Resources

McKinsey’s State of AI research surfaced a striking finding: AI high performers are nearly three times as likely as others to fundamentally redesign their workflows when deploying AI.

The top performers are not using more sophisticated technology than everyone else. They understand the work before they try to change it.

Three examples show what happens when that understanding is missing.

A company deploys AI prospecting tools to increase pipeline and close rates. Adoption goes up, but sales leaders still spend much of their time handling fulfillment escalations and rescuing accounts. The result is more AI-generated leads without more time to sell, because the post-sale workflow was never addressed.

Another company deploys a generative AI assistant to reduce average handle time in a customer care center. Agents still escalate constantly because decision rights are unclear. The AI can draft the answer, but the agents do not know which answer they are authorized to give.

A third deploys an AI HR assistant to reduce support tickets. Managers now choose among four channels, the chatbot, SharePoint, email to their HRBP, or a ticket, and get conflicting answers from all of them. The assistant added a fourth door to an already confusing hallway, so people fall back on whichever channel they trust.

In each case the AI performed as designed and the deployment still failed. The surrounding workflow was the obstacle, and no one had mapped it before go-live.

The question that would have helped in each case is a simple one: what reality are we dropping this into?

What does the workflow look like today, as workers live it rather than as it was designed? Where does time go, where do people get stuck, what are they working around, and what will the AI touch that no one has accounted for?

Most organizations do not answer that question rigorously before deployment. They build the use case, scope the tool, and stand up the training, but they never get a quantified, worker-informed picture of the workflow they are about to change.

The cost of skipping that step shows up in weak adoption, in flat productivity numbers six months after launch, or in an AI blamed for a problem that predated it.

The organizations that do answer it treat it as a readiness gate: a short set of questions every AI deployment must answer before scale. What are the conditions the AI will land in, and which owners can change them? What is the baseline performance and time cost of the workflow today? What will prove the workflow improved after deployment? And which owner is accountable for each fix the data surfaces?

The gate does not stop deployment. It disciplines it. If the workflow has not been measured, baseline it before scale. A company that does this once improves one deployment. A company that does it repeatedly builds the practice into how it deploys AI, and accumulates a record of how work changed under each deployment that no retrofit can recreate.

Asking the question does not have to slow the program down. With the right approach, a quantified, statistically robust picture of a workflow takes under three weeks to build. Against the cost of a failed deployment, that is a small investment.

By now you have probably seen the statistic: the MIT report finding that 95% of generativeAI pilots fail to deliver a return.

The figure invites reasonable pushback. Definitions of failure vary and sample sizes matter. But the underlying finding holds up, because the reasons for failure keep pointing to the same place.

From the research: “Most fail due to brittle workflows, lack of contextual learning, and misalignment with day-to-day operations.” And: “What’s really holding it back is that most AI tools don’t learn and don’t integrate well into workflows.”

The researchers did not blame the models or the use cases. They blamed integration.The tools do not fit into workflows.

This tracks with what we observe directly in client organizations. Leaders are deploying powerful tools into employee workflows they do not fully understand. The tool works, but the work around it does not change. The handoffs, exception handling, unclear decision rights, and tool-switching carry as much friction as before, and sometimes more.

When an AI rollout underperforms, the instinct is to add change management: more training, more communication, a champions program. These help at the margins, but they do not fix a workflow that was broken before the AI arrived.

So what does fix it?

Workflow redesign. Process redesign addresses the organization’s intended flow. Workflow redesign looks through the worker’s lens: what changes in how they work now that AI is part of the picture, where friction shifts and concentrates, and what the AI was supposed to make easier that has instead become harder.

The organizations seeing real productivity gains answer these questions before deployment. They go in with a clear picture of the work they are dropping AI into. They know where effort concentrates and why, and they can identify in advance where the AI will create friction as well as reduce it.

When something underperforms, they do not have to guess why. They have the data to diagnose quickly and correct course.

The high failure rate in AI pilots is real, but it is not inevitable. The path around it runs through clarity about the human work surrounding the AI, and that clarity starts with taking workflow intelligence seriously before deployment.

Right now, inmost large organizations, three teams are simultaneously redesigning the same employee workflows.

The AI team is shipping copilots and agents. HR is redesigning roles and skill mixes. Ops is changing the operating model: sites, channels, hand offs. These teams are not siloed by choice. They know parallel work is underway and they want to align, but they struggle to.

What they lack is a shared picture of how work gets done today, and of how each team’s changes land in the workflow of the person trying to deliver an outcome. Alignment conversations stall, and the work progresses anyway, because it has to.

Each team falls back on what it can measure. The AI team tracks adoption of its copilot. HR tracks role coverage and skill maps. Ops tracks channel mix and handle time. That leaves three dashboards and three definitions of better, none of them centered on the employee workflow where the friction of all these concurrent changes lands.

The problem is that friction does not show up in any one team’s metrics. It shows up later, as a customer CSAT dip, an unused AI tool, or an attrition spike in a critical role. By that point nobody can quite explain why, because each team’s data looked fine.

This is one of the most under appreciated costs of modern enterprise transformation. Organizations are redesigning work faster than ever, and without a shared picture of the workflow, that speed compounds the problem: three teams changing the same work at once, with no way to see the combined effect.

Governance committees and alignment processes help, but the more fundamental fix is a shared data layer: a common view of how specific employee workflows are actually performing, from the perspective of the workers running them. With it, all three teams can debate from the same evidence and measure the interventions against it.

When that shared picture exists, alignment shifts from a negotiation about whose data is right to a conversation about what to do next.

The workflow evolution now underway will run for a decade or more. Organizations that get the visibility infrastructure right early will hold an advantage over those that wait.

‍

Most organizations have a structural gap in how they understand work.

They have process maps, org charts, job descriptions, system logs, and KPI dashboards. What they lack is an accurate picture of how work actually unfolds for the people doing it: the workarounds, the rework, the undocumented exceptions, the informal coordination that happens in a chat thread or a hallway, and the judgment calls that no process map captures.

Leaders know this anecdotally. A business leader will tell you, “I know my people are frustrated. I know something’s slowing them down.” But they cannot tell you where the problem is largest or most damaging, or what to prioritize in a sea of friction.

So, with this incomplete picture, cross-functional leaders do what they can. Tech leaders improve digital tools and deploy AI wherever they can find a use case. Business leaders revamp the operating model. HR leaders redefine roles and roll out change readiness surveys. Everyone is changing the same work at once, without a shared and accurate view of what is happening inside it.

This has always been a problem, and it is becoming a more expensive one.

AI is reshaping workflows faster than most organizations can measure. New tools and agents go live every month, and roles get redesigned around them. As work changes faster, the gap between what leaders think is happening and what workers experience keeps growing.

The current methods are not keeping pace. Process mining reveals system-level patterns but misses the out-of-system reality, and it does not explain why friction exists.Engagement surveys capture frustration but not root cause. Interviews and focus groups provide depth, but they are slow and expensive, and they cannot run continuously across an enterprise.

The result is that organizations are making major decisions about AI, workflow redesign, and investment based on a partial picture. Work is a black box.

The way out requires a different kind of data: quantitative, workflow-level, gathered from the people doing the work, and collected continuously rather than once at the start of a transformation. Workflows are not static. They change every time anew tool ships or a process changes.

Organizations that build this capability, increasingly called workflow intelligence, gain something most of their peers lack. They can see the work clearly enough to prioritize, and they can verify whether the changes they make are improving anything.

The goal is to move from guessing to knowing. As AI accelerates the pace of change in work, that move is worth more than it has ever been.

‍

‍