Why frictionless enterprises break first

All friction has a cost, but only some of it serves a useful purpose. Escalating a customer service call can lengthen a manager’s queue, but it can also protect a vulnerable customer or keep a valuable one.

Earlier robotic process automation technology tended to follow fixed rules and processes and left people to handle exceptions, so the balance between cost and purpose changed slowly and predictably. AI upsets this balance because it produces output and decisions faster, cheaper and in greater quantity. AI agents also act across multiple processes once separated by people or functional boundaries, which means an error made in one place can now spread further and faster before being detected and resolved.

Making tasks faster does not automatically make the enterprise faster or the output more valuable because removing resistance from one step can create a bigger bottleneck at the next. Professor Feng Li, Chair of Information Management at Bayes Business School, puts it in practical terms. “As producing and processing information becomes easier,” he says, “framing the question, verifying outputs, handling exceptions and taking accountable decisions become the real constraints.”³

For example, in software development, the engineering analytics firm Faros AI studied telemetry from more than 10,000 developers and found that those on teams with high AI adoption completed 21% more tasks and merged 98% more pull requests than those with low adoption.⁴ Yet the time taken to review those pull requests rose by 91%, and there was no measurable efficiency improvement at company level.

What separates value from speed is less the technology and more the context and judgment wrapped around it. “As AI takes on more of the work, the temptation is to strip out every bit of friction we can find,” says Katherine Boiciuc, EY Regional Chief Technology and Innovation Officer for Oceania. “But some of that friction is the challenge, the evidence, the pause that gives a fast decision the chance to be a good one.”

How much friction can AI remove cleanly?

To find out, EY analyzed all 18,796 tasks that the US Department of Labor's O*NET database uses to describe 923 occupations, weighted by US employment. A task could be cleanly removed only if the friction it contained was routine, easy to check and largely wasteful, requiring little judgment or creativity. Three further tests identified work that AI could support, but would need to be redesigned around human judgment, built with stronger safeguards, or have new controls added where AI could make results hard to check or produce errors threatening other systems.

Fewer than 3% of tasks passed the test for clean removal, against 24% for redesign, 27% for stronger safeguards and 23% for new controls. Many tasks passed more than one test but across all scenarios we modeled, clean removal remained between 0.4% and 5.5% of tasks and was always the smallest response. About 65% of the employment-weighted task inventory failed to meet any of the four friction-response tests. These tasks tend to involve more physical, lower-consequence and less AI-supportable work than tasks that met at least one test.

Clean removal is the narrowest response

Previous studies have examined the degree to which tasks are exposed to large language models (LLMs), but whereas exposure asks whether AI could make a task faster, clean removal asks whether the delay can be eliminated without losing any value.⁵ This is why so much work can be exposed to AI but so few of the bottlenecks removed cleanly.

Clean removal is also the smallest response in every sector, ranging from about 7% of tasks in finance and insurance to 0.3% in accommodation and food services. Even in finance and insurance, roughly a third of tasks supported by AI would need new controls. Most of the variation in the results arises from the nature of occupations and tasks themselves, many of which are common across sectors, functions and use cases, which is one reason why friction should be managed at the level of the whole value stream.

Clean removal is the smallest response in every sector

Removing congestion in one task or team can make the wider system harder to govern, more expensive to assure and slower to recover. This effect is easy to miss when AI is introduced one use case at a time. Whereas used across a whole value stream, AI can make these risks easier to see and manage. As reported by InfoQ in August 2026, Rootly, a small US-based software company, found that its old rule of keeping code changes small stopped working once AI was writing most of its code.⁶ Instead of just trying to review these changes faster, it added an agentic AI reviewer to flag to human reviewers how risky each change was, and began releasing new features in a switched-off state, so they could be introduced gradually and switched off again quickly if something went wrong.

Because clean removal is rare, leaders need more options than simply keeping or cutting friction. The right combination of challenge, assurance and expertise depends on the risk, uncertainty and impact on stakeholders, and how closely the work is linked to other systems. A single value stream may need friction removed at one stage, redesigned at another, strengthened at a third or added at a fourth. The table below sets out the four responses, when each applies and who should own it.

Four responses to friction

The four responses to friction, in action

The case studies below illustrate real-world applications of the four responses.

Picture an enterprise in 2030 whose AI agents negotiate contracts, allocate inventory and resolve customer requests across its global operations. A geopolitical shock suddenly changes the risk profile of several critical suppliers. Can the organization bring the right experts into the decision, and act on their advice, before a problem with a few suppliers becomes a system-wide failure?

The answer depends on how much agency the organization gives AI, from tools that assist people to agents that start and complete actions on their own. It also depends on how well it governs judgment and value, which means deciding where people stay involved, tracking whether AI creates value, and choosing what to scale or stop. Together, these two factors produce the four futures, each of which uses the four responses to friction differently.

Maurice Conti, futurist and EY AI Advisory Council member, expects people rather than technology to set the pace. “People think the technology is hard. And it is,” he says. “But humans are the real challenge. They’re messy, take forever to change and adapt, and each one has their own agenda. The big differentiator will not be the technology. It will be getting the humans right.”⁹

Four plausible futures shaped by AI and the governance of judgment and value

Why frictionless enterprises break first

The failure pattern behind frictionless fragility predates generative AI, as Knight Capital discovered on 1 August 2012. That morning, a faulty software deployment sent more than 4 million orders into the US stock market in 45 minutes while the firm was trying to fill hundreds of customer orders, at a cost of more than $460 million.¹⁴ Its systems had sent nearly 100 automated emails flagging an error before the market opened, but no one acted on them, and the firm lacked adequate controls to prevent the erroneous orders. Agents that act across whole value streams can reproduce this same mix of speed and tight coupling, as well as an absence of meaningful human intervention, but on a far larger scale.

Therefore, organizations that expand agentic AI faster than their governance and their capability to monitor agents and recover from mistakes may drift towards frictionless fragility. For organizations still building the controls and expertise required for greater autonomy, guardrailed augmentation may be the most credible next step.

Design around the value stream

Because so much of AI’s potential value sits between functions, according to previous EY analysis, point solutions tend to improve the performance of the most visible tasks but miss those elsewhere in the workflow. An insurance claims model may classify simple cases in seconds while exceptions wait days for an investigator, and a sales tool may generate strong recommendations while account teams are still measured on activity volume rather than deal margins.

“Leaders are missing the opportunity to drive real business value with AI,” says Kim Billeter, EY Global People Consulting Leader. “Too few are stepping back and rethinking their operating model, the design of their organization and the design of the actual work. Instead of coordinating AI into workflows with the right guardrails, many are still counting on the provision of tools to bring the value. That comes with risk and cost.”

Leaders should start with the most important outcome or growth opportunity and work backward into the value stream that delivers it, mapping the decisions, risks and accountabilities at each step. This is often where ownership breaks down, because no single leader feels able to make a call that spans several departments. Only at this level, though, can leaders see whether a gain in one team or function improves the outcome or simply relocates the bottleneck.

Get the right hands on the loop

Adding a “human in the loop” has become the standard reassurance for agentic AI, but it says nothing about whether these humans have the context, data, expertise, independence, time or authority to make good decisions. Many CEOs see a gap here, with 47% of those in the October 2026 EY-Parthenon CEO Outlook Survey saying their organization is not developing AI skills quickly enough to keep pace with innovation.

Removing the wrong resistance can widen this gap by taking away the work through which people learn. “Some apparent friction is where judgment is formed rather than merely exercised,” says Professor Li. Junior employees learn judgment by watching experts frame ambiguous problems and handle exceptions, so treating that work as waste may make decisions faster now and worse later.

Early evidence from Stanford researchers suggests that the damage could start with entry-level hiring. Employment of 22- to 25-year-olds in AI-exposed occupations is now 19% below where it would have been had it kept pace with less-exposed peers, mainly because fewer are being hired.¹⁵ Cutting entry-level roles may save money now, but in a few years, no one may be ready to supply the judgment AI systems depend on.

Professor Shannon Vallor, Chair in the Ethics of Data and Artificial Intelligence at the University of Edinburgh and EY AI Advisory Council member, sees a wider problem. "If you put people in a situation where the machine is doing all the work the mind used to do, they become habituated into behaving like mindless machines — and mindless machines aren't loyal, aren't invested in anything, and don't have the motivations you actually need in a workforce."¹⁶

Putting the right hands on the loop therefore becomes important, meaning people with the context and authority to manage an AI-enabled system. Where a human in the loop typically approves outputs one at a time, the right hands on the loop oversee how the whole system behaves and step in when judgment is required. They are needed at every level, from the executive who owns the outcome to the specialist who knows when an agent’s answer is wrong, and they come from business, technology, finance, risk, and people functions alike. Their roles should be explicit before any AI-enabled workflow is allowed to scale.

They also need clear rules that set the thresholds and triggers for handing a decision from an agent back to a person. Authority alone is not enough, Professor Vallor adds. “Even if you have the authority to overrule the machine, you need criteria, you need to know what counts as a mistake.” These rules should be visible and testable through audit trails, incident reviews and independent challenge.

Whatever the rules say, accountability for an agent’s decisions remains with the organization that deploys it. Singapore’s government made this explicit in January 2026, when it published what it describes as a first-of-its-kind governance framework for agentic AI.¹⁷ It states that the organizations that deploy agents, and the humans who oversee them, remain accountable for what those agents do, and it advises that agents should seek human approval at defined checkpoints and before specified actions.

The right hands will also be leading a different kind of team, in which managers direct the work of agents as well as people, and employees treat an agent as a colleague whose work they shape and check. “We're already seeing hybrid human-agent workforces with clients,” says Kim Billeter. “Many organizations are experimenting with them but haven't yet redesigned their structures or workflows to take full advantage of AI, or prepared their managers to lead these hybrid teams.”

Measure value and act on the evidence

In the October 2026 EY-Parthenon CEO Outlook Survey, AI-enabled tools were the most-cited driver of productivity gains, ranked among the top three by 50% of CEOs, but 23% say these gains are largely absorbed by organizational complexity, regulation, risk management or the requirements for new work. This is what we would predict, since faster tasks tend to expose the next constraint in the process.

Before leaders can manage AI investment consistently, they need a shared definition of AI economics that sets out which costs and benefits are important and over what period. Without it, an organization may have sufficient evidence to run experiments with individual use cases, but too little on which to redesign an entire operating model.

The way the economics are framed also changes how people respond to AI. “When AI is framed primarily as a way to reduce headcount, people go on guard,” says Dan Diasio, EY Global Consulting AI Leader. “Their creativity shuts down. They are less likely to help rethink how the work should change.”¹⁸

Organizations therefore need a regular decision-making cycle that measures what AI has changed and uses the evidence to decide which opportunities to fund or stop and where to extend or pull back AI’s autonomy. An AI value management practice is one way to run it because it gives leaders a common basis for deciding where to invest and what to scale.

How evidence about value informs the next decision in the cycle

Treating friction as a design choice can start with the value stream in which AI’s agency is growing fastest. The five priorities below form a checklist for this value stream and for every significant AI-enabled program that follows. They broadly track the stages of the cycle in Chapter 4, and each ends with the questions its owners should be able to answer.

1. Start with the outcome and map the value stream that delivers it

Primary owners: CEO and COO, working with leaders across the value stream.

Identify the outcome the enterprise needs to change and trace the work that delivers it from start to finish, recording the decision rights, risks and accountabilities along the way. Check whether improvements in one function create additional work or exposure in another. Name one owner for the outcome and design the operating model around the value stream instead of the functions it crosses.

Which end-to-end outcome are we trying to change?

2. Use the full range of responses to friction

Primary owners: CEO and COO, with the CRO, general counsel, CISO, and CIO or CTO.

Find where delays exist in the value stream and what each instance does, whether it wastes effort, provides assurance, builds skills, sustains relationships, or supports recovery. Then decide what to remove, redesign, strengthen and add, knowing that a single response will rarely fit a whole workflow.

Where is friction present, and what, if anything, does it do? Which response fits each source of friction?

3. Get the right hands on the loop

Primary owners: CEO, with the CFO, CHRO, CRO and CIO or CTO.

Assign responsibility for outcomes, judgment, stewardship, economics and challenge. Make sure these people can stop and change the system, and help redesign where necessary, rather than merely approve its output. Write down the conditions under which an agent must hand a decision back to a person.

Who are the right hands on the loop?

4. Make the role of judgment explicit and build the workforce to supply it

Primary owners: CHRO, with business unit leaders.

Identify the moments where better context could change the outcome of AI-enabled work and assign that judgment to people with the evidence to exercise it. Redesign roles, incentives and career paths to promote the development of judgment, particularly in teams where people work alongside agents.

What skills do these people need, and when should AI hand a decision back to them?

5. Govern value continuously

Primary owners: CFO, with the CEO.

Establish an AI value management practice or similar capability as a standing management discipline that connects investment to the full cost, business outcomes and workforce effects. Let the evidence decide whether AI’s autonomy should expand, contract or be redesigned as conditions change.

What evidence do we have that value is being realized, and what will we do next?

The authors would like to thank Shivani Ghildiyal, Assistant Director, Thematic Insights, Clients & Industries, Ernst & Young LLP, for her contributions to this article.