UiPath’s 2026 AI and Agentic Automation Trends Report just reported a figure that will feel familiar to anyone who has sat through an agentic AI strategy meeting lately: 78% of executives say their organisations need to fundamentally reinvent their operating models to get full value from the technology.
What makes it sit awkwardly is what McKinsey found when they surveyed the same pool of leaders: more than 80% of companies are seeing no measurable bottom-line impact from their AI investments at all.
So the agreement on what needs to happen is almost universal. The evidence that it is actually happening is rather thin. Something is getting in the way. And the research – across McKinsey, KPMG, Deloitte, and Google Cloud – points to a fairly consistent answer. It is not the technology.
The starting point problem
McKinsey’s research on the agentic organisation found that 89% of companies still run on fundamentally industrial-age structures: functional hierarchies, sequential hand-offs, processes designed around humans completing every step. Only 9% have reached the product-led, agile operating models that came out of the digital era. A mere 1% operate as the kind of decentralised network that agentic AI actually suits.
That is the structural reality most organisations are starting from – and it matters because McKinsey’s own framework for what an agentic transformation requires describes change across five pillars at once: business model, operating model, governance, workforce and culture, and the technology and data foundation. Most organisations have, at best, moved on one or two of those. Almost none are working across all five in any coordinated way, and that coordination gap is where most programmes quietly come apart.
Agentic AI does not slot into existing structures. It demands a different answer to who does the work, how tasks move between humans and systems, and what oversight looks like when autonomous agents are making multi-step decisions without someone watching every move. The 78% of executives who understand that reinvention is required and the organisations that are structurally positioned to deliver it are, at the moment, quite different populations.
Why the readiness gap is so wide
The most consistent finding across the research is that organisations know where they want to get to. What they are far less certain about is how to get there. McKinsey’s 2026 AI Trust Maturity Survey identified knowledge and training gaps as the leading barrier to responsible AI implementation – ahead of budget constraints, ahead of competing priorities. KPMG found the same pattern from a different direction: 65% of business leaders have cited agentic system complexity as their top barrier for two consecutive quarters.
A lot of the conversation around AI readiness focuses on infrastructure investment and talent. Those things matter. But underneath them is a more basic problem: many organisations simply do not have a clear picture of what a well-sequenced transformation looks like end to end. The path from ‘we need to reinvent our operating model’ to ‘here is how we do that, in this order, with these dependencies’ is not obvious, and most teams are working it out as they go.
The second barrier compounds the first. Google Cloud’s analysis of failed agentic deployments is pointed: the most common mistake is introducing autonomous systems into environments with unresolved technical debt. AI does not fix a fragile foundation. It runs faster on one. An agent working across inconsistent data pipelines and poorly governed cloud infrastructure will not produce better outcomes than the humans it replaced – it will produce the same errors at much higher volume, often before anyone has noticed something has gone wrong.
Data privacy, legacy system integration, and cost control remain the top concerns cited by leaders when asked about their AI programmes. These are old problems that agentic AI has made newly urgent. The third barrier tends to get underweighted. McKinsey’s AI Trust survey found nearly two-thirds of organisations cite security and risk concerns as their primary block to scaling agentic AI, well above regulatory uncertainty or technical capability gaps. Only around 30% have reached any meaningful maturity level in governance and agentic AI controls. More striking: despite rising deployment numbers, confidence in how organisations would actually respond to an AI incident has fallen year on year. More agents running, less trust in the ability to manage them.
What the leaders are doing differently
A minority of organisations are pulling ahead, and the data on what they are doing differently is reasonably clear. Deloitte’s 2026 enterprise AI research found that 34% of companies are using AI to genuinely transform how their business operates – new products, redesigned core processes, reshaped business models. The other two-thirds are either optimising around the edges or still largely uncommitted. McKinsey’s data shows that the organisations seeing material returns are not using better technology. They are running the programme differently. Senior leadership is directly shaping AI governance rather than delegating it to technical teams. They have cleared foundational architecture debt before scaling agents rather than after. And they have avoided the agent sprawl pattern that Google Cloud identifies as a primary driver of negative ROI: decentralised teams building disconnected agents without a central governance layer, creating what ends up being a much more expensive version of the mess they already had.
The timeline for getting the structure right is shortening. Multiple research firms now project that 40% of active agentic AI programmes will be abandoned by 2027, primarily because of inadequate governance and business cases that were never properly defined. Deloitte’s finding that 84% of companies have not yet redesigned their workflows around AI capabilities suggests a significant portion of current deployments are structurally exposed to exactly that outcome.
Q&A: Navigating the Agentic AI Shift
What is the functional difference between an AI assistant and an AI agent?
Traditional assistants are reactive, requiring constant human prompting for every sequential step. Agentic AI is goal-oriented and autonomous; it can plan, use external tools, write code, run tests, and self-correct to execute multi-step workflows with minimal human intervention.
How does “architectural governance” prevent AI risks?
Governance establishes the boundaries of autonomy. By using a modular architecture and strict design patterns, you restrict an agent’s access to specific, pre-approved functions. This ensures the agent cannot circumvent security protocols or hallucinate its way into unauthorised system modifications.
Is it safe to let AI agents autonomously refactor legacy code?
Only if the target environment has a high degree of test coverage and clear service boundaries. Unleashing an agent on tightly coupled, undocumented legacy code will likely accelerate technical debt, as the agent will create fast, brittle workarounds rather than sound architectural solutions.
What is agent sprawl, and why does it kill ROI?
Agent sprawl is what happens when individual teams build and deploy agents independently, without a central governance layer. You end up with dozens of redundant, siloed agents that duplicate work, create conflicting outputs, and multiply your security surface area. The cost compounds quietly until someone does an audit and finds they have been paying for four agents doing the same job with no oversight of any of them.
How do you know whether your data infrastructure is ready for agentic deployment?
Ask whether you would trust a human contractor, given access to your current data pipelines, to make autonomous decisions on your behalf. If the answer is no – because the data is inconsistent, access controls are patchy, or lineage is poorly documented – an agent will hit the same walls, faster. Readiness means the data is governed and trusted, not just technically accessible.
Why should AI governance sit at executive level rather than within the technical team?
Because the decisions that matter most – which workflows agents can own, what the escalation paths are, how accountability is assigned when something goes wrong – are business decisions, not engineering ones. Technical teams can implement governance frameworks, but they cannot set the organisational risk appetite that those frameworks need to reflect. McKinsey’s research is fairly direct on this: organisations where senior leadership actively shapes AI governance generate materially better business value than those that delegate it downward.
What does McKinsey mean by the five pillars of an agentic organisation, and do you need to tackle all of them at once?
The five pillars are business model, operating model, governance, workforce and culture, and technology and data. You do not need to move on all five simultaneously, but you do need a clear view of how they connect – because changes in one create dependencies in others. Redesigning workflows without addressing governance produces agents nobody trusts. Investing in infrastructure without redesigning the operating model produces expensive infrastructure that nobody uses differently. The sequencing matters as much as the scope.
Closing the gap
The 78% figure from our intro is not evidence that the industry has worked this out. It is evidence that the industry knows what it needs to do and, largely, does not yet have the foundation to do it. What distinguishes the organisations making real progress is that they have treated operating model transformation as a delivery problem rather than a technology selection problem. Governance is embedded into how things get built, not reviewed afterwards. Architecture decisions are made before agents are scaled, not in response to the problems that scaling creates. And there is a coherent view of how the five pillars of an agentic organisation connect – so that work on data, platform, workforce, and governance is happening in a coordinated way rather than as a set of parallel, disconnected efforts that nobody has formally joined up.
That is the combination Vertex Agility is built around. We work with technology leaders across AI, Cloud, Data, and Platform Engineering, and what we find consistently is that the organisations struggling most are not short of ambition or budget. They are short of a coherent delivery framework that connects those investments into something that actually compounds.
If your initiative is at that point – either trying to map the roadmap or trying to get an existing one unstuck – we would be glad to talk. Get in touch to find out how we can help.