AI's Real Business Value Is Optimizing Friction

A meeting starts on time. Forty minutes later, the decision is still out of reach because the group is rebuilding context the decision needed before the meeting began. The budget does not capture those forty minutes. The project plan treats them as if they never happened. Everyone feels the drag.

For years I’ve thought about business decisions as a balance among revenue, cost, and risk. That frame is right as far as it goes. Revenue says why an opportunity is worth pursuing. Cost says what it will take. Risk says what could go wrong. Lately I’ve been convinced the frame is missing a fourth thing: friction.

Friction is easy to confuse with cost, because both make work more expensive. The difference is where they surface. Cost surfaces in the budget. Friction surfaces in the path the work has to travel to get anything done.

Friction appears wherever work should move and doesn’t. The answer exists somewhere, just not where the person who needs it can reach it. The customer has already explained the problem, but the company remembers the transaction better than the situation. The evidence is sitting in a system, and someone still has to rebuild it by hand before the review will accept it.

Some of that earns its place. Review slows the work because judgment belongs there. A control is awkward because the evidence has to hold up. Other friction has a less defensible origin. Systems were never connected. Definitions drifted. Decision rights moved, but the committee stayed. A workaround solved a problem once and then became the process.

That mix has always been partly visible, but hard to act on, because it lives in language, handoffs, search, judgment, and local knowledge. It has been too scattered to automate cleanly and too ordinary to earn its own line in the business case. AI changes the discussion because it can operate in many of the places where the drag collects, from search and drafting to comparison, handoff, evidence, and explanation. Used well, it gives leaders a way to ask which drag protects the business, which drag slows it down, and which drag should be redesigned.

Where the business case goes soft

AI business cases run through the same categories as any other investment decision: revenue, cost, and risk. Those categories are good at getting a decision approved. They are worse at showing what has to happen after approval. The cost case misses the labor buried in daily work. The risk case misses the control that passes only because someone repairs it by hand every week. The revenue case assumes a customer journey that looks clean on the slide and breaks the moment it meets execution.

Friction tends to evaporate as a business case moves upward. By the time an idea reaches a steering committee, the drag has been translated into timeline, staffing, a status color, or a few anecdotes from the field. The people doing the work know more than the business case can hold. They know which system can’t be trusted, which handoff will come back, which approval is mostly ceremonial, and which meeting will lose its first half-hour to catch-up before anyone can decide.

The business case treats the outcome as a fact. In practice, the work has to travel through missing context, mismatched definitions, side channels, manual fixes, and the people who know how to make the official path function.

AI changes the work around the job

The fear around AI usually starts with the job title. Customer service representative. Developer. Analyst. Underwriter. Architect. The question is whether the role is exposed, and how much of it a model can take.

That is a fair concern, and it oversimplifies how work happens. Inside a business, a job is rarely a clean stack of tasks. It is a set of responsibilities moving through systems, policies, customer history, exceptions, meetings, handoffs, and habits that never made it into the documented procedure. A task can look easy to automate from a distance while the work around it still leans on context and judgment.

AI tends to change the work before it replaces the role.

A support agent no longer opens three knowledge bases and asks the person at the next desk which answer is current. A control owner can start from an assembled comparison of policy language, evidence, and prior review comments. A developer can start from a working test case instead of a blank file. A new hire can see the pattern sooner, because the system surfaces examples that used to live only in the heads of people who had been there ten years.

None of this means AI owns the outcome. A person is still accountable for the answer, the decision, the code, the exception, the customer. What changes is the starting point. The first usable version arrives sooner, the search burden is lower, translation across domains takes less effort, and the work begins closer to something a person can actually judge.

Customer support gives a concrete version of the pattern. Erik Brynjolfsson, Danielle Li, and Lindsey Raymond found that generative AI assistance raised productivity by 15 percent, measured by issues resolved per hour, with the largest gains going to less experienced agents. The gain came from changing the route through the work. Newer agents reached usable answers faster, carrying more of the shortcuts and exceptions that experienced workers had learned by repetition. The job didn’t disappear. The route through the job changed.

The same pattern is surfacing across software, operations, compliance, finance, architecture, product, and sales. AI drafts the first version, prepares the background, compares artifacts, finds the inconsistency, explains the system, assembles the evidence. The value isn’t that a task got automated. It’s that the drag around the task changed.

That should change the management conversation. When the question is replacement, the analysis collapses into headcount math. When the question is friction, it moves toward operating-model design: where the work stalls, where expertise arrives too late, where people spend their effort getting ready to do the real work, and where AI could shorten the path without weakening judgment, accountability, or trust.

The old theory underneath the new tool

Friction belongs in the AI conversation, and the idea isn’t new.

Ronald Coase argued that firms exist partly because transactions are costly, not only in dollars but in the effort of finding information, negotiating terms, coordinating work, monitoring performance, and confirming that commitments were met. A firm pulls an activity inside its own walls when that is easier than coordinating it through the market.

Coase was explaining why firms exist. For the AI discussion, the same idea points somewhere more immediate: the friction inside the work itself. As AI lowers the cost of searching, translating, monitoring, and verifying, the texture of work starts to shift. People spend less time hunting for the answer. Less meaning is lost as work crosses business, technical, legal, and operational lines. Some coordination can move closer to the edge, because the organization has better ways to see what happened and check whether commitments held. Other work stays inside tighter boundaries, because it leans too heavily on trust, context, and accountability to be treated as a simple coordination problem.

The practical strategy work begins where the work catches.

A friction map traces the places where work loses speed, context, or clarity. It follows a value stream from the original request through the information needed to understand it, the decision, the handoff, the execution, the review, and the result, marking every point where the work slows, loops back, or changes shape.

That is where AI strategy should start, before the conversation turns into platforms, models, and vendors.

The dangerous kind of efficiency

Once friction is visible, it is tempting to treat all of it as waste. A review looks too slow. A handoff looks unnecessary. A human decision looks like a delay the model could delete.

The reaction makes sense. It is also where the argument goes wrong.

Some friction is doing real work. Legal review feels slow, but exposure has to be understood before the business moves. Control testing feels awkward, but the evidence has to survive scrutiny. Audit trails, segregation of duties, model validation, and human approval all add drag exactly where accountability matters.

The trouble starts when inherited drag and necessary review look identical from the outside. A checkpoint might exist because the decision is consequential. It might also exist because two systems never connected, a past failure scared everyone, or no one has revisited the approval path in years. Given enough time, both kinds acquire the same institutional confidence, and the same reluctance to touch them.

AI moves through both at the same speed. Before treating speed as the goal, leaders have to know what the friction is doing. That usually means sorting it into three kinds: the waste that can come out, the protection that has to stay, and the friction worth designing in on purpose, where failure would matter.

Waste friction slows the work without improving the result. It comes from broken handoffs, duplicated controls, unclear ownership, disconnected systems, and steps that have outlived their reason. AI can often remove, automate, or simplify it.

Protective friction slows the work because something real is at stake. The review exists because judgment matters; the control is awkward because the evidence has to hold up. AI can support that friction. It shouldn’t erase it.

Designed friction puts evidence, escalation, review, or human judgment at the points where a failure would carry real consequences.

The goal is an enterprise where friction has a job. A business with no resistance anywhere may feel fast, but speed without judgment is just another kind of exposure.

When speed just moves the mess

AI can make the visible part of a process faster while leaving the real problem untouched. A policy summary arrives in seconds and the hard part remains, because the exception still has to be understood in context. A draft looks finished while carrying a bad premise forward. A workflow speeds up only to land downstream on someone who now has to check it, repair the gap, or explain to a customer what happened.

Most of those failures were already in the process. Speeding up the first step doesn’t retire them. Verification, exception handling, customer explanation, control evidence, audit review, and cleanup don’t disappear because the opening move got easier. They arrive sooner, and sometimes with less warning.

The jagged-frontier research helps explain why this fools people. AI can do well on one knowledge task and badly on another that looks almost identical from the outside, and to the user both feel easier. The answer comes faster, the draft reads cleanly, the workflow appears to move. Ease is not the same as good execution. The faster answer can be wrong faster. The cleaner draft can carry a weaker assumption. The smoother workflow can leave judgment, accountability, or evidence behind.

The real test is whether AI improved the path the work has to travel, or only made one step on it feel easier.

How to use the frame

The place to start is rarely the job title. Titles invite arguments about exposure and headcount. The better unit of analysis is the path the work travels.

Take a real flow and follow it past the process document. The friction that matters usually shows up after the formal step has been described and before the outcome actually arrives. A request gets routed, and its context goes missing. A product idea gets approved, and the handoffs stay vague. A control has an owner, and the evidence still has to be rebuilt from scratch. The official map and the real work stop matching, and the gap is where the resistance lives. That gap often turns on one analyst who knows which version of the data to trust, or an approval that is compensating for unclear decision rights.

Once you can see it, the AI conversation gets less abstract. Some drag can go, because it adds nothing. Some should stay, because it protects the result. And some exists for a good reason but has been built in a needlessly painful way. The most interesting opportunities tend to sit in that last group, the awkward middle, where the step still earns its place but the way it is performed no longer makes sense.

A control shouldn’t require a scavenger hunt for evidence. Legal review shouldn’t begin with the lawyer doing the first comparison of contract language, policy, precedent, and exception. An approval meeting shouldn’t spend its first half-hour reconstructing facts the request should have carried in with it. That’s the space where AI is most useful: cutting the preparation burden, making context travel with the work, and letting the human step spend itself on judgment instead of assembly.

What this says about jobs

The jobs question doesn’t vanish. AI will replace some work. Some people will be affected directly. Some roles will be reorganized until the old title hides a very different job underneath.

But the clean substitution story doesn’t match what is actually showing up. Anthropic’s Economic Index found observed Claude usage tilting slightly toward augmentation, 57 percent of tasks augmented against 43 percent automated. Brookings, summarizing recent research, reports that AI hasn’t yet produced broad job loss, and that adoption so far tracks with firm growth, employment growth, and innovation. That isn’t reassurance. It is a sign of a messier transition than the headlines promise.

AI changes jobs by changing the friction around them: what people search for, what they start from, what they review, what they escalate, and what they’re expected to catch. As first drafts get cheap, judgment carries more of the weight. As systems get better at producing fluent language without understanding the situation, context carries more too.

The headcount question is always tempting, because it’s easy to count. The operating-model question is harder and more worth asking: where execution is slow, where knowledge is trapped, where controls have become theater, where customers feel the company’s internal seams, and where judgment needs strengthening before speed turns into risk.

The friction question

Revenue, cost, and risk are still the necessary tests of a business case. They say why a decision is worth making, what it might cost, and what it exposes the organization to. They say very little about whether the organization can actually move.

A company can pick the right market, fund the right initiative, approve the right strategy, and still watch the value leak out through slow execution, broken handoffs, unclear ownership, brittle controls, and customer experiences that put the company’s internal seams on display. AI won’t fix that on its own. Sometimes it only helps the organization race faster down a path that should have been rebuilt years ago.

That’s why friction deserves a seat in the frame, examined with the same seriousness as the other three. What does the current path cost in delay, rework, supervision, customer effort, and lost momentum? How much of that cost is buying something worth having, such as trust, evidence, safety, or better judgment? How much is just the tax of old systems and process memory?

AI makes the question practical, because it can work close to where the drag collects. The answer won’t always be more automation. Often it’s a cleaner handoff, better evidence, clearer decision rights, or a human review moved to the one place it actually changes the outcome.

The question worth adding to revenue, cost, and risk is also the plainest: what is the drag doing, and what is it costing?

Selected References

Ronald H. Coase, “The Nature of the Firm,” Economica, 1937.

Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, “Generative AI at Work,” Quarterly Journal of Economics, 2025.

Fabrizio Dell’Acqua et al., “Navigating the Jagged Technological Frontier,” Organization Science, 2026.

Anthropic, “Introducing the Anthropic Economic Index,” 2025.

Tania Babina and Anastassia Fedyk, “The Effects of AI on Firms and Workers,” Brookings, 2025.