Every vendor pitch, board deck, and strategy offsite right now revolves around one intoxicating promise: AI is going to eradicate friction from your enterprise. We’re going to automate the bottlenecks, gut the bureaucracy, and execute at machine speed.
The problem isn’t that AI can’t do these things. The problem is that it absolutely can.
When you give a company the power to instantly vaporize operational friction, the most dangerous thing it can do is start blasting without asking what that friction was actually doing in the first place.
Enter G.K. Chesterton’s 1929 book The Thing, specifically a chapter titled “The Drift from Domesticity,” which introduced a principle that has quietly become the patron saint of systems engineering: Chesterton’s Fence.
The premise is simple. You’re walking down a road and find a fence blocking your path. You don’t see any reason for it to be there. The instinct of the modern reformer (or the modern AI project lead) is to tear it down.
Chesterton’s response?
“If you don’t see the use of it, I certainly won’t let you clear it away. Go away and think.”
His point wasn’t that the fence must stay. It was that you had no business removing it until you understood why it had been built. Once you understood its purpose, you could make an informed decision about whether that purpose still mattered.
In the rush to deploy enterprise AI, we are ripping out fences before we know why they were built. We’re mistaking structural resilience for bureaucracy.
In the age of autonomous agents, friction isn’t necessarily a bug. Sometimes, it’s the most important feature in the system.
The “Useless” Rule May Be a Hidden Circuit Breaker
Friction in a business process can serve purposes that never appear in the process documentation. Sometimes those purposes were intentional. Sometimes they emerged entirely by accident.
Take standard procurement. A manager needs a software license, but the ticket requires three approvals and sits in a queue for four days. An AI vendor swoops in: “Our agent validates the budget, checks the general ledger, and issues the PO in seconds.”
The CIO signs the check. Friction gone. Process optimized.
But what if those four days were doing something nobody had thought to measure?
Perhaps a procurement officer would have noticed that the vendor was under regulatory scrutiny. Perhaps somebody would have remembered that marketing bought a nearly identical tool last week. Perhaps an experienced reviewer would have questioned why a seemingly routine software purchase required access to sensitive customer data.
The process wasn’t just doing math. It was creating opportunities for people to notice things that weren’t part of the formal decision criteria.
Of course, sometimes the queue really is just a queue. Nobody reviews anything meaningful, and the ticket spends four days collecting digital dust.
That’s precisely the point.
You cannot tell which kind of friction you’re dealing with simply by measuring how long it takes.
Here is the twist: sometimes bureaucracy wasn’t put there by a wise architect intentionally trying to ensure safety. Sometimes a process is just clunky, annoying, and poorly designed. Yet, entirely by accident, that irritating bottleneck produces unintended positive consequences.
It creates a natural buffer. It slows the propagation of mistakes. It gives people a chance to reconsider decisions before those decisions become expensive or irreversible.
Whether a rule was designed to protect the system or merely happens to protect it, removing that rule can eliminate safeguards you never realized you had.
The fact that nobody intended a process to provide protection doesn’t mean it isn’t providing any.
Amplifying the Blast Radius
AI prototypes tend to showcase the “Happy Path”—clean data, perfect prompts, flawless execution. In a testing environment, friction just feels like a lack of imagination.
But the real world isn’t a testing environment.
The real world is where messy reality hits rigid automation, where exceptions are more interesting than rules, and where the most consequential failures often emerge from conditions nobody thought to include in the test suite.
There is a massive gap between intelligence and authority.
Intelligence is analyzing data and reasoning through a problem. Authority is changing core systems, rewriting security rules, or spending real company money.
By optimizing strictly for speed, we risk giving AI both, assuming that because an agent produces convincing answers, it can be trusted to act on them.
But a system’s ability to recommend an action tells us very little about whether it should have permission to execute that action.
A ticket sitting in a human’s queue for four days might be terribly inefficient. But a meaningful review can give a seasoned operator time to cross-reference recent events, challenge assumptions, and notice consequences that weren’t included in the original request.
An autonomous agent might execute the same flawed decision across an entire operation before the error even registers on a dashboard.
And unlike a human employee making one questionable decision at a time, a coordinated network of agents can propagate a mistake through systems, departments, and even organizations simultaneously.
Applying an AI agent to a broken or misunderstood business process doesn’t fix it. It can simply help you make terrible mistakes at the speed of light.
By tearing down the fence of human oversight, we haven’t just cleared a bottleneck.
We may have amplified the blast radius.
Testing Industry Ambitions Against the Fence
The temptation to strip away friction is strongest in complex, high-stakes environments. These are also the places where an apparently inefficient human process may be performing several functions simultaneously.
Consider healthcare.
The dream is to have AI handle patient intake calls, match symptoms against clinical guidelines, and route patients toward appropriate care.
There is genuine value in that ambition. Better triage, faster responses, and easier access to care are worthwhile goals.
But nurses aren’t just walking decision trees.
Patients are often terrified, embarrassed, confused, or simply terrible at articulating what’s actually wrong. A clinical professional may recognize that a routine complaint conceals something more serious, or that a patient is withholding information because they don’t yet feel comfortable sharing it.
An AI system might ask clinically appropriate questions. Some patients may even find it easier to disclose sensitive information to a machine. But that doesn’t eliminate the importance of relationships, context, and professional responsibility.
The risk isn’t that AI participates in patient intake. It’s that we mistake successful information collection for the entirety of patient care.
Or think about the supply chain.
If a supplier delays a shipment, an autonomous agent could scour the globe, evaluate alternatives, negotiate terms, and execute a replacement contract before anyone in procurement has finished their morning coffee.
Sounds fantastic.
But procurement isn’t just a mathematical exercise in matching specifications against prices.
Experienced procurement professionals understand relationships, supplier histories, contractual ambiguity, geopolitical risk, and the difference between what a vendor promises and what it can reliably deliver.
Sometimes the crucial question isn’t answered by the available data. It emerges when somebody asks why a supplier’s explanation has changed, requests evidence of an uncomfortable claim, or remembers a problem from three years ago that never made it into the vendor database.
AI can help analyze these risks. It may even identify patterns that humans miss. But optimization against known criteria is not the same as knowing whether the criteria capture the actual risk.
Picture the insurance industry.
The ambition is to build an app where customers upload photos of property damage, an AI assesses the image, verifies the policy, and pays out in seconds.
For simple, well-understood claims, that could be a tremendous improvement.
But when a human adjuster visits a burned-down kitchen, they’re not simply assessing the replacement cost of cabinets. They’re dealing with a person whose home may have been destroyed.
They’re evaluating evidence, investigating inconsistencies, understanding circumstances, and explaining decisions that can have enormous consequences for somebody’s life.
That doesn’t mean human adjusters possess some magical ability to detect fraud through intuition or body language. They don’t. Human judgment is fallible, and suspicion is not evidence.
It means the claims process involves more than calculating damage. It involves investigation, discretion, communication, fairness, and the ability to challenge an initial conclusion.
When we automate the transaction, we need to be certain we haven’t automated away those responsibilities.
In finance, AI can monitor millions of transactions and identify patterns of suspicious behavior at a scale no human team could match.
That’s progress.
But money laundering is a human behavior problem as well as a data problem. Patterns evolve. Criminals adapt. Legitimate transactions can look suspicious, and suspicious transactions can look perfectly ordinary.
A human analyst investigating an anomaly may understand local economic circumstances, business relationships, or unusual but legitimate behavior that isn’t represented in the model.
The important question isn’t whether humans or machines are better at detecting suspicious transactions.
It’s whether the architecture combines their strengths while preserving the ability to investigate, challenge, and correct decisions.
Finally, consider human resources.
An AI chatbot could handle initial employee disputes, summarize complaints, evaluate performance dashboards, and draft performance improvement plans without waiting for a meeting with HR.
For routine administrative work, that might be exactly what employees need.
But workplace disputes aren’t simply disagreements about policy.
They involve power, fear, personal histories, unspoken tensions, and people who may be taking considerable risks just by asking for help.
A system can accurately quote an employee handbook while completely misunderstanding the human situation in which that handbook is being applied.
An experienced HR professional may recognize when a formal disciplinary process is inappropriate, when a conversation needs to happen privately, or when the apparent problem is merely a symptom of something deeper.
Humans can mishandle these situations too. That’s not an argument against automation. It’s an argument against confusing procedural correctness with organizational wisdom.
Across all these industries, the pattern is the same.
We identify an expensive human interaction, define it as friction, and design a system that can bypass it.
Only afterward do we discover that the interaction was doing more than the transaction we measured.
The Bottleneck Worth Keeping
The real challenge isn’t deciding whether AI should replace human work.
It is deciding where human judgment creates value, where it merely introduces delay, and where an automated system should be allowed to act independently.
Consider software engineering.
An AI agent can generate a database migration, modify an authentication service, or rewrite a deployment configuration with extraordinary speed.
It may produce excellent code.
But the senior engineer who insists on reviewing the migration before it touches production isn’t necessarily an obstacle to productivity.
They’re asking questions that don’t always appear in the specification.
What happens if we need to roll this back? What assumptions are we making about existing data? What other systems depend on this behavior? What happens when this change interacts with a failure that isn’t covered by our tests?
The engineer isn’t valuable simply because they’re human. They’re valuable because they understand consequences, dependencies, and the history of decisions that shaped the system.
That senior engineer is the friction.
And their friction may be the most valuable asset the company has.
The goal should be to automate away the tedious work that prevents experts from exercising judgment—not to automate away the judgment itself.
Sometimes the right architecture removes ten approval steps and strengthens the eleventh.
Sometimes it lets agents make thousands of routine decisions while preserving human authority over the handful that could cause irreversible harm.
Sometimes it replaces a four-day queue with an automated decision in seconds.
And sometimes it deliberately introduces a pause.
The point isn’t that every process needs a human. It’s that every consequential decision needs an appropriate mechanism for challenge, correction, and accountability.
Speed is a capability.
Judgment is a responsibility.
Confusing the two is how organizations turn technological progress into institutional fragility.
How to Approach the Fence
The lesson here isn’t to freeze your company in amber.
Chesterton’s Fence is not an excuse to hoard bloated bureaucracy or defend legacy systems just because they’re old.
Some fences are obsolete. Some were badly designed. Some protect entrenched interests rather than the people they’re supposed to serve.
And some should have been torn down decades ago.
The lesson is that understanding has to precede optimization.
Before you replace a human workflow with an autonomous agent, you should understand what that workflow actually accomplishes—not just what the process diagram says it accomplishes.
What decisions are people making that never get recorded? What mistakes are being caught informally? What relationships are being maintained? What risks are being absorbed by delays, reviews, and opportunities to reconsider?
And if those functions matter, how will the new system preserve them?
Perhaps AI can perform them better. Perhaps a redesigned process can eliminate unnecessary approvals while providing stronger safeguards. Perhaps the system can become both faster and safer.
That’s the outcome we should be pursuing.
But you have to demonstrate that you’ve accounted for the consequences, not simply that you’ve accelerated the transaction.
Stop asking, “How do we use AI to make this faster?”
Start asking, “What disaster was this slow process preventing, what human connection was it preserving, and how will we know if we’ve lost either?”
If you can answer those questions, and you can demonstrate that your new AI architecture accounts for the risks, then you’ve earned the right to tear down the fence.
Until then, respect the friction.
It might be the only thing keeping the roof from caving in.
Sources and further reading
- G. K. Chesterton, The Thing (1929), “The Drift from Domesticity”.
- C. F. Bond Jr. and B. M. DePaulo, “Accuracy of deception judgments” (2006), a meta-analysis of human lie detection.