Seventy-six percent of organizations say their current operations and infrastructure cannot support AI agents. That number comes from a Celonis process optimization report cited this week in MIT Technology Review. The same organizations, 85% of them, say they want to be agentic within three years.
Those two numbers do not add up. And the gap between them is not a technology problem. It is an operating model problem.
PwC UK’s global CTO for workforce consulting, Prasun Shah, put it plainly: companies are “embedding AI employees into what is a human operating model.” He called it adding sticky tape to parts of an operating model that is already breaking.
That is the most accurate description of enterprise AI adoption I have seen this year.
The sticky tape pattern
Here is what it looks like in practice. A company buys an AI agent platform. They plug it into customer service, or HR, or sales. The agent sits inside the same queue, the same escalation path, the same approval chain that was designed for a human worker doing the job manually.
The agent can handle a thousand customer interactions in the time it takes a person to handle ten. But the escalation rules, the handoff points, the approval gates, the reporting metrics were all built for human speed and human judgment. The agent is faster than the system it sits in.
So what happens? The agent finishes in seconds and waits. It routes to a human who routes back. It generates output that nobody reviews because the review process was designed for a tenth of the volume. Or worse, the team measures success by “calls handled” and declares the agent a win while customer satisfaction stays flat.
BCG estimates that AI agents could accelerate business processes by 30% to 50% and cut low-value work time by 25% to 40% when deployed at scale. But “at scale” means inside a workflow designed for agents. Not bolted onto a workflow designed for people.
What redesign actually looks like
The MIT Technology Review piece covered a framework called Agentic Business Transformation, or ABT. The useful part is not the acronym. It is the three areas it forces you to look at: your technology stack, your workforce structure, and your success metrics.
Start with the stack. Your current systems were built for humans clicking through applications one at a time. An AI agent operates across multiple systems simultaneously, at machine speed. If your architecture forces the agent through the same single-threaded paths your team uses, you have not deployed an agent. You have deployed a fast human.
Then workforce structure. McKinsey predicts that by 2030, three-quarters of current jobs will require redesign, upskilling, or redeployment. That is not a layoff number. That is a job-description number. Managers who spent their careers coordinating execution will need to manage hybrid teams where some of the workers are agents. That requires different skills: trust calibration, explainability, knowing when to override and when to let the agent run.
Then metrics. This is where most companies fail first. One enterprise customer profiled in the report switched from measuring cost per query and AI accuracy to measuring the percentage of contracts reviewed without human escalation. Their measured ROI from AI agents tripled in two quarters. Same agents. Same technology. Different measurement of what success means.
Why waiting makes this harder, not easier
Goldman Sachs Research published a report in May projecting a 24-fold increase in AI token consumption by 2030. But they also forecast that only 12% of knowledge workers will be using agentic AI by that point, rising to 37% by 2040.
That is a 14-year adoption curve. And the reason is not the technology. Goldman Sachs is explicit about this: enterprise adoption takes longer because it requires testing, integration, compliance alignment, and organizational change. The companies that start redesigning their operating model now will be running agents in production workflows while their competitors are still running pilots.
The gap compounds. Every month an organization spends bolting agents onto a human operating model is a month where the return stays marginal and the skeptics gain ammunition. “We tried AI agents and they didn’t move the needle” is almost always a statement about the workflow, not the agent.
The one question to ask
If your team is evaluating AI agents right now, or already running them, there is one question that separates the companies getting results from the ones adding sticky tape: Did you change the workflow before you deployed the agent, or after?
The organizations reporting real ROI changed the workflow first. They mapped the process, identified where human speed and human judgment were the actual bottleneck versus where they were just the default, and redesigned before the agent touched production.
The ones reporting disappointment deployed first and hoped the agent would force the change. It does not work that way. An AI agent will run whatever workflow you give it. If the workflow is broken, the agent will run the broken workflow faster.
Seventy-six percent of companies already know their operating model is not ready. The question is whether they will redesign it before they deploy, or spend the next two years wondering why the agents are not delivering.