Most Companies Are Just Speeding Up Broken Workflows With AI

Only 21% of companies using generative AI have redesigned at least some of their workflows. That single number, buried in McKinsey's latest State of AI report, explains why most companies are not seeing the productivity gains they were promised.

The promise sounded simple. Plug AI into the work and watch output rise. The data tells a different story. McKinsey tested 25 different organizational attributes for their effect on EBIT impact from generative AI. Workflow redesign ranked at the top. It correlates more strongly with financial returns than headcount changes, model selection, training budgets, or executive sponsorship. Yet four out of five companies skip it entirely.

The 21% Doing It Differently

Among the small group of high performers, defined by McKinsey as the roughly 6% of organizations attributing 5% or more of EBIT to AI, the redesign rate jumps to 55%. That is almost three times the rate of the rest of the field. The high performers are not buying better models or hiring more data scientists. They are doing the unglamorous work of reorganizing how the work itself flows.

Meanwhile, only 39% of organizations report any enterprise-level EBIT impact from generative AI at all. The gap between "we use AI" and "AI is changing our P&L" is precisely the gap between speeding up individual tasks and rebuilding the workflows those tasks sit inside. Most companies are sitting on the wrong side of it, and the side they are on is widening into a structural disadvantage.

Why Faster Tasks Do Not Add Up to Faster Work

The math behind this is uncomfortable. Picture a ten-step process where each step takes roughly 10% of the total cycle time. Drop AI on one step and double its productivity. The overall workflow time falls by 5%. That is the entire return on a tool that was supposed to transform the operation.

This pattern shows up everywhere once you look for it. A marketing manager uses AI to turn a two-week briefing exercise into an afternoon. The creative director still has one weekly review slot. Legal still takes five days. Procurement still needs four sign-offs. The campaign brief still ships in three weeks, just with the marketing manager bored for most of them. The bottleneck moved. It did not vanish.

Research on Asana's AI productivity paradox shows the same dynamic at scale. AI succeeds on roughly 58% of straightforward tasks but on only 35% of multi-step processes. The harder the workflow, the more obvious it becomes that wins at the task level do not survive contact with the operating model. The bottleneck is rarely where the AI was pointed.

What Redesign Actually Looks Like

The companies getting this right do not start with AI. They start with the workflow. Leaf Home, a US home services business, used task mining to study 13 business areas before deciding where to apply automation. The exercise surfaced inefficiencies worth $120,000 in savings. The AI tools came in second. The diagnosis came first.

Nestle took a similar route inside SAP Concur for expense management. Rather than bolting an assistant onto the existing process, they rebuilt the flow around what AI could now do. The result was a threefold lift in the efficiency of producing expense reports, not because the AI was three times smarter than humans but because the process around it stopped wasting most of its output. The procurement steps, the approval routing, and the audit checkpoints were redrawn at the same time as the AI went in.

The pattern is consistent. Redesign means breaking a workflow into its component tasks, deciding which ones are best handled by AI, which by people, and which can be eliminated entirely. The workflow is then reconstructed around the new division of labor. It is closer to industrial engineering than to software adoption, and that is why most companies avoid it. Industrial engineering needs an owner, a budget, and a cross-functional mandate. Software adoption can be approved in a Tuesday meeting.

The AI Value Gap Is Widening, Not Closing

The cost of skipping redesign is rising fast. BCG's Widening AI Value Gap report, based on a survey of 1,250 senior executives across 25 sectors, found that AI leaders are now growing revenue twice as fast as laggards and capturing 40% more cost savings. The leaders are not slightly ahead. They are compounding.

Only 5% of companies in the BCG sample qualify as future-built for AI, the group that systematically reshapes operations rather than layering tools on top. These companies plan to spend 26% more on IT and direct up to 64% of that budget into AI. The investment is not what causes the gap. The willingness to restructure work is what causes the gap. The investment is what restructuring work happens to require.

For everyone else, the trajectory is harder to recover from than it looks. A laggard that adds more AI tools to the existing operating model produces marginal gains. A leader that has redesigned its workflows can now plug new AI capabilities into a process built to absorb them. Each new wave of AI compounds for the leader and barely registers for the laggard. The gap is not just present. It is widening every quarter.

The Org Chart Problem

There is a structural reason most companies miss this. AI tools are typically owned by IT, a function, or an innovation team. Workflow redesign requires authority across functions, since real workflows cross departments. Few companies have an existing role with both the mandate and the capacity to redraw a cross-functional process. The result is that AI gets deployed where someone has the budget to deploy it, not where the process has the most slack to remove. The most consequential decision in an AI program, what gets redesigned, ends up being made by whoever happened to have authority over the smallest unit of the work.

This problem also explains why agentic AI is unlikely to fix the situation on its own. An agent that rides existing rails will still be bottlenecked by them. The constraint is not the intelligence available at each step. The constraint is the operating model the steps live inside. Adding smarter agents to an unchanged operating model accelerates the bad pattern, it does not break it.

How to Catch the Pattern Before You Sign

For business owners deciding whether their AI investments are going to produce returns, the diagnostic is straightforward. Three questions to ask before approving the next AI tool.

First, which step of which workflow does this tool target, and what percentage of the total cycle time does that step represent today? If the answer is 10% and the tool doubles the speed of that step, the maximum benefit to the whole workflow is 5%. That is your ceiling, before integration costs and adoption friction eat into it.

Second, where does the work go after this step is faster? If the next stage cannot absorb a higher throughput, the savings stay theoretical. The bottleneck has moved, not closed. Map the absorption capacity of every downstream handoff before you sign the contract.

Third, who owns the redesign of the workflow itself, not just the tool deployment? If the answer is nobody, or the answer is the same person who owns the tool, the project will join the 79% that never reach measurable EBIT impact. A workflow without an owner is a workflow that will not be redrawn, no matter what software gets bolted onto it.

These three questions filter out the projects that look exciting on a vendor demo and produce nothing on a P&L. They also reveal where redesign would actually pay off, which is almost always upstream of where most AI conversations start.

The Reframe

The companies pulling away with AI are not the ones with better models or bigger budgets. They are the ones who treated AI as an excuse to ask a harder question. Is this workflow worth speeding up, or is it worth rebuilding? Most of the time, when a workflow has accumulated enough manual steps to make AI attractive, the honest answer is rebuild. The 21% who acted on that answer last year are now the 6% capturing meaningful financial returns. Everyone else is making a broken process run faster, and watching the bottleneck move one step downstream.