Enterprise AI is crossing an important threshold as we move from AI that answers questions, summarizes information and helps people make decisions to agents that can make decisions and execute work themselves.
But this changes the risk equation considerably – an AI assistant that gives someone a poor answer creates an issue whereas an agent that takes the wrong action inside a critical business process can create an entirely different class of problem.
Yet many organizations are deploying agents without giving them a complete understanding of the environment they are being asked to operate within. The agent may have access to the right data, applications and tools. But does it understand the process, the approvals, the exceptions or the controls? And does it know when it should act, as opposed to when it should stop and involve a human in the loop?
In other words, we are giving increasingly powerful AI access to enterprises whose processes and workflows it doesn’t fully understand yet.
The enterprise has a context problem
For years, the AI conversation has focused heavily on data and as a former data leader at Google, Tesco and Kantar Media, I understand why – models need accurate, relevant and well-governed data to produce useful results. But data tells AI things about your business. It doesn’t necessarily tell AI how your business works.
Consider something as seemingly straightforward as approving an invoice. An agent might have access to the supplier, purchase order, invoice amount and payment terms. But questions arise such as which approval thresholds apply? What happens if the purchase order and invoice don’t match? Is there a compliance control that needs to be satisfied? And when should the agent escalate rather than act?
Much of this knowledge exists somewhere inside most organizations. The problem is that it is fragmented across process models, policies, applications, documentation and, very often, people’s heads. Humans have traditionally compensated for those gaps but agents don’t arrive with that institutional knowledge.
New research suggests this gap is already becoming visible. In The Hackett Group’s recently published 2026 Process Context Study, developed in collaboration with ARIS, 86% of respondents agreed that AI agents cannot be deployed reliably without process context, while 76% expect process context to become very important or critical within three years.
Yet only 22% of organizations report comprehensive, real-time visibility into their business operations, while 59% describe their visibility as fragmented. Businesses increasingly understand the context agents need, but most don’t yet have a complete enough view of their own operations to provide it.
From data context to process context
Process context is the business understanding an agent needs around the task it is performing: workflows, roles, rules, systems, approvals, controls, dependencies and exceptions. It connects what an organization says should happen with what actually happens when work is executed.
And that distinction becomes ever more important as agents become more autonomous. If AI recommends an action to an experienced employee, that person can provide missing context and decide whether the recommendation makes sense. If the agent executes the action itself, that context needs to be available to the agent.
The fundamental question changes from “What can this agent do?” to “What should this agent do in this situation?” Answering that requires an understanding of the business around the technology.
The visibility gap becomes a governance gap
This changes how we should think about AI governance. A great deal of attention is rightly being paid to which systems agents can access, what data they can see and which actions they are authorized to perform. Those controls are essential but authorization alone doesn’t provide operational understanding.
Without end-to-end process visibility, an agent can optimize the task immediately in front of it while creating a problem somewhere else. It can miss an important exception or take an action that is technically permitted but inappropriate in the wider business context.
The Hackett Group research reinforces the connection between context and control. Organizations with greater process context experience reported substantially greater confidence in their ability to govern AI-driven decisions.
But governance doesn’t necessarily have to slow down the impact of agentification. Quite the opposite, in fact – process context connects agent behavior to known workflows, rules, approvals and exception paths. The clearer you are about what an agent should do, what it shouldn’t do and when it needs to escalate, the more confidently you can allow it to operate autonomously within those boundaries. Now, governance becomes an accelerator rather than a brake.
Build the map before deploying the agent
So what should enterprise technology leaders do? I would start with three steps.
First, understand the process before automating it. Don’t begin by asking where you can deploy an agent. Start with the business outcome you want to improve and understand the process responsible for delivering it. Combine process models, policies, roles and controls with evidence of how work actually flows through systems.
Second, identify where autonomy creates genuine value. Not every task needs an agent, and the easiest process to automate isn’t necessarily the most valuable one. Assess opportunities based on business outcome, complexity, available context, risk and potential value. Deploying hundreds of agents is not an AI strategy whereas improving business outcomes is.
And third, make process context available when the agent acts. An agent needs relevant context at runtime: process state, rules, roles and permissions, dependencies, controls, exceptions and escalation requirements. And that context needs to remain governed and current as the business changes.
This is where the architecture around AI becomes just as important as the intelligence of the model itself.
Process is becoming part of the AI architecture
There is also a compelling business case. The Hackett Group study found that organizations with high process context experience are five times more likely to report very successful AI outcomes than organizations with low experience.
That doesn’t prove process context causes better AI performance. But the association is striking and points to an important distinction between experimenters and scalers when it comes to AI.
Small-scale pilots can survive with narrow context and significant human oversight but enterprise-wide deployment is more complex as agents encounter process variation, systems, exceptions and accountability requirements. The research found organizations consequently prioritizing real-time context, testing and guardrails as they prepare for greater autonomy.
For decades, organizations have invested in understanding and improving processes primarily so people can work more effectively. In the Agentic AI era, that understanding increasingly needs to become machine-consumable.
For CIOs and enterprise architects, that means process can no longer sit solely within an operational excellence team. Instead process models, execution data, governance rules and organizational context need to become part of the infrastructure required to deploy AI reliably and cost-effectively.
As the underlying AI models become increasingly powerful and increasingly accessible to everyone, the differentiator will not simply be who has access to the smartest AI rather it will be the companies who can give agents the deepest understanding of their business and put them to work most effectively.
##
ABOUT THE AUTHOR
Guillaume Bacuvier, CEO of ARIS
Guillaume joined ARIS from Kantar Worldpanel, where he served as CEO for four years, leading the transformation and separation of the business from Kantar Group. Prior to this he was CEO at dunnhumby, Tesco’s global retail data analytics firm, and held multiple senior executive roles at Google prior to this. He also has extensive non-executive board experience at CCEP, Veon, and Berger-Levrault.
