Put the question to any executive committee and watch what happens. What does AI-native look like for us, specifically?
The answers arrive in two flavours. Some are technology: agents, copilots, a data platform. Some are aspiration: faster, leaner, data-driven. Neither is a description of a company. Both describe things you would install, or qualities you would like to possess, and a destination is neither.
This is not because the team is unserious. It is because the question cannot be answered from where they are standing. And a great deal of what goes wrong afterwards follows from the fact that nobody says so out loud.
What a plan does when it cannot name its endpoint
The previous generation of transformation was navigable. Cloud migration, ERP, collaboration suites: in every case you could visit a company that had already arrived, see the endstate, and reason backwards to a critical path. The destination existed before you set out. A roadmap was the right instrument, and the discipline of using it well was the whole job.
Set the same instrument against a destination nobody can describe, and the plan does not fail loudly. It substitutes. Unable to specify an outcome, it specifies the things it can name with confidence: environments provisioned, licences deployed, staff trained, pilots launched, a steering committee convened. Every item is real. Every item ships. The plan completes.
Checkboxes are not laziness. They are the residue of a question nobody could answer.
Which produces the outcome now visible across the sector: substantial AI programmes are reported everywhere, and correspondingly few organisations report any movement in earnings because of them. The dashboards are honest. They are simply measuring the parts of the work that could be specified in advance — and those were never the parts that mattered.
What AI-native actually means
Start somewhere unglamorous. Why does a purchase over a certain value need three signatures?
Not because three is a natural number. Because at some point in the company's history, judgment was scarce and expensive. Reviewing a purchase required a person who understood the category, the budget, and the counterparty, and such people were few and busy. Three signatures was a rationing mechanism for a costly resource, dressed as a control.
Now look at the rest of the operating model with the same eye. Every process your company runs encodes an assumption about what used to be expensive.
Here, then, is a definition worth putting in front of a board:
An AI-native organisation is one whose processes have been re-derived from what is now cheap, rather than inherited from what used to be expensive.
It follows that adding AI to an existing process cannot produce one. A process built around scarce judgment does not improve when judgment becomes abundant — it becomes the wrong shape, and the wrong shape executed faster is still the wrong shape, now at a marginally higher unit cost. The approval still takes three signatures and four days. The emails are better written.
Why you cannot know your own answer in advance
Nobody wrote the constraints down. The four-day approval does not carry a comment explaining that it exists because analytical judgment was scarce in 1998. The assumption was true when the process was designed, so it was never stated; and once unstated, it became invisible, and then it became simply the way we do it here.
This is why the destination cannot be specified at the outset. It is not hidden in a strategy the team has failed to write. It is distributed across a thousand undocumented assumptions, and the only reliable way to discover whether one is still load-bearing is to remove it and observe what happens. You find these things by subtraction, not by analysis. No consultant can hand you the answer, because the answer is a property of your own accumulated history — and anyone who arrives with your destination already drawn is selling you someone else's fossil record.
Two further unknowns compound it. You do not know what the technology will do in nine months, and neither does anyone else. And you almost certainly do not know where your real bottlenecks are, because you have been measuring the parts of the business that were easy to measure.
How to move when you cannot see the destination
The response to an unspecifiable endpoint is not to abandon planning. It is to stop planning the destination and start planning the search.
Four disciplines replace the roadmap. None of them is soft.
- A thesis, not a target Name the constraint you believe has lifted, and what it would mean if you were right. We believe underwriting judgment is no longer scarce; if true, the queue disappears and the role changes. A thesis can be wrong, which is what makes it useful. A target of 40% adoption by Q3 cannot be wrong, only unmet.
- Probes aimed at the economics Send the probe where the money is, not where the consent is. The reason flagship pilots cluster in HR and internal support is that those functions have the least to lose by saying yes — which is precisely why nothing found there will move the margin. Go to the process that carries cost, volume, or cycle time, and expect resistance proportional to the value at stake.
- A learning rate faster than the technology's change rate If the models improve materially every quarter and your organisation completes one learning cycle a year, you are permanently reasoning about a world that no longer exists. The cadence of probe, measurement, and re-orientation is not an operational detail. It is the single variable that determines whether any of this compounds.
- Substrate that is true regardless of who wins Some investments pay off under every future: a gateway that keeps the choice of model reversible, ingestion standards that make internal knowledge retrievable, an evaluation suite that tells you when a change made things worse. Frontier models are converging and falling in price; standardising on one vendor's suite buys convenience and little durable advantage, since whatever it does for you it does for your competitor next quarter. What cannot be bought is your organisation's own accumulated judgment. Make certain it accumulates in your layer, not a supplier's.
Deciding at two speeds
Uncertainty is an argument for moving fast, and also an argument for moving slowly, depending entirely on whether the decision can be undone. Most organisations get this exactly backwards: they deliberate for a quarter over which model to use, then reorganise a department on a hunch.
What management should watch for
Four signals, all of them cheap to check.
- A plan that cannot fail If the roadmap contains no proposition that could turn out to be false, it is an inventory of activity. Ask which item, if the underlying belief were wrong, the organisation would abandon.
- Adoption presented as an outcome Seats, licences, pilots and tokens are inputs. They rise reliably, they are easy to report, and they can be delivered without anyone else's consent — which is exactly why they crowd out cycle time, cost per transaction, and margin by function.
- A review cadence that punishes the correct strategy Redesign consumes two quarters before it returns anything. Deployment returns a dashboard in six weeks. At the six-month review, the organisation doing the shallow work will be visibly ahead of the one doing the deep work. Governance that does not anticipate this will select against the right answer, in good faith, on schedule.
- A destination someone else drew A vendor's reference architecture describes what worked in a company whose constraints were not yours. Useful as a catalogue of what is now possible. Worthless as a description of where you are going.
The aspiration, stated properly
Completing the technical foundations is not optional. Secure environments, data pipelines, compliance frameworks — this is the floor, and it must be built. It is also inert on its own, and it is easy to mistake for progress precisely because it is the part with a dashboard.
The aspiration is not a state you arrive at, and any attempt to render it as a milestone will convert it back into a checklist. It is closer to a habit: the standing capacity to notice that a constraint has lifted, to test whether the process built around it is a fossil or a load-bearing wall, and to rebuild when the answer is the former. Some organisations will do this three times and stop. The ones that compound will still be doing it in a decade, on constraints that have not lifted yet.
You will know it has taken hold when someone proposes an AI project and the room treats it as a category error — the way a proposal for an electricity project would land. Not because the technology stopped mattering, but because it stopped being the thing you do, and became the thing you do everything with.