When AI makes expensive work cheap, incumbents lose their cover

Written by Russell Pereira

Jun 10, 2026
8 min

Benedict Evans argues that AI's real disruption is not the productivity gain it delivers, but the way it makes once-expensive capabilities cheap enough to dissolve the cost and complexity barriers that protected incumbents.

The efficiency case for artificial intelligence (AI) is familiar to most senior executives. Tasks become cheaper, headcount requirements shift and workflows accelerate. In financial services, banks already apply that logic to customer service, compliance monitoring and relationship-manager support, including client summaries and suggested next steps, as they respond to cost pressure, regulation and digital competition. At SuperAI, technology analyst Benedict Evans argued in his “AI Eats The World” keynote that this framing misses the bigger question. As he put it: "What tasks used to be expensive that now become free?" The issue is not only what AI saves. It is what AI dissolves when activities that once required large teams, complex systems or institutional knowledge become cheap enough for others to attack.

Financial services makes the question concrete because much of banking advantage has historically come from scale, licence, trust, data, risk management and the ability to coordinate complex processes across products and markets. Some advantages remain defensible because they depend on regulation, balance sheet strength and customer confidence. Others may rely more heavily on the cost of review, interpretation, monitoring and relationship coverage than banks would like to admit. When AI lowers those costs, it changes more than the economics of a process. It can also change who performs that process and who captures the margin.

Across Asia’s financial markets, this question has become more immediate as institutions test AI in multilingual service centres, anti-money laundering reviews, credit-document processing and relationship-manager productivity. The strategic significance emerges after the pilot stage, when a tool that improves productivity also changes the cost of serving customers, monitoring risk or making decisions at scale. An incumbent may use AI to make an existing process cheaper. A competitor may use the same capability to remove the need for that process, rebundle it into software or attack the margin that the process once protected.

Cheap information can redraw the competitive map

Evans drew on the history of the barcode to make the point concrete. When the United States (US) grocery industry deployed universal product codes in the mid-1970s, the stated purpose was to save money at checkout. Accurate inventory records then allowed retailers to carry less stock, manage far more product lines, lower prices and redesign the operating model. "Just automating that simple process changes the structure of the industry," Evans noted. Walmart showed that process automation can do more than improve an industry. It can restructure the industry and hand the restructured version to the competitor that understands the second-order consequence first.

The barcode made inventory visibility cheap enough to support a different retail model. AI may do something similar for judgement, monitoring and interpretation in financial services. Complex onboarding, credit assessment, compliance review and relationship coverage have often been expensive enough to act as cost barriers, limiting who can compete at scale. Evans framed the risk plainly: "Maybe it was that cost-based barrier to entry that protected you from competition." When AI makes an expensive process cheap or close to free, it does not merely reduce the cost of doing business. It removes part of the entry cost that kept others out.

Every previous wave of automation addressed tasks that could be described in logical steps. What a human could write down as a process, a machine could eventually execute. Evans described AI as operating on a different principle: "AI lets you automate anything where there's enough training data and where verification is scalable." In regulated industries, scalable verification means outputs can be checked repeatedly against rules, records, approvals, transaction histories, customer outcomes or risk controls. That scope reaches into judgement-heavy work, where staff interpret incomplete information rather than only follow fixed steps.

The new value lies in questions institutions could not ask before

The practical illustration Evans offered maps directly onto financial services operations. Monitoring every customer call to flag an angry interaction is one kind of automation, and many institutions already use versions of it in contact centres, complaints handling and conduct monitoring. The more significant capability is different in kind: "Listen to every single call, and then notice stuff that no person could ever have noticed before." In banking, that could mean linking complaints, repayment stress, adviser behaviour, service failures and churn signals across millions of interactions. The value does not lie only in checking whether a call followed a script. It lies in making patterns visible that no human team could detect at institutional scale.

The same logic applies to enterprise analytics. Cross-referencing customer interactions against churn data, pricing structure, product usage, credit behaviour and profitability can surface what should change in a product or relationship model. Evans extended the same logic to the enterprise, using telemetry, or product-usage data, as part of the question: "Listen to every Zoom call with every customer, then look at our telemetry, customer analytics and churn data, and tell me how we should change our pricing or our product." That is not an efficiency gain. It changes the questions management can ask. An institution may have always owned the data, but the cost of interpreting it across silos, channels and products often made the question impractical.

When a cost barrier falls, industries do not only do the same thing more cheaply. They ask different questions and design different products. Evans pointed to recorded music, where the first phase of digital change removed the cost of buying a single track. The second phase created something categorically different: all the music in the world for a flat monthly fee, a product that would have been impossible before the cost barrier fell. "No one even thought about doing that," Evans said. "The internet means that you change the question." In financial services, the equivalent is not only making existing products cheaper to deliver. It is asking which products or services become conceivable when the cost of interpretation, monitoring or personalisation approaches zero

Incumbents need to know what is genuinely hard

Evans raised a deceptively simple question that large organisations often struggle to answer: "What's the hard part?" Is the hard part writing the code, processing the transaction, delivering the product or maintaining the spreadsheet? Or is it "knowing which task to apply, and what else you should do instead?" Once AI lowers the cost of coding, reviewing, monitoring and analysis, institutions have to identify what remains genuinely scarce. The answer may lie less in activity itself and more in trust, risk judgement and balance sheet strength.

This question forces incumbents to separate activity from advantage. AI will not weaken every source of advantage equally, but it can expose which activities were protected mostly by operating cost rather than by true strategic advantage. A fintech, platform company or software provider may not need to replicate the whole institution if it can unbundle one profitable workflow, stripping it out of the wider business at much lower cost.

History shows that incumbents often misidentify which part of the value chain is truly defensible. Evans noted that newspapers "didn't think of themselves as printing companies" and record companies "didn't think of themselves as making little pieces of plastic", yet both depended on physical production and distribution more than they realised. Financial institutions should ask which of their current activities fit the same pattern. The equivalent may not be branches or paper forms alone. It may be the cost of interpreting every customer interaction, assessing small-business risk or embedding financial services into another company's workflow.

The board question is what remains defensible

The future remains hard to predict because each market absorbs technology differently, and the shape of AI’s impact will only become clear as institutions build with it. Evans noted that it was easy to describe mobile in 2015, but much harder in 2005, and the AI market now sits closer to the earlier moment. "Google didn't exist, Mark Zuckerberg was in junior high school, Jeff Bezos sold books," he said of the early internet era. Financial institutions should therefore expect some of the most important AI competitors to emerge from places that do not yet look like full competitors, including workflow software that manages specific banking tasks, compliance technology and embedded finance, where financial services are built into non-bank customer journeys.

Evans offered two tests that cut through the uncertainty. The first is whether AI breaks "some fundamental industry assumption" in the way an institution thinks about its business. His second is to act on the assumption that AI will change "out of all recognition in the next two, and three, and four years", as mobile and the internet did before it. Financial institutions cannot manage the next phase of AI only as a portfolio of pilots, tools or productivity targets. The strategic test is where cheaper interpretation changes the economics of revenue, risk and customer ownership, and where the institution still has advantages that cannot be easily replicated. The harder task is not to automate more work, but to decide which parts of the institution remain genuinely defensible when the cost barrier falls.