You’ve seen the all-hands. Leadership clicks through a few slick demos. Someone announces a company-wide ChatGPT license. A chart goes up: “73% of employees now use AI weekly”. The room nods. We’re an AI company now.
And then everyone goes back to work — the same work, routed the same way, through the same approvals, the same handoffs, the same bottlenecks. Nothing about how the company actually delivers has changed. A few people draft emails faster.
That gap — between the announcement and the operating model — is the most expensive illusion in business right now. Because “we use AI” has become a vanity metric. It measures how many people opened a tool. It tells you nothing about whether the company is built differently from a year ago.
The right question isn’t whether your company uses AI. It isn’t even how “actively” it uses it. The question is:
How deeply has your company rebuilt the way it works around AI?
Depth of operating-model change — not tool count, not usage stats — is the line between hype and transformation. And once you start measuring depth, an uncomfortable thing happens: almost every company turns out to be lower on the ladder than it thought.
We’ve been building these systems since 2017. The same pattern shows up everywhere, so we turned it into a framework — four questions and a six-stage ladder. This is that framework, and an honest account of how to use it, including where we ourselves land on it.
A company that has genuinely rebuilt around AI is not an old company with AI tools added on top. It sees its own data differently. It makes decisions differently. It distributes responsibility differently. It turns knowledge into action differently. The tools are the least interesting part.
This is why the “which subscription should we buy” conversation is mostly a distraction. The operating model change is the same whether it runs on Claude, Gemini, OpenAI, or an open-source stack underneath. The substrate you build — the way your work is described, connected, and made actionable — is what you own and what compounds. The model is a swappable engine. The transformation is everything around it.
And here’s the part most leaders underestimate: this kind of change doesn’t happen by license rollout. It requires people and teams to actually learn how to work in a new way. Skip that, and you’ve paid for capability you can’t capture.
Look at the picture above, which greatly represents the trend we have right now. The AI is not just some technology that only coders should use; it’s a ground-breaking change that we should master.
For two years, the move was to bolt AI on. Buy the licenses, run the pilots, put out the press release. That wave has crested — almost everyone has done it, which means it no longer distinguishes anyone. What’s opening up now, in 2026, is a gap. On one side, the companies that demoed AI and stopped. On the other side, the companies that used the last two years to rebuild — to make their work legible to machines, to let AI take real action, to push the ability to build solutions out to ordinary employees. The first group is hitting a ceiling. The second is compounding. The distance between them is widening fast, and it is much harder to close than it looks, because it isn’t a purchasing decision. It’s an operating-model decision.
The ladder below is a way to figure out, honestly, which side you’re on.
Before the stages, the lens. These four questions are the whole diagnostic. You apply them to any unit — the whole company, a single function, a single department, or a single process — and together they measure depth rather than surface activity.
-
What can AI see? Is the work described, structured, machine-readable — or does it live in people’s heads, scattered across chats and tools nobody connected? (Substrate)
-
What can AI do? Can it only help write, search, and summarize — or can it take real action: trigger a process, update a record, finish a task? (Agency)
-
Who can build with it? Only specialists and a few enthusiasts — or can ordinary employees assemble useful solutions themselves? (Ownership)
-
Has the organization itself changed? Or did the old structure get a new tool layered on top? (Structure)
Most “AI strategies” only ever touch on a subset of them, not all 4. That’s something to think about.
The bottleneck rule: AI can’t do more than it can see
Here is the single most useful thing in this entire framework, and the line worth being present on your strategy deck:
AI can rarely(question 2) more than it cando(question 1).see The substrate — what AI can see — is almost always the bottleneck. When people say “AI isn’t good enough for our business,” what they almost always mean, once you dig in, is ”AI can’t see enough of our business”. The work isn’t described. The data isn’t connected. The context lives in someone’s head.
This matters because it changes where you spend your money. Reaching for a stronger model buys you a temporary bump. Structuring what AI can see is cheaper, and it sticks — swap the model later, and the gain survives, because you fixed the foundation, not the engine. Fix what AI sees before you expect it to act.
With the lens in hand, here are the six stages a company moves through. Each one has a characteristic look, a trap that keeps companies stuck there, and the move that climbs out.
Lots of talk about AI; almost nothing changes in the real work. Data isn’t organized, processes aren’t described, and key information is fragmented. At best, AI summarizes a meeting or drafts some text — after a human has set everything up. Leadership can talk about AI, but can’t point to a single process it changed.
The trap: mistaking visible enthusiasm — demos, announcements, “we’re an AI company now” — for change in how you deliver.
** Measure:** the number of recurring processes AI can complete end-to-end
The move: put capable AI into people’s hands, remove access and billing friction, and describe at least one real workflow so AI finally has something concrete to see.
Now there’s real, tangible benefit — but only as individual productivity. People write, analyze, and produce faster. Each person has their own tricks and prompts. The gains are real, and they belong entirely to individuals, not to the company as a system. When a strong employee leaves, their way of working leaves with them.
The trap: vanity metrics. “Most of our employees use AI weekly” measures tool adoption, not organizational transformation.
Measure: per-person task cycle time and rework rate — not “what share of people use the tool”. Speed of the work, not adoption of the app.
The move: capture those individual tricks into shared, team-owned playbooks and conventions — so the method survives any one person, and the goal shifts from “people use AI” to “a team’s work is measurably faster”.
AI embeds into team workflows, not just individuals — but each function evolves on its own. Sales has its AI scenarios, support has theirs, and each team grows its own shared playbooks. There are real, team-level productivity gains. But there’s no shared environment across functions: the company is efficient in parts, not holistically new.
The trap: local optimization. Each function improves independently, and the gains don’t compound because nothing connects them.
Measure: team** **cycle time, time-to-productive for a new hire, and the share of work that runs on shared playbooks rather than personal habits.
The move: connect function-local data and systems into a shared substrate — expose each as a programmatic interface rather than copying it around — and start giving AI the cross-functional picture, with non-specialists beginning to co-author.
This is where change becomes structural rather than cosmetic. AI gets the whole picture — data across functions, key systems, decision history, institutional knowledge — and it starts to act, not just suggest: updating records, launching tasks, gathering information across sources, preparing cross-functional decisions. Crucially, non-technical people — in sales, service, finance, product — shift from being users to being co-authors of internal solutions.
The trap: technological theater. It can look advanced while the foundation stays weak — ordinary employees still can’t shape solutions, and AI still only suggests. The tell: if non-technical people aren’t becoming co-authors, the foundation isn’t there yet.
**Measure: **the number of solutions built by non-specialists, cross-functional lead time, and the share of tasks where AI acts rather than only suggests.
The move: build the substrate that makes autonomy trustworthy — shared memory, evaluation, the ability to observe what AI does — and set clear bounds so core cycles can act safely while humans move toward direction and exceptions.
Core work cycles run themselves within allowed bounds, and people focus on direction and exceptions. The system notices changes, gathers context, acts within its bounds, and escalates only where genuinely necessary — driving bounded work to completion without being told each step along the way. Management itself changes: less manual routing, fewer intermediaries, humans supervising and handling the edge cases. For ongoing production work, value usually peaks here.
The trap: automation chaos — over-trusted autonomy. A pile of autonomous actions is volume, not correctness. ”It ran” is not “it ran right”. Grant autonomy only where you can both see it and stop it.
Measure: the share of bounded work completed autonomously, the escalation rate, incidents, blast radius, and cost per completed unit of work.
The move: close the loop so the system improves over time, and track quality against a captured baseline so any improvement is real and measurable — not just motion.
The frontier on top of a working autonomous system: it improves itself. It learns from past cycles, reuses prior solutions, cuts duplication, and refines its own way of working — getting measurably better over repeated cycles, and proposing its own improvements for humans to approve. This is rare, and it is not where most companies should aim.
The trap: mistaking a stronger base model — or polished automation — for genuine self-improvement; and self-degradation, when a system edits itself without anything checking whether it’s getting better or worse.
Measure: the quality trend over many cycles against a fixed baseline — is it rising? Plus a falling defect/rework rate and falling cost per unit.
The move: only attempt this on a foundation that already works, with hard gates: nothing self-proposed ships without measurement, and without a human approving the change.
Here’s the quick version you can run in your head right now:
Answer the four questions for your company. What can AI see? What can it do? Who can build with it? Has the org itself changed? Take your lowest score across the four — that’s your real stage, not your best one.
Running a bit further, your real stage is the lowest of three things, not just your weakest answer. First, the type of work sets a ceiling: one-off research and pilots top out around Stage 3 no matter how slick the tooling — only repeatable production work can justify Stages 4 to 6. Second, your weakest of the four questions. And third, the highest stage whose guardrails you can actually pass— if you can’t yet see and stop what AI does, you don’t get to count yourself at the stage that assumes you can (each stage requires additional security/observability guidelines, without which you don’t have the cards to conrol the agent). Take the lowest of the three. That’s your floor, and your floor is your stage.
Expect a mixed profile. Most organizations straddle stages — strong on one question, weak on another — and that’s normal. Score each question independently, then anchor on the weakest link, because that’s what’s actually holding the system back. The bottleneck rule predicts what you’ll usually find: the first question, what AI can see, is the one dragging the rest down.
This is the part that separates a useful framework from a maturity-model treadmill.
You do not need to reach Stage 6. From Stage 3 onward, any stage can be a perfectly good place to stop. The work you do sets a natural ceiling — some processes top out at Stage 3, and that’s correct; pushing them to Stage 5 would cost far more than it returns. Climb only where the value justifies the standing cost of staying there.
There’s exactly one rule that isn’t optional: Stages 1 and 2 are never the place to stop. At those stages, the gains belong to individuals, not the company, and they walk out the door when people do. Everything above that is a genuine strategic choice. The goal isn’t to be the highest. It’s to be deliberately placed.
A few lessons that don’t fit neatly on the ladder but shape everything on it:
It would be easy to end a piece like this from the summit. We won’t, because the most honest thing we can tell you is also the most useful.
Mindcraft sits at around Stage 3, building toward Stage 4**.** Shared, team-level AI workflows are in place; we’re now connecting them into a unified, cross-functional setup. The Stage 3 → 4 work is active, not finished. We currently sit earlier on this ladder than the setups we build for our clients — a real gap, and one we’re closing as we climb our own ladder in parallel.
So: how deep does your AI actually go?
Of course, many technical aspects remain behind the scenes. Things like actually implementing these 6 stages, the many trade-offs, and the why & how questions we answered when applying these stages to real-world companies. I may post a more detailed article on this topic in the future.
In addition, I want to thank my colleagues: Alex Simkiv, Andy Bosyi, and Nazar Savchenko for productive conversations, collaboration, and valuable advice, as well as the entire MindCraft.ai team for their constant support during the development of this framework.
How Deep Does Your AI Transformation Actually Go? was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.