Everyone Says AI Kills Apps. The Opposite Is Closer to True.

Everyone says AI turns everyone into a tool-builder — that soon every worker just prompts their way to bespoke software, apps as we know them are dead, and enterprise software vendors are toast. Benedict Evans published a clear-eyed rebuttal this week. He’s right. And the implications run deeper than most people are following.

The premise fails at the first empirical checkpoint: most people don’t think in tool-builder terms. They think in task terms. A junior analyst doesn’t want to build a pipeline. She wants her pipeline built. The model can do that — but someone still has to specify what the pipeline should do, catch when it does the wrong thing, and route its outputs into a workflow that other humans depend on. That’s not app-building. That’s still knowledge work, just with a smarter assistant.

Here’s the more important structural point. Software in companies doesn’t exist because individuals wanted personal tools. It exists because organizations needed coordination mechanisms — shared state, audit trails, permission structures, integration contracts between teams. A model-generated app that lives in one person’s context window solves nothing at the organizational layer. The org still needs a system of record. The system of record still needs a vendor, a contract, and an IT owner.

So the “apps are dead” thesis confuses the individual layer with the institutional layer. They are genuinely different markets with different buying motions and different switching costs. AI compresses the individual layer. It hasn’t touched the institutional layer — and might entrench it, because now the expensive thing isn’t the software itself, it’s the integration layer and the governance structure around it.

The companies that should be nervous are the ones whose value proposition was “we make it easier to do the thing” rather than “we are the authoritative system for the thing.” Workflow tools, lightweight productivity apps, form-based internal tools — those are genuinely under pressure. Salesforce is not. SAP is not. The hospital EMR is not. Not because those systems are good. They’re not. But because ripping them out requires an organizational decision that no single LLM prompt can make.

Evans’ framing is specifically useful on transformation. He argues AI doesn’t automatically change how companies work — it changes what’s possible for individuals, and whether that compounds into organizational change depends entirely on whether leadership makes deliberate decisions to restructure around the new capability. That’s historically true. Spreadsheets made individual analysts dramatically more productive in the early 1980s. It took fifteen years for that to show up as meaningful changes to how finance departments were organized. The technology moved fast; the institutional adaptation moved slow.

The implication: the AI productivity gains that are being measured right now are mostly individual-layer gains. The organizational-layer gains are still in front of us — and they require management decisions, not model releases. That’s a very different timeline than the one most AI coverage assumes.

The companies best positioned for this transition are the ones that can do three things simultaneously: sell an individual-layer product that gets adopted bottom-up, demonstrate enough organizational value to get into the institutional procurement conversation, and hold the integration moat long enough to be the system of record when the restructuring happens. That’s a hard combination. Most companies can do one. Very few can do all three.

The dominant prediction in AI coverage right now is: models get smarter, apps get obsolete, incumbents lose. The more defensible prediction is: models get smarter, individual productivity rises, organizational structures change slowly, and the companies that own the authoritative data and workflow layers capture most of the value anyway — just with LLMs under the hood. The winners look a lot like the current winners, moving faster on a better substrate.

That’s less exciting. It’s probably more accurate.