which records, which values, which recipient. A before-and-after view lets the reviewer catch a wrong customer or a wrong amount in seconds.

4. Set hard limits

Some safeguards should not need a person at all. Cap how many records one run can touch, how much it can spend, and which domains or accounts it may contact. An allow-list beats a block-list, because it limits damage even when the AI does something you never imagined.

5. Keep a log and a way back

Every automated action should leave a record: what the AI decided, what it did, and when. If you can undo it, build the undo button before you need it. When something goes wrong, a clear log turns a panic into a ten-minute fix.

A small code sketch

Here is the shape of a simple approval gate. It is a sketch to show the idea, so adapt it to your own tools.

RISKY_ACTIONS = {"refund", "delete_record", "send_contract"}
AMOUNT_LIMIT = 5000  # set your own number

def run_action(action, payload, ask_human, execute, log):
    needs_review = (
        action in RISKY_ACTIONS
        or payload.get("amount", 0) > AMOUNT_LIMIT
    )

    if needs_review:
        decision = ask_human(action, payload, timeout_minutes=30)
        if decision != "approved":      # includes "no answer"
            log(action, payload, status="blocked")
            return "blocked"

    result = execute(action, payload)
    log(action, payload, status="done")
    return result

Notice one line in particular: anything other than a clear "approved" blocks the action. A timeout counts as a no.

How this looks in n8n

If you build with n8n, the idea maps neatly onto workflow steps:

  1. The AI step prepares the draft or decision.
  2. A rule step (an IF node) checks the risk: amount, action type, retrieval score or validation result.
  3. Risky items go to a message in Slack, Telegram or email asking for a decision. n8n has a Wait node that can pause a run until a webhook call resumes it, and some app nodes offer a "send and wait for response" option. Check which one your version supports.
  4. Approved items continue. Rejected or unanswered items stop and are logged.
  5. A final step writes the result to a log sheet or database.

The same structure works for a voice agent, a WhatsApp bot or a RAG assistant. Only the risky actions change.

Five mistakes that make approval useless

Asking for approval on everything. People start clicking "yes" without reading. Keep reviews rare enough that each one gets real attention.

Showing too little. "Approve action 4521?" tells the reviewer nothing. Always show the content and the effect.

Treating silence as a yes. If no one answers, the safe result is nothing happening. Auto-approval after a timeout defeats the purpose.

Never reviewing the reviews. Check your logs now and then. If humans approve 100% of requests, either the rules are too strict or nobody is looking.

Giving the agent more power than the job needs. Approval steps are a second line of defence. The first is giving each automation only the access it truly needs.

A simple checklist before you ship

  • Have I listed every action my automation can take?
  • Is each action marked as safe, logged or approved?
  • Is there a hard cap on spend, volume and recipients?
  • Does "no answer" mean "do nothing"?
  • Can I see a log of what the AI did last week?
  • Can I undo the actions that matter?

If you can answer yes to all six, you are ahead of many production systems.

Frequently Asked Questions

What does human-in-the-loop mean in AI? It means a person reviews or approves an AI decision before it takes effect, at least for the actions that matter most.

Will approval steps slow my automation down? Only for the risky actions. If you sort actions by risk, most run instantly and only a small share waits for a person.

Is a human review step required by law? It depends on your country and your industry. Rules differ and are still changing, so check the rules that apply to you or ask a professional.

Can I replace the human with a second AI? A second model can catch some errors and is useful as an extra filter. For money, legal or permanent actions, a human is still the safer final check.

How do I decide which actions need approval? Use three questions: can it be undone, does it cost money, and does an outside person see it? If any answer is yes, add a check.

Key takeaways

  • Agents are moving into real business work, and control is becoming the central question.
  • Sort actions by risk. Automate the safe ones, log the medium ones, and require a person for the dangerous ones.
  • Use measurable rules for routing, not the model's own confidence.
  • Make silence mean "do nothing," and always keep a log.
  • Rare, well-explained approvals work better than constant ones.

Read Next

Sources: Axios's report on OpenAI's planned policy push (November 1, 2026), as summarised by Implicator and AI Weekly, and The Next Web's report on the ADNOC, OpenAI and G42 agreement. Details may change as the stories develop. This article is general guidance, not legal advice.

✳ ✳ ✳