Why Enterprise AI Breaks Down Between Business Units

Deloitte's 2026 Global Technology Leadership Study, based on a survey of more than 660 technology executives, found that 81% of leaders are confident they can scale AI. In the same survey, 75% said their operating model has to fundamentally change to get real value from it, and 42% reported low or no return on their AI investment so far. Confidence in the technology and confidence in the organisation around it are pulling in different directions. That mismatch likely shows up faster and costs more for a diversified enterprise running AI across a dozen business units than for a single-business company, though Deloitte's survey wasn't measuring that comparison directly.

What "operating model" actually means here

An AI operating model is the answer to a simple question: who decides how AI gets built, governed and used across the organisation. Enterprises building one generally choose between three structures. A centralised model puts one team, often called a centre of excellence, in charge of tools, standards and governance for everyone. A federated model lets each business unit build and own its own AI, moving faster locally but risking duplicated work and inconsistent standards. A hub-and-spoke model sits between the two: a central function sets shared standards, infrastructure and governance, while business units own their own use cases within those guardrails.

None of the three structures is inherently right or wrong. A holding company with ten genuinely different businesses, different regulators, different customers and different risk profiles, has a different answer to this question than a single bank with several product lines. Deloitte's research doesn't tell us which structure enterprises are actually choosing, or whether they've chosen one at all. What it does show is that most surveyed leaders think their operating model needs to change, and that few report having redesigned at scale. Read together, those two findings point to a large share of enterprises running AI initiatives across business units without a settled answer to the ownership question, though that reading is ours rather than something the survey measured directly.

The evidence that structure alone doesn't solve this

A separate Deloitte AI Institute survey, polling nearly 3,700 professionals across functions, found that 48% of respondents said their organisation had introduced AI without redesigning the workflows or roles it sits inside. Only 12% reported redesign at scale, with a genuine operating model behind it. Those two figures alone show that introducing AI and redesigning around it are different things for a substantial share of organisations, and that only a small minority report having done the latter properly.

This matters more for a diversified enterprise than for a single business, because adding AI without redesigning the workflow around it produces a different AI in every business unit that does it. One division's version of a customer service assistant learns whatever policies and shortcuts that division's team happened to feed it. Another division's version, built independently, learns something else entirely. Neither is wrong on its own terms. Together, they mean the group has no single, coherent answer to a basic question a board might ask: what does our AI actually do, and does it do the same thing the same way everywhere we operate.

A concrete illustration

Picture a conglomerate with ten operating businesses spanning retail, logistics, financial services and real estate, each running its own pilot programme because each pilot looked sensible in isolation. The retail business built an AI assistant trained on its own return policy. The logistics arm built one trained on its own delivery exceptions process. Both are reasonable, useful tools inside their own business unit.

The problem shows up the moment a customer, an auditor or a regulator asks a question that spans more than one of those businesses, or the moment the group tries to apply one governance standard, one data policy or one risk threshold across all ten. There is no single team that can answer for how AI behaves across the group, because nobody was ever responsible for that question. Fifty systems and ten businesses each solved their own local problem competently. None of them solved the group's problem, because it was never assigned to anyone.

Why this is a different problem from scaling a single pilot

This isn't the same challenge as taking one AI pilot into production, which is a question of whether a single system can handle an organisation's real data, workflows and governance once it leaves the sandbox. The operating model problem shows up even when every individual business unit has already done that work well. Ten business units can each successfully scale their own AI pilot into production and still leave the group with ten incompatible answers to what AI is allowed to do, who's accountable for it and how it's supposed to reflect the group's actual standards. Scaling within a business unit and coordinating across business units are two different problems, and solving the first doesn't touch the second.

What this means for enterprise and conglomerate leaders

The choice between centralised, federated and hub-and-spoke models matters less than actually making the choice deliberately, and then making sure two things travel with it regardless of which structure gets picked. The first is governance: who decides what an AI system is allowed to do, and who's accountable when it does something it shouldn't. The second is context: whether each business unit's AI actually reflects that business's own policies, precedent and risk appetite, or a generic version borrowed from wherever it happened to be built.

Deloitte's figures speak directly to the structure and redesign half of this: 75% say their operating model needs to change, and only 12% report having redesigned at scale. The context half, whether each business unit's AI actually reflects that business's own policies, precedent and risk appetite, isn't something the survey measured; it's the piece we'd add to complete the picture. Both halves matter together: an enterprise that fixes structure but not context will end up with a well-governed set of AI systems that still don't understand the businesses they're meant to serve, and one that fixes context but not structure will end up with pockets of genuinely useful AI that nobody can coordinate or govern consistently at the group level.

AinTools works with diversified enterprises building AI that understands each business unit without losing a coherent view across the group. Our Platform page sets out how we approach it.

Sources

• Deloitte, "From Operators to Orchestrators: Deloitte's 2026 Global Technology Leadership Study" (official press release, deloitte.com): deloitte.com
• Deloitte AI Institute, AI Pulse Check series, "Enterprise AI trends 2026: AI transformation strategy": deloitte.com
• Deloitte Tech Trends 2026 (17th annual report), supporting context on enterprise operating models: deloitte.com