Over the past year, the conversation around enterprise AI has shifted dramatically. We’ve all seen that AI is extremely competent when it comes to generating content or answering questions with advanced reasoning, but what G2000 companies are now asking is whether agents can be trusted to execute work efficiently and effectively across the organization.
This is where most of the value creation will lie for large global companies, but many make the mistake of prioritizing quick deployment of agents, expecting instant results that can reassure shareholders, customers and other stakeholders that the company is not “falling behind” in the AI race.
The fact is that most organizations that skip the initial hard yards of standing up detailed operational foundations for agent deployment discover they’ve simply automated complexity, with high failure rates, increased risk and disappointing results.
Large language models are extraordinarily capable, but they do not understand your business context. They do not know how work should flow across departments, where exceptions arise, who has authority to approve decisions, which policies take precedence, how compliance is maintained or what success looks like for each business process.
I also remained amazed by the number of large organizations that still do not have a detailed and documented understanding of how work truly happens within their business and therefore try to deploy AI within processes they are not even able to describe accurately in the first place.
Without that understanding, every agent action or decision becomes more uncertain, which is why so many enterprises are discovering that while deploying agents is the easy part, getting ROI and controlling risk is much harder.
One of the biggest misconceptions about AI is that business processes are predictable and consistent. In fact, they’re not — our typical G2000 enterprise customer has thousands of variations, exceptions and dependencies that have evolved over years, sometimes decades. For example, a customer order follows a different path depending on geography, customer type, inventory availability, regulation, contractual obligations and dozens of other variables. This is intrinsic to large, complex operations operating in many jurisdictions, and cannot be removed just by deploying AI agents.
This operational complexity is what makes enterprise AI so challenging. An agent may perform flawlessly in a controlled pilot, but real businesses rarely operate under ideal conditions. Without a clear understanding of how work is designed to flow — and, crucially, how it flows when exceptions occur — agents struggle to make consistent, reliable decisions in production. The challenge isn’t the intelligence of the AI: it’s giving agents the context they need to navigate the day-to-day realities of enterprise operations.
This brings us back to one of the core principles of process improvement: don’t automate a process you can’t describe in full in all its variations, and don’t automate a broken process either. And the same rule applies to Agentic AI — agents learn from the environments they’re deployed into so they will amplify dysfunction and error rates.
In contrast, I’ve seen organizations achieve remarkable results by taking the opposite approach. Rather than rushing to deploy new technologies, UK-based retailer Boots first created a connected process architecture spanning more than 2,000 business processes. That visibility enabled the company to redesign core finance processes, reducing one process from 220 steps to just 40, improving efficiencies by up to 75% and creating the foundation for starting to roll out agents to take on key responsibilities.
As Boots’ Head of Finance BPM and Continuous Improvement, Lee Oates, puts it: “It’s very easy to think that you can just chuck AI in and it can solve all your problems. But if we don’t prepare our foundations, our data and our processes, that’s a fundamental mistake. Preparing those foundations helps us pick the right AI solutions and put the right governance around them “
Similarly, Lockheed Martin has made operational foundations a core part of its AI strategy. CIO Maria Demaree says the company is first standardizing business processes and building a model-based enterprise before scaling AI across the organization — recognizing that AI is most effective when it operates on trusted operational foundations.
As organizations move from automation to autonomous agents, governance becomes just as important as intelligence. After all, every enterprise operates within clearly defined boundaries covering rules such as who can approve a payment, when decisions should be escalated, which policies take precedence and what controls must be followed.
These aren’t questions an AI model can infer from transactional data alone, so they require explicit operational knowledge. Without those guardrails, businesses face an impossible choice: agents that escalate every decision and deliver little new value, or rogue agents that act too freely and create unacceptable levels of risk.
Organizations that successfully deploy AI at scale won’t be the ones who bolt governance on afterwards but instead embed it into the way work is designed from the outset, giving agents clear boundaries within which they can operate confidently and safely.
But another challenge emerges as organizations move beyond isolated AI pilots and use cases. Business processes rarely exist in isolation, and every decision creates downstream consequences across multiple teams and systems. If agents operate with different assumptions about how work should happen or do not have a full understanding of upstream and downstream consequences of the work they are undertaking, inconsistency quickly becomes enterprise risk.
That’s why organizations are investing in a governed Digital Twin of the organization to act as the single source of truth and to understand in detail interdependencies between processes. Leonardo, one of the world’s leading aerospace and defence companies, recorded more than 5,000 process models across its global operations to establish a common operational foundation; now it provides the context, rules and governance needed for agents to operate reliably and efficiently across its highly complex engineering and manufacturing environments.
And Hyundai has built a Digital Twin of operations at its $7.6bn Georgia hub — the largest car manufacturing factory in the US — that mirrors the physical plant in real time to predict optimized outcomes and identify the root causes of production issues, reducing costs and being able to respond more effectively to disruption on the line.
Perhaps the biggest misconception about Agentic AI is that success can be measured by the number of agents deployed. This is a fallacy — the real question is whether business outcomes improve. Are customer journeys faster and costs lower? Has compliance improved and risk been reduced?
Without operational baselines and process KPIs, organizations have no reliable way to determine whether AI is genuinely improving performance or simply changing how work is executed.
The enterprises that will create new business value from AI all share one characteristic: they understand that successful AI starts long before the first agent is deployed. They are investing in building a clear operational understanding of how their business works — connecting processes, systems, people, governance and performance into a trusted foundation for execution.
That foundation gives AI the context it needs to operate reliably and at scale, delivering the speed to value, reduced risk and new productivity that all CEOs are under pressure from their shareholders and boards to demonstrate.
Put simply — the winners in the Agentic AI era won’t be the companies that deploy the most agents the quickest but those whose agents understand their businesses the best.