There is an irony in the way we learn about AI. We spend a lot of time asking it what it can do. I certainly do. “Can you analyse these survey results? Can you research the top articles related to the future of work? Can you create my outreach dashboard?”
A more useful way to think about it is: Even if AI can do it, should it? Do I want it to? Or are there certain aspects I want to protect?
My research found that leaders getting results from AI were deliberate about both where AI adds value and what they rwanted to preserve and protect.
At an AI governance roundtable that I attended last week there were mentions of “group therapy”. Senior decision makers valued hearing from others that they face similar pressures. That the more time we spend with AI the harder it becomes to strike a balance between innovation and guardrails, and that there’s a major gap between using AI and using AI well. Some worried about cognitive atrophy and losing their distinct advantage. Under all of the discussion there was an important question: what do we feel comfortable delegating to AI, and what do we still want to rely on ourselves to do?
AI forces leaders to define both value and values
Across the successful AI adopters I interview in my research there is a common pattern: they have drawn clear personal lines, whether about data privacy, creative ownership, or where their own judgement remains essential.
Sue Lacey Bryant, ex-Chief Knowledge Officer of the NHS in England, quipped during our interview:
“AI forces you to identify both value and values”.
That really stood out for me because in business we instinctively focus on the first word value, as in ROI, efficiency, outcomes. The second word, values, tends to come later, usually after something goes wrong, or as part of a marketing exercise.
AI is forcing all of us not just to identify where the value of AI will come from, but also what we consider valuable enough to protect as human.
I see this as setting your own “red lines”. Where are the lines you will not cross, whether AI can replace what you do or not.
Another way to think about this is as as setting your own rules of the game. AI is your teammate or “colleague” and, to play well together, you need to have a clear idea of who is meant to do what, what is allowed and what is not.
How do high AI-Q leaders find their own red lines?
My research showed that they approach this in two ways. One group dives straight in and figures out their red lines as they go along, during their experimentation. The other group feels more comfortable to experiment by having those red lines, their non-negotiables, in place to begin with, and then iterating them as their confidence grows.
Either way works. The common denominator is being explicit and being able to articulate what those boundaries are. Ultimately you bear responsibility for the way you use AI, and this clarity is essential as it provides the psychological safety to continue with experimentation, an important component of AI adoption, given that the technological capabilities are constantly evolving.
Where does human control still matter?
Finding your own red lines about what should stay human-only can seem a rather abstract exercise, but one of the most powerful ways to get started is to get personal.
As a writer, if I use AI to generate my ideas I would not feel that I own those ideas – that they are mine. Even if the output is identical, that ownership is an essential part of the process and many writers I speak to feel the same. In this case AI can do the writing, but I feel it should not, because doing so does not align with my values, nor with my identity. On the other hand I am quite comfortable using it to research company examples for me as a starting point, which I then need to verify, as the trade-off I am making there is aligned with my values.
One of the clearest examples I heard on red lines came from a wealth manager whom I met during a friend’s birthday dinner. Her clientele was barristers and she had spent fifteen years building client relationships based on discretion and personal trust. She was using AI agents for financial modelling but was absolutely clear about where the line was: scrubbing client data and the safety of the model she used. She removed names, addresses and other identifying details, before using their broader data on income and expenses to model different scenarios for them. She also chose AI tools with a safety-first approach to their data handling.
A different kind of red line, but equally clear, came from the CEO of a publishing company, who had worked with his marketing team to train an AI agent to write speeches and emails in his voice and tone. He never allowed the agent or the team to send those out without personally editing every single one to ensure the match was perfect. Nothing goes out under his name without his personal sign-off. Authenticity of voice is something he does not compromise on, even when AI can replicate it convincingly.
The red line is different in each case: mine is ownership, the wealth manager’s was trust and privacy and the publishing CEO’s authenticity.
How red lines can enable AI adoption
Your red lines don’t stop you - they give you an edge.
For some the boundaries will emerge from regulatory requirements in your sector. The EU AI Act, for example, places specific obligations around human oversight of AI systems, and several industries are developing their own codes of practice. What feels like a personal boundary today may become a compliance requirement tomorrow, which is another reason to be explicit about where you stand now.
If you lead a team or function, you will also want to think about how to help your teams establish their own red lines, and also what those look like at the organizational level as you govern how AI is used.
In the context of your organization you can use red lines as enablers of AI adoption. In the public sector, reproducibility is often a non-negotiable. In a project that I worked on with the UK Government Office for Science, a key constraint was being able to reproduce the findings with the same methodology. This meant we had to rule out the use of certain generative AI tools for deep research because their probabilistic nature meant that someone else running the same prompt would get different results. In this case AI tools were unsuitable for this particular piece of work. We knew where generative AI was appropriate and where not.
There is no universal answer. What matters is clarity about what cannot be compromised. And importantly, leaders who can clearly articulate their own boundaries give their teams permission to do the same.
Three questions to ask before delegating work to AI
What data are you sharing with AI tools, and would you be comfortable if a client, regulator, or board member knew?
What decisions or outputs are you delegating to AI, and are you still scrutinising the results?
Where would you refuse to use AI even if it could do the job, and why?
If you’d like to share your own red lines and how you identified them as part of your AI adoption journey, I’d be interested to hear what shaped them.
Recently on The Athenian Intelligence:
