AI, tools and transformation
Why cheaper tools do not eliminate workflow discovery or organizational change
Benedict Evans argues that AI makes tools easier to build without resolving which problems to solve, how shared workflows should change, or who takes responsibility for them. He explains the movement between improvised spreadsheets and institutionalized software, then separates deployment choices from operational change and competitive strategy. An analytical essay using examples and historical analogies, not a measured study of enterprise AI outcomes.
Benedict Evans argues that making software cheaper to create does not automatically transform a company. People still need to discover which problems matter, design useful solutions, and coordinate changes across the organization. AI changes the available tools and when they become worthwhile; it does not remove those decisions.
Evidence basis: This is a summary of Evans's September 3, 2026 essay. Its business examples and historical analogies support his interpretation of technology adoption, not a measured comparison of enterprise AI deployments. Expectations about future applications remain forecasts.
Building a tool is only part of solving a problem
Evans separates three obstacles that easier code generation does not resolve:
- Recognizing an opportunity. A lawyer or salesperson can be excellent at their work without habitually looking for ways to redesign it. A forward-deployed engineer—a builder working with people in their workplace—may spot opportunities they miss. But some problems are hidden inside other processes, and even builders may struggle to identify them.
- Choosing what the tool should do. Seeing a frustrating task does not establish the right product or workflow. Evans points to successful software businesses preceded by failed attempts that framed the problem differently. Faster implementation helps build a proposed solution; it does not establish that the proposal is useful.
- Getting a shared process changed. An accounts-payable improvement can involve many people, departments, systems, and regulatory obligations. One employee's working tool cannot by itself authorize a change in how everyone works. Procurement and organizational decisions remain part of adoption.
These are the essay's reasons for rejecting the leap from “anyone can make a tool” to “companies no longer need their many applications.” Original essay(opens in a new tab)
Improvised workflows and dedicated software keep exchanging tasks
Evans describes a spectrum from improvised work in spreadsheets, email, and shared folders to institutionalized work in dedicated applications. Institutionalized means that the organization has agreed on a repeatable process and needs people to follow it consistently. Improvised tools handle exceptions and one-off questions that established systems do not accommodate well.
As an improvised task becomes frequent, widely shared, or important to revenue and risk, it needs audit, security, maintenance, and accountability. That creates a reason to adopt dedicated software. The movement is not one-way: a team inside a large company may return to spreadsheets because the official application is too inflexible for its particular task.
His hiring example makes the threshold concrete. A small firm hiring a few people can use email and Google Sheets; a firm recruiting thousands needs a managed process. AI might let the small firm keep using Sheets longer, generate a custom tool, or make a new specialist application attractive. Chatbots become another flexible workspace alongside existing tools, while those tools also gain AI capabilities. Evans expects tasks to move between them, rather than all software collapsing into a chatbot. Original essay(opens in a new tab)
Deployment, operational change, and business strategy are different questions
Evans argues that broad chatbot access and targeted pilots each address only part of transformation. Access reaches many employees but does not tell them how AI fits this week's work. Pilots can test specific processes, but a handful of successful trials does not explain how to change hundreds of workflows. Training and change management matter; neither substitutes for deciding what the company should do differently.
He asks companies to separate three questions:
- How do we acquire and deploy the technology? Choose among bundled products, internal development, contractors, and specialist vendors.
- How does it change operations? Identify which processes can be reorganized, rather than merely giving employees another interface.
- How does it change the business's economics and competition? Consider new competitive pressure, opportunities, or threats to the business itself.
His closing forecast goes beyond automating existing work: previous platform shifts eventually enabled products and activities that had not been possible before. He expects AI's larger consequences to include such new uses, without specifying which will succeed or when. Original essay(opens in a new tab)