- The model is the last layer: Fable 5, GPT-5.6, the next one — you swap them in an afternoon. Value is decided earlier, at the table with the domain experts.
- The gold is in the processes the experts never mention: solidified years ago, safe, beyond discussion. Feature requests are a symptom, not the project.
- The impossible question — “what would change your life, but we haven’t even mentioned because it’s impossible anyway?” — bypasses the mental model, with trust already in the room.
- Build the first project on the answer: buy-in for free and ROI in the “impossible before, feasible now” category.
This week Fable 5[1] and GPT-5.6[2] came out, and the question I get most often is “which one should I use?”.
It’s the wrong question, or at least the last one to ask. Let me tell you why the biggest value of an AI project stays invisible precisely to the people who know the work best, and the single question I ask at every first table to surface it.
Why experts can’t see the biggest value
My job is putting AI agents into production. The technology side — agents, MCP, integrations, long-horizon agentic flows — I know well enough that I can afford not to think about it during discovery. The real focus is something else: understanding how the domain experts work and how their workflows should be redesigned to be AI-centric or human-AI hybrid.
Here I always hit the same obstacle, and it’s human, not technical. Domain experts don’t naturally shift their mental model from traditional to AI-first. They’re excellent at their craft precisely because their way of working solidified over years of practice. And that solidification is a synonym for quality: a true expert knows what they’re doing, how, and knows how to transfer their knowledge to others.
The breaking point is speed. Today those years of expertise have to be communicated in days, to iterate with the innovation team and give fast feedback. You need the same quality standard and the same ownership, achieved in a new way. Experts are very good at this. They just don’t know yet that they can do it.
The practical result at the first table is always the same: they ask for features they think they want (almost never the core), they ask to automate what they can see (the process they already master), and they never mention the solidified processes, the ones that became “the standard”. That’s where the gold is.
Solidification is a synonym for quality. But it’s also why the biggest value stays invisible.
The gold is in what they don’t tell you
The parts of a workflow most in need of AI are almost always the ones the experts don’t even realize they run sub-optimally. They solidified years ago. They’re repetitive, safe, beyond discussion. Nobody puts them back on the table because in the old system they were the best possible. It’s the same mechanism as tacit knowledge: whoever is good at a process has automated it inside themselves, and no longer sees it.
Almost never the core of the value: they’re the old system’s answer to the new question.
The process they already master, the one they know how to tell.
Repetitive, safe, beyond discussion: in the old system they were the best possible.
Nobody puts them back on the table, so nobody brings them to the table.
That’s where the gold is.
And when you ask “what can we do with AI?”, experts answer with what they imagine AI can do. The truth, every single time: what AI can do is far more than they imagine. AI has already moved faster than most domain experts, and the gap widens with every release.
From here, a precise responsibility for whoever leads innovation: push the quality bar always much further than where the domain experts set it at the start. If the project stops where the expert believes the limit is, you’ve automated the visible and left the gold underground.
What AI can do is far more than experts imagine. And the gap widens with every release.
The impossible question
At every first table I always ask the same question. It works for three precise reasons.
“What’s the thing that would change your life, but we haven’t even mentioned because it’s impossible anyway?”
Normal questions (“what do we automate?”) get answers from the old system. This one explicitly asks to step outside it.
The expert has self-censored that need for years, because in their system it really was impossible. The question legitimizes the desire.
The answer almost always comes fast and with energy. If someone answers right away, they’ve been thinking about it for years.
The answer is the natural candidate for the first project: it almost always lives in the “impossible before, feasible now” category.
Then you build on it. The answer to the impossible question is the natural candidate for the first project: it has high perceived value for the expert themselves, so you get the buy-in for free. And it almost always lives in the “impossible before, feasible now” category, where the ROI changes by an order of magnitude.
The question only works if there’s trust
The impossible question only works if the person answering trusts you. Communication is almost always the most critical thing in AI adoption, more than any technology choice.
The fears I meet are always the same two: “AI will replace me” or “I’ll work more, with tighter deadlines”. A project that feeds these fears is the wrong way to do AI: it opens a token-shaped hole in the budget, makes people more distant and less productive, and misses the biggest point. How to build adoption that doesn’t feed those fears is what we wrote in how to get your team to adopt AI.
The biggest point: AI is a multiplier. It opens up what used to be impossible. Using it only to shave a few points of margin is the fastest way to have your margins eaten by costs. If instead you aim at the “impossible before, easy now” category, the bottleneck moves: it becomes go-to-market, and production no longer is.
No LLM handed us this knowledge. We learned it in the field, with real people and real projects, making mistakes and building on top of them. The feedback that confirmed the route: “with this AI-native workflow we comfortably handle deadlines that used to be structurally impossible, or that took hours of overtime”.
AI is a multiplier. Using it only to shave a few points of margin is the fastest way to have your margins eaten by costs.
What to take away
- 01The model is the last layer. Fable 5, GPT-5.6, the next one: you swap them in an afternoon.
- 02The gold is in the processes the experts never mention. Feature requests are a symptom, the starting point for understanding.
- 03Ask the impossible question, at the first table, with trust already in the room.
- 04Build the first project on the answer: buy-in for free, ROI in a different category.
How you turn that answer into a deliverable in production, and who signs off on quality, is the next piece in this series. In the meantime: ask your team the impossible question this week, and see what comes out.
Bring the impossible question to your first table.
We start from your process, not the model: discovery with the domain experts, the first project built on the answer that’s worth the most, and an agent that reaches production. Fixed price and timeline, the code is yours.
The impossible question, in practice
Why can’t domain experts see where AI brings the most value?
Because their way of working solidified over years of practice, and that solidification is a synonym for quality. The parts of a workflow most in need of AI are almost always processes that solidified years ago: repetitive, safe, beyond discussion. Nobody puts them back on the table because in the old system they were the best possible, so they never even get mentioned.
What is the “impossible question” to ask your team?
“What’s the thing that would change your life, but we haven’t even mentioned because it’s impossible anyway?”. Ask it at the first table of an AI project, with trust already in the room: it’s the most direct way to surface the value that normal questions never reach.
Why does the impossible question work?
For three precise reasons. It bypasses the mental model: normal questions get answers from the old system, this one explicitly asks to step outside it. It grants permission to say the unsayable: the expert has self-censored that need for years. And it captures the desperate need: if someone answers immediately and with energy, they’ve been thinking about it for years.
Fable 5 or GPT-5.6: which is better for an enterprise AI project?
That’s the wrong question, or at least the last one to ask: the model is the last layer and you can swap it in an afternoon. What decides whether an AI project produces value or burns budget happens earlier, in the discovery with the domain experts: understanding how they work and how their workflows should be redesigned to be AI-centric or human-AI hybrid.
How do you choose the first AI project in a company?
The natural candidate is the answer to the impossible question: it has high perceived value for the expert themselves, so you get the buy-in for free, and it almost always lives in the “impossible before, feasible now” category, where the ROI changes by an order of magnitude compared with automating the visible.
Does AI replace domain experts?
No: AI is a multiplier, it opens up what used to be impossible. The typical fears (“AI will replace me”, “I’ll work more, with tighter deadlines”) come from badly framed projects. A project that feeds those fears opens a token-shaped hole in the budget and makes people more distant and less productive: it’s the wrong way to do AI.
Sources
- [1]Anthropic, Claude Fable 5 and Mythos 5 announcement, July 2026. www.anthropic.com
- [2]OpenAI, “Previewing GPT-5.6 Sol: a next-generation model”, July 2026. openai.com
This is a point of view by Simone Bova, AI Engineer and co-founder of Yempik, with editing done with Claude. The cases and feedback quoted come from his work in the field; where we cite an announcement, you’ll find the source below.