Human in the loop. On the loop. Out of the loop.

Earlier this month I was in Dublin at IAPP AI Governance Global Europe, the gathering where the people writing the next decade of AI governance frameworks argue with each other. While I was there, the ICF and the EMCC were holding their own global annual meetings. Three rooms, the same question. Who stays in charge of an AI that increasingly does not need them.

Both coaching bodies have called from their stages this year for the profession to engage. Neither has set a compliance standard yet. The window in which coaches set their own standard, rather than waiting for one to be set for them, is open right now.

The three positions

These terms come out of AI governance, originally from work on autonomous systems. They have now reached coaching. Vendors use the word supervised for any of the three. Knowing which one you actually bought is the only protection you have when a client, an insurer or a regulator asks.

Human in the loop. A human is a required step. The AI cannot act until a person approves. Nothing reaches the end output without a human hand on it. Slow. High accountability. The simplest position to stand behind, because someone signed.

Human on the loop. The AI runs on its own. A human supervises, watches, and can intervene. The system does not wait for the human. The supervision lives in the architecture and the review loop rather than in every keystroke. Faster. Defensible only if the supervision is real and designed, not theoretical.

Human out of the loop. Full autonomy. No human is positioned to intervene during the operation. Whatever happens, happens. The audit comes afterwards, if anyone does one at all.

Where you sit on this map decides three things at once. Whether the human stays in the loop through architecture, design and ethics rather than only through real time approval. What you can honestly tell your client. And what kind of practice you have.

The human can stay in the loop in all three positions. The mechanism changes. The commitment does not.

Where each one already lives in coaching

You see all three positions in coaching right now, even if no one names them.

Out of the loop is the weekly conversation volume on consumer AI that has no coach, mentor or therapist in it. A coachee opens ChatGPT at 11pm, types out what they cannot bring to the next session, and takes the answer at face value. If that conversation goes somewhere harmful, no one will know until it shows up in the room weeks later. Your clients are in this category whether they tell you or not.

On the loop is where the enterprise platforms are quietly moving. The chatbot speaks to the coachee. The coach in the network watches a dashboard, in theory. The conversations happen without them, and most will never be read. The supervision becomes a claim.

In the loop is what CoachNova does for everything that carries the coach's name, and what most credentialled coaching practices do when they bring AI in carefully. The output reaches the coach before it reaches the client. The coach is the gate.

The difference between in the loop and on the loop is what you have to show, and to whom. Both can be legitimate. The architecture decides which is which. Designed, not claimed.

Where CoachNova sits on the map

People expect a lawyer to say stay in the loop and do not move. That is not what I say to CoachNova, and it is not what I would say to any company building real AI in 2026. A working AI practice lives across the whole map. The discipline is matching the lane to the action, and proving the floor under each one.

The EU AI Act, phased in through 2026 and 2027, treats each lane differently. Article 14, which sets out human oversight requirements, is the part to watch. Which is why this map matters legally as well as ethically.

Across all three lanes, the coach holds the covenant with the client. The mechanism changes. The covenant does not.

In the loop, by default

Everything that goes out in your name. Session notes, prep briefs, drafted client messages. These reach you before they reach anyone else. You read, you edit, you send.

In this lane the AI is doing the work of a thinking partner who has read every session you have ever had with this client. Before a high stakes session, it pulls together what the client has been working on, surfaces the question you asked last that landed hardest, and drafts the brief. You walk in prepared. Afterwards the notes are ready in minutes. The agent of the practice is still you. The AI gives back time, signal, and a sharper version of the question you were already asking.

This lane stretches as far as your reading time stretches. Which is why the next two exist.

On the loop, for Nova for Clients

This is the layer where your client talks to your AI directly between or after sessions. You will not be approving each message in real time. The AI runs. You supervise. Telling you this is still in the loop would be marketing language, and I would not advise CoachNova to use it. The word matters because the architecture has changed.

The supervision is real if, and only if, five things are true. Expect to see all five anywhere a vendor uses the word supervised.

1. Your methodology, voice and guardrails go in first. Yours, locked before any client gets access. You are not approving each message now. You are approving the rules every message has to live inside.
2. One client, one context. Always. Your work with one client is not in the same context as your work with another. The retrieval boundary is the client. Drift across clients is closed off at the architecture level, not by promise.
3. Built on your sessions, not on the internet. Tightly scoped to the context the two of you have built together. This is what we mean by the coach's AI rather than a coaching chatbot.
4. You see what is happening, in your own time. Summaries of what your client is working through, the themes as they emerge, and a flag the moment something needs you. Not a transcript of every message, because your client is entitled to think out loud. Enough signal that you are genuinely supervising, with a one tap path to step in, change the instruction, and have it take effect within a day.
5. Crisis paths designed in, with you at the other end. Signals of self harm or harm to others trigger a response to the client immediately and an escalation to you and to us immediately. You are the human at the other end of that escalation, not a checkbox in a log. The Canadian case where a major AI provider was held responsible for a chatbot that noticed signs of harm and did not escalate is the cautionary tale every privacy lawyer working in this space carries in their head. We do not want to be that story.

On the loop, engineered. Not on the loop, theoretical. In this lane the AI is working on your behalf when you cannot be in the room. Same thinking partner, different chair.

Out of the loop, only by your client's choice

Your client can decide to make their conversations with your AI private from you. We honour that choice, because it belongs to them. When they make it, you no longer see the content of those conversations. What you still see is the signal minimum: how often the AI is being used, what topics are coming up at the category level, and any crisis flag.

The five items above do not switch off in privacy mode. The retrieval boundary still holds. The scoping still holds. The crisis paths still fire to you and to us.

We do not put coaches out of the loop. Clients can put themselves there, with the floor still holding.

Vendors who quietly move you out of the loop to protect their margin are a different category, and you should be able to tell them apart from us by reading the architecture rather than the marketing copy.

What you can say to your client

A coach asked me how to talk about this with a client. The version that lands is short.

"I use AI in two ways. For the work between us, I read and approve everything before it reaches you. For the moments between our sessions, when you want to think out loud and I am not available, you can talk to an AI built on my methodology, in my voice, shaped by the work we have done together. I see the themes and I am told the moment something needs me, unless you ask me not to be. You can change anything, at any time. If you would rather not use it, you do not have to."

Short, precise, and more useful than saying you use AI for notes. Your client will hear the difference.

What to ask any AI tool in your practice

1. Action by action, where am I on this map. In, on, or out? Make the vendor say one of the three words for each action their product takes. If they reach for a softer word, that is information.
2. Where I am on the loop or out of it, who chose. Me, my client, or you? Vendor chosen out of loop is a different risk from client chosen out of loop. If the vendor cannot tell you the difference, walk.
3. Show me the architecture that proves the floor holds when I am not in the loop. Scoping, crisis paths, signal minimum, intervention. The five items above are a fair starting list. If you cannot get a clear answer to any of them, the supervision is being claimed, not delivered.
4. What happens if I disagree with something the AI did in my voice? If the answer involves opening a ticket, you are out of the loop without having consented to it.

Closing

There is one question that comes up in every conversation I have with a coach about this. Will I still be the coach if my AI is in the conversation when I am not?

The answer is yes. The lane the AI is sitting in matters less than who put it there, and what is holding the floor underneath.

Out of the loop without your client's consent, without crisis paths, without a supervision floor anyone can read, is where the covenant breaks. Out of the loop because your client asked for privacy, with the architecture still holding, is where the covenant holds. The coach holding the covenant is still you. The map is bigger than one lane.