The line you cannot see until you cross it

There is a moment I keep noticing in conversations with communication teams. Someone describes handing a task to an AI tool, a first draft of an internal announcement, a summary of an engagement survey, a translation of a message from leadership, and then they stop for a second. Sometimes the pause is relief, because what came back was better than they expected. Sometimes it is a small, uneasy laugh, because what came back was fluent, confident and wrong, and they only caught it because they happened to know the subject well enough to see the error. What stays with me is not the success or the failure. It is that almost no one in the room could have told you beforehand which of the two it would be.

Ethan Mollick, in Co-Intelligence, gives this experience a name. He describes what he calls the jagged frontier: the uneven line between the things AI does astonishingly well and the things it gets quietly, plausibly wrong. But as I sat with that idea, thinking about what it meant inside organisations, the metaphor stopped being enough for me. In the margin of my copy I wrote something different. There is an invisible wall between what the technology can do brilliantly and what turns into an aberration the moment we leave it alone. The trouble, and perhaps the most important part of the story, is that the wall stays invisible until you walk into it.

For someone working alone, this is mostly a question of efficiency. You try things, you learn where the tool is strong, you adjust. For those of us whose work is a function rather than a task, whose responsibility is the voice, the sense and the trust of an organisation, the invisible wall becomes something else entirely. It is no longer a question of productivity. It is a question of responsibility. Of understanding what we can reasonably let go of, and what still demands human interpretation when the cost of being confidently wrong is not a weak paragraph but a small fracture in how people understand the organisation they belong to.

What makes this genuinely difficult is that the wall never stands in quite the same place twice. It does not run through every organisation at the same point, because organisations themselves are not interchangeable. An AI tool might draft a policy update for one company in a tone that lands perfectly, and produce something for yours that is technically correct yet culturally tone-deaf because it has no access to the history your people carry. The restructuring two years ago. The promise that was made and not kept. The word that means one specific thing inside these walls and something entirely different everywhere else. Context changes where the wall appears. Which means no one can hand you the map. There is no definitive list of what to delegate and what to keep, because the answer is never simply technological. It is organisational.

We spend a great deal of time asking what AI can and cannot do, as though the answer existed independently of the people using it. But capability is only half of the equation. The other half is context. Every week there seems to be another article listing the tasks AI can now perform, yet those lists begin ageing almost as soon as they are published. Not because the technology changes overnight, but because organisations do. A task is never just a task. It carries history, relationships, promises, language and trust. The same activity can sit safely on one side of the wall in one organisation and on the wrong side of it in another. The question was never simply what AI can do. It has always been what AI can do here.

So the real question is not whether to use these tools. It is where, in your own work, the wall stands today, knowing that it will probably stand somewhere slightly different tomorrow. And that is something no framework, no maturity model and no consultant can determine for you. It can only be discovered by testing.

This is uncomfortable for a profession that has often preferred to look composed. It asks us to experiment in the open, to try the tool on real work, to look carefully at what comes back, and to admit when we have just walked straight into the wall. Over time, those experiments produce something far more valuable than a set of rules. They develop judgement, not in the abstract, but about your own organisation, your own culture and your own context. That kind of judgement cannot be downloaded, standardised or borrowed from another company. It is built the slow way, through repeated encounters with the line that only becomes visible after you have crossed it.

The more I think about it, the less interested I become in the list of tasks AI can perform. What remains remarkably stable is the work that sits on the human side of the wall. Knowing which message this culture can absorb right now and which it cannot. Holding the relationships that tell you what is really happening beneath a survey result. Recognising when a technically correct message will land as organisationally wrong because the context has shifted in ways no model could possibly see. Carrying responsibility for what gets said, and for the trust that is either spent or strengthened each time we say it. The tool can produce language. It cannot hold the context in which that language will be interpreted. And it cannot answer for what happens once it is.

This is also where the invisible wall stops being an individual discovery and becomes a cultural one. If the only way to find it is to experiment, then an organisation that punishes experimentation will never learn where it actually is. Its people will either avoid the tools altogether and quietly fall behind, or they will use them in secret, without the conversations that allow judgement to develop. Creating permission to test, to fail on something low stakes, to say out loud that a particular attempt sounded convincing but missed the mark, is not an IT decision. Nor is it simply another governance policy. It is culture in practice. And helping organisations create that culture has always been one of Internal Communication’s most important responsibilities, even if we have not always described it that way.

Perhaps that is why this idea has stayed with me longer than many of the others in Co-Intelligence. Read through the lens of organisational life rather than individual productivity, it tells a very different story. It does not tell me that communication is becoming less human. It tells me that the human part of communication is becoming easier to see. The more fluent these tools become at producing language, the more valuable become the capabilities that language alone can never replace: reading context, recognising what people carry with them, distinguishing between what sounds plausible and what is actually true, and accepting responsibility for the consequences of saying one thing rather than another.

We will never be handed the map. Every organisation has to draw its own, one experiment at a time, paying closest attention precisely when the technology sounds most convincing. Because the invisible wall is not where AI stops being useful. It is where our responsibility begins.