Most AI programmes are not failing because the technology underperformed. They are failing because the organisation around the technology never changed.
Somewhere this week, a chief executive is going to walk out of a quarterly review wondering whether anyone is going to say the thing she has been thinking for three months.
Her firm has spent real money on AI. The demos have been good. There are dashboards. There is a head of AI. There is a partnership with one of the big model vendors. There are pilots in three divisions. The board has been told, in writing, that AI is now operational at scale.
And yet. The income statement looks remarkably like last year's. Revenue per employee has not moved. Margins are flat. Decisions still take as long as they used to take. New products are launching on the same cadence they did before any of this began.
If that sentence has been said in your boardroom in the past six months, your organisation has plenty of company. In Green Everest's language, it is stuck in State 2 of the Intelligent Organisation Pipeline. BCG's latest AI value-gap research found that only about five percent of firms are "future-built" and capturing substantial AI value. The other ninety-five percent are working hard. Most of them are in exactly the trap I want to describe.
State 2 is the most expensive place to be in 2026, and almost nobody who is there knows they are there.
Quick answer
State 2 is where AI looks busy but the business has not changed. Tools have arrived, people are experimenting, and pilots are visible, but workflows, roles, decision rights, governance, and incentives remain mostly untouched. The way out is not more pilots. It is five parallel shifts: rewrite jobs, rebalance spend toward work redesign, govern agents as a portfolio, design workflows around agents doing work, and reward the behaviours that make the new operating model real.
What State 2 actually feels like
It feels like progress. Tools have arrived. People are using them. Time-saved figures are real - somebody really is finishing their Tuesday afternoon work by lunchtime. Executive sponsors can point at things. The slide deck for the next board meeting almost writes itself.
What is missing is harder to see. The work itself has not been redesigned. The roles have not been rewritten. The way decisions get made has not changed. The way people are paid and promoted has not changed. The way the agents are governed - who owns which one, who can switch it off, what happens when it goes wrong - has not been thought through, because the agents are still treated as software rather than as a class of decision-maker that now operates inside the firm.
The result is the pattern Harvard Business Review described in 2025 as workslop: AI output that gets generated, gets circulated, looks like work, and produces no economic value because the system around it was never redesigned to convert it into value.
On our five-level AI work-design maturity view, State 2 sits in the left half of the curve. AI is present as a thought partner or contextual assistant, but the organisation has not yet crossed into redesigned process territory.
The trap is not that anyone made a bad decision. Every individual choice - buy the licences, run the pilots, appoint the head of AI, sign the cloud partnership - was reasonable. The trap is that all of that adds up to a technology rollout, when what the moment actually requires is something different in kind.
It requires an organisational passage. Closer to going public than to procuring a new platform. Closer to entering a new market than to launching a new product.
That is the shift in framing that the next eighteen months will reward.
The five shifts
The way out is not more pilots. It is not a bigger model. It is not, despite what most vendors will tell you, a new platform. The way out is five specific shifts, and they have to run in parallel. Trying to sequence them - architecture first, change management second - is the most reliable way to remain at State 2 for another eighteen months.
-
Shift one. Tell people their job has changed.
For every role whose work is now being done with an agent in the room, the job description needs to be rewritten in three honest columns: what new skills the role requires, how the person should be spending their time differently, and what new behaviours will define a good performer.
This sounds like an HR exercise. It is not. It is a leadership artefact. Until a person has been told, in writing, that their job has changed and what it has changed to, no amount of technology investment around them will produce a return. This is the single most consequential thing a CEO can do this quarter.
-
Shift two. Spend the money on the work, not the technology.
BCG's evidence on this is, by now, unambiguous. The firms that capture AI value do not treat the model as the whole programme. They invest heavily in people, work redesign, training, governance, and the operating-model changes that allow AI to change actual business outcomes.
Most firms in the trap are spending the other way around. Reversing that ratio is uncomfortable, because the work-design spend has no glossy pitch deck behind it. It is internal work. The CFO who funds it anyway is the one whose firm gets out.
-
Shift three. Put the agents on the calendar.
The firms that have crossed into the next state share one simple discipline: they treat agents as a governed portfolio, not as software. There is a weekly forum where every production agent is reviewed. Every agent has a named owner. Every agent has an off-ramp - a documented way of switching it off when something goes wrong, with a person whose name appears next to the button.
The board sees the agent portfolio on the same cadence as it sees the financial portfolio. None of this is exotic. It is the calendar discipline a firm already applies to the things it considers important. The shift is to apply it to this.
-
Shift four. Design for the agent doing the work, not the human prompting it.
State 2 systems mostly look like a chat box: a person asks a question, the model answers, the person decides what to do next. The next state looks different. An agent is set in motion against a defined goal, draws on the firm's data through a governed connection, takes some actions on its own, escalates the ones that need a human, and is evaluated overnight against whether it actually produced the outcome it was supposed to produce.
This is a deeper change than it sounds. It is not about a new tool. It is about redesigning the workflow so that the agent - not the human - is the unit doing the work.
-
Shift five. Pay and promote for the new behaviours.
This is the shift that closes the loop, and the one that most firms get to last and treat as optional. It is neither. As long as the performance review still rewards individual output, the high-volume personal decision-maker, and the manager who supervises humans, the organisation will quietly remain at State 2 no matter what else changes.
The firms that have crossed tell us the moment a senior promotion turned on the new behaviours - on whether someone could direct a mixed team of humans and agents, on whether they redesigned their function rather than just digitising it - was the moment everyone else believed the change was real.
What lies on the other side
Firms that complete these five shifts over the next eighteen to twenty-four months reach the state in which AI begins to show up in the income statement. First as margin. Then as new propositions. Then as the cost-to-serve advantages that competitors at State 2 simply cannot match.
This is not a story about technology any more. The technology has done its part. The next part is ours to do. It is organisational, behavioural, and unglamorous. It is the work that most firms in 2026 are putting off and that a small minority are doing on purpose.
That minority is the one whose names we will be writing about, three years from now, in the case studies that explain what separated the firms that captured the AI decade from the firms that watched it go past.
The question is no longer whether the company has AI.
The question is whether the company has redesigned itself so AI can create economic value.
FAQ
What is State 2 in AI transformation?
State 2 is the stage where AI activity is visible but not yet structural. The organisation has tools, pilots, and usage, but the work, roles, governance, and incentives have not changed enough to produce measurable business value.
Why do AI programmes stall after pilots?
They stall when leaders treat AI as a technology rollout rather than an operating-model change. The pilot may work, but the surrounding workflow, decision rights, data access, accountability, and performance system are still designed for the old way of working.
How should leaders move beyond State 2?
Start by rewriting affected roles, funding work redesign, governing agents as a portfolio, designing workflows around agents doing bounded work, and changing performance systems so new human-AI behaviours are rewarded.
Daniel Retief Maritz is the chief executive of Green Everest, a strategy and AI transformation consultancy based in Cape Town. Green Everest works with mid-market organisations on the passage from State 2 to the re-architected, AI-native firm.
Sources: BCG, The Widening AI Value Gap; BCG, Where's the Value in AI?; Harvard Business Review on AI-generated workslop; Gartner AI Maturity Assessment.