The AI Stack.
Exploring why nine in ten enterprises are running an AI strategy with one engine in the bay — and what it costs to keep it that way as the field moves to five.
In December 2025, McKinsey released its ninth annual State of AI report. Inside the dense pages of statistics, one number stood out for its quietness. 88% of executives surveyed planned to increase AI-related budgets in 2026. The accompanying breakdown of what they intended to spend that money on was where the strangeness began.
Roughly nine in ten of those budget increases were earmarked for one specific AI capability. Generative — the family of large language models that produce text, summaries, drafts, code, images. The engine everyone has been touching, demonstrating, procuring, and proudly listing on quarterly earnings calls since 2023. The other four engines barely featured.
This piece is the first of a five-part series I’m calling The Organisation of the Future. Over the next five Wednesdays I’ll lay out — across roles, structures, and stack — what the 2030 enterprise actually looks like once multiple AI capabilities are integrated and orchestrated. The pieces stand alone, but they compound. To talk usefully about how AI restructures work, we need to start with the parts.
There are five. Most leaders are treating them as one. That is the most expensive vocabulary mistake of the decade.
“Traditional AI adoption has climbed to 72% over the past eight years, but organisations are adopting agentic AI rapidly, well before they have orchestration strategies in place.” — MIT Sloan Management Review × BCG, 2025
That sentence describes a particular kind of breakage. You have one type of AI everywhere. You bolt on another type of AI without a plan for how the two interact. Then you wonder why the productivity numbers don’t add up.
This piece names the five engines, what they do, where they’re real, where they’re brittle, and — most importantly — why the field’s collective fascination with one of them is the structural mistake that will determine which firms compound advantage by 2030 and which firms get stuck.
A short, honest history of why we got here
The vocabulary collapse — the way “AI” came to mean “generative AI” in mainstream business conversation — has a date. It happened over roughly 18 months between late 2022 and mid-2024. ChatGPT launched. Then GPT-4. Then Claude, Gemini, Mistral, Llama. Boards started asking about AI. The market for “AI strategy” presentations exploded. Vendors learned that the word “AI” sold faster than “machine learning” or “computer vision” or “operations research” had ever sold. The semantic field shrank.
The shrinkage was efficient. It made the technology procurable. A board could be told “we are deploying AI” and understand, instantly, that there would be a chatbot. The procurement department could be told “buy the AI” and know which licences to negotiate. The McKinsey-deck-driven AI strategies of 2023 and 2024 worked because the word had narrowed to a single referent.
What was lost was the rest of the field. Predictive AI — running quietly in fraud detection, credit scoring, demand forecasting since the 90s — became unfashionable, almost embarrassing. Computer-vision projects struggled for budget because they didn’t fit the chatbot shape of the conversation. Reinforcement-learning teams were quietly disbanded at firms that “moved their AI investment to generative”. The legacy AI infrastructure of most enterprises was treated as old, even when it was producing more measurable value than the new generative initiatives.
By 2026 the cost of that shrinkage is becoming visible. The firms that “won” with generative AI are the ones who already had the other engines running and integrated the new one cleanly.
The firms that bet on generative as a single solution have spent two years discovering that a chatbot, on its own, does not transform an operating model.
To get the rest of the decade right, we need to recover the vocabulary. Five engines. Each with a distinct shape of value. Each with a distinct cost of ownership. Each with a distinct failure mode. None of them, on its own, is the strategy.
Engine 1: Generative
What it does. Produces content. Text, code, images, audio, video, structured outputs. By predicting likely continuations of an input.
Where it is real. Knowledge work involving drafting, summarising, translating, restructuring, idea generation. Customer support. Marketing first drafts. Code completion. Internal-search-with-explanation. Anywhere a human used to stare at a blank page or a long document.
Where it is brittle. Anywhere accuracy matters more than fluency. Generative AI does not have a concept of true. It has a concept of likely. Those are different things. The Deloitte fabrication scandals of late 2025 — A$440,000 worth in Australia, a near-million-dollar healthcare report in Canada — are not anomalies. They are what generative AI does when nobody verifies. It produces fluent, plausible, well-structured wrongness.
The deeper truth. Generative AI is a median engine. It pulls work toward typicality (I wrote about this last week in The Originality Tax). That makes it perfect for low-stakes drafting and dangerous for high-stakes thinking. The professionals who treat it as a writing assistant will be fine. The ones who treat it as a thinking partner are quietly outsourcing their judgement to a probability distribution.
Engine 2: Predictive
What it does. Forecasts an outcome from historical patterns. A churn score, a demand curve, a credit risk, a fraud likelihood, a delivery ETA, a sensor failure window.
Where it is real. Insurance, banking, supply chain, energy, healthcare diagnostics, fraud detection, predictive maintenance. Most of the AI value already in production at Fortune 500 firms is predictive — and has been since long before “AI” meant chatbots. The numbers are quietly enormous: 15-18% reductions in inventory cost from hierarchical reinforcement-learning approaches, 20-30% reductions in working capital from AI-driven demand forecasting and inventory optimisation. Nobody puts those on a slide because they are not interesting.
Where it is brittle. Distribution shifts. The world changes; the model was trained on what came before. The 2020 demand-forecast model that survived COVID is famous because nearly all of them did not. Predictive models also offer a single number with high confidence, which encourages humans to defer instead of think. That is its own failure mode.
The deeper truth. Predictive is the unsexy engine that has paid every bill in enterprise AI for the last twenty years. It does not write you a memo. It tells you what is likely to happen, how confidently, on what basis. It is the engine generative AI conspicuously cannot replace — though many vendors have tried.
Engine 3: Perceptive
What it does. Turns raw sensor data — pixels, audio waveforms, depth maps, vibration signatures, electrocardiograms — into structured states. The defect on the assembly line. The tumour on the scan. The shoplifter at the self-checkout. The fatigue in the operator’s voice.
Where it is real. Manufacturing (32% of computer-vision deployment by 2025, per Markets and Markets), healthcare imaging, automotive driver assist, agriculture, retail loss prevention, security. Visual AI systems can now detect assembly defects in under 200 milliseconds and reduce unplanned downtime by 50%. The market — $23 billion in 2025, projected $63 billion by 2030 — is one of the fastest-growing in enterprise tech, and almost none of that growth is in chatbot-shaped form.
Where it is brittle. Edge cases. The model was trained on what was photographed. The condition that wasn’t photographed because nobody knew it was a condition is the condition the model misses. Perceptive AI also fails silently — a vision system that should detect a defect and doesn’t has no way of telling you it failed. It just confidently says “no defect.”
The deeper truth. Perceptive AI is what turns a software product into an operational one. It connects the digital model of the business to the physical state of the business. The firms that have not yet wired up perceptive AI to their physical operations are running the digital twin of an organisation they cannot see.
Engine 4: Agentic
What it does. Takes actions in the world. Opens browsers. Calls APIs. Writes to databases. Dispatches emails. Escalates tickets. Runs multi-step workflows that previously required a human in the loop at every junction.
Where it is real. Tightly-bounded enterprise workflows: ticket triage, lead enrichment, basic customer-service resolution, internal-search-and-action, meeting prep, scheduling. The number of organisations experimenting with agentic AI is high — 62% by McKinsey’s count. The number actually scaling agents in production is much smaller — 23%. The number running agents reliably enough to take revenue or compliance risk is smaller still.
Where it is brittle. Open-ended environments. Long task chains where errors compound. Anywhere the cost of a mistake is high relative to the cost of the action. Gartner forecasts that 40% of agentic AI projects launched in 2025-2026 will be cancelled by 2027 — not because the technology fails, but because the orchestration around it fails. Most enterprises do not yet have the verification, recovery, or escalation patterns required to deploy agents safely.
The deeper truth. Agentic AI is the youngest, most fragile, and most over-promised of the five engines. It is also, by 2030, likely to be the most consequential. The firms that learn to deploy agents safely — meaning with clear boundaries, fast fallback, and human verification at the high-stakes nodes — will reset what one mid-level employee can accomplish in a day. The firms that deploy agents naively will appear in postmortems.
Engine 5: Optimisation
What it does. Decides — given constraints, objectives, and dynamic state — what action minimises cost or maximises return. Reinforcement learning, mathematical optimisation, dynamic pricing, route planning, scheduling, bid optimisation, network management.
Where it is real. Logistics (route and inventory), pricing (e-commerce, ride-share, airlines), advertising (real-time bidding), energy (grid management), telecoms (network optimisation). Hierarchical reinforcement learning approaches now reduce inventory costs by 15-18% on real industrial deployments (OpenReview, 2025). Distributed AI-enabled control systems integrating IoT sensors, deep-learning forecasts, and RL decision policies are moving from research papers into operational pipelines.
Where it is brittle. Reward design. The optimisation engine maximises whatever you tell it to maximise. If the metric is wrong, the optimisation makes the wrong thing happen faster and at scale. Optimisation is also brittle to environments it has not seen — a pricing model that has never seen a recession is a model that has never seen a recession.
The deeper truth. Optimisation is invisible because it lives upstream of the user. Its outputs are decisions, not artifacts. You cannot demo optimisation in five minutes. The board cannot see what it does. Which is why it is consistently underbought in firms whose AI strategy is procurement-driven — and consistently the engine that produces the highest measurable ROI in the firms that have it running.
Why one engine is not a strategy
Most enterprises in 2026 are running a generative-AI strategy. They have a Copilot. They have a custom GPT. They have a chatbot or three. They have, in McKinsey’s framing, “scaled an agentic system” — meaning they have one workflow that calls a model. That is one engine.
A one-engine stack is a stack that can do one thing. It can write but cannot see. It can summarise but cannot predict. It can suggest but cannot decide. It can draft a fraud report — but it cannot detect the fraud. It can summarise a maintenance log — but it cannot tell you the bearing is two weeks from failure. It can write a customer email — but it cannot decide which customer to email first.
The firms that are compounding real advantage in 2026 are the ones running multiple engines in choreography. A perceptive model identifies the defect on the line. A predictive model estimates downstream production impact and supply implications. An optimisation model reschedules production around it. A generative model drafts the customer-comms package. An agentic model dispatches the field engineer with the right parts.
None of those engines is the headline story. The choreography is the story. And the choreography is the thing most enterprises in 2026 do not yet have a vocabulary for, much less a role to own.
What this means
For early-career people: stop building deep skill on one engine. The professionals who will compound value through 2030 will know how to interrogate at least three. Pick one to specialise in deeply — but be conversant in all five. The career risk of being “the prompt engineer” five years from now is identical to the risk of being “the Excel macro expert” was in 2002 — when Excel macros were still cool.
For hiring managers: the most undervalued profile in 2026 is the candidate fluent across multiple engines. They are rare because the market hasn’t named the role yet. Look for engineers who have shipped both an ML model and an LLM workflow. Look for product people who can talk fluently about both confidence intervals and prompt evals. They cost more. Hire them anyway. They will be the operators and workflow designers I describe in the next four pieces.
For leaders: audit your AI strategy for engine balance. If your line items are 90% generative, you have a chatbot strategy, not an AI strategy. Insist that at least two of the other four engines be on the roadmap inside twelve months. The competitor that ships a perceptive-and-predictive workflow before you ship your second chatbot is the one taking your margin in 2027.
The uncomfortable truth
Generative AI was easy to buy because it was easy to demo. Type something. Get something. The procurement decision became trivial. The success criteria became “did the demo work.” The strategy became “deploy a chatbot.”
The other four engines are hard to buy because they are hard to demo. They require integration with the messy parts of the business — sensors, logs, ERPs, OMS, dispatch, finance systems. They cannot be demonstrated in five minutes. They cannot be wrapped in a chat interface. They require, in the most literal sense, doing the work.
Most enterprises in 2026 will not do the work. They will buy more generative licences and call it a strategy. The boards will be told that AI investment has tripled. The press releases will land. The Copilot deployment numbers will be cited at the next earnings call. None of that constitutes an AI strategy. It constitutes a generative-AI strategy.
The firms that build the full stack — the ones running all five engines in choreography by 2028 — will, by 2030, be running operating models the rest of the field cannot match. The gap will not look closeable, because the rest of the field will still be staffing for a one-engine world. They will keep hiring prompt engineers when they need verifiers. They will keep buying chatbots when they need optimisation pipelines. They will keep procuring AI as a product when they should be building it as a stack.
The organisation of the future is not a one-engine organisation. It is not a chatbot wrapped in a corporate brand. It is a coordinated stack of generative, predictive, perceptive, agentic, and optimisation capabilities, run by people whose roles do not yet exist on most org charts, supervised by structures most boards have not yet drawn.
This series, over the next four weeks, is about who those people are, what they do, and how the org around them takes shape.
Start with the vocabulary. Five engines. Most leaders are treating them as one.
The strategic question of the next 36 months is not “which AI vendor.” It is: how many engines are you actually willing to run?