Enterprise v2
What if the smartest play for a legacy enterprise isn’t to iterate with AI — but to rebuild from scratch?
The thought experiment
The most likely path for widespread enterprise adoption of AI is slow and iterative. But other companies might take a different approach. What I’m calling “Enterprise v2”.
Big companies like to test things in small, safe environments before they roll them out across the organisation. Microsoft Copilot is a good example. Its abilities are limited, not game changing. It’s controlled in such a way that people are somewhat protected from themselves and unlikely to bring down the company with it.
The latest agentic tools from Anthropic, OpenAI and others are a different story. These can really do damage if not used well — data leaks, system failures, huge swaths of mission critical infrastructure being deleted... Companies will eventually deploy them in limited scope with limited functionality, or they’ll toggle such restrictive enterprise-level controls that they won’t be much better than Copilot. Over time this will spread across the company as teams become more capable, and more problems will be solved with AI. Slow and steady.
Block is already going further. Jack Dorsey cut 4,000 employees — nearly 40% of the company — and published a piece with Sequoia’s Roelof Botha arguing that corporate hierarchy exists to route information, and AI now does it better. He’s not adding AI to the existing org. He’s removing the org and replacing it with AI. But this kind of transformation is like rebuilding the aeroplane while flying and not all executives will have the gumption to try it.
The rewrite strategy
Instead of iterating slowly or rapidly transforming, some companies will just start again from scratch. Think companies with epic Frankenstein legacy systems — like banks. Many banks have hugely complicated legacy infrastructure, some of which still runs code written in the 1970s. COBOL, a coding language developed in the late 1950s, still processes around 95% of US ATM transactions. When Anthropic recently demonstrated AI tools capable of automating COBOL modernisation, IBM’s stock dropped 13% in a single day, wiping out over $30 billion in market value. That’s how deep the dependency runs.
For these companies relying on extensive legacy technological processes built over many years, it may make more sense to just build a new company. Enterprise v2.
Why it’s usually failed
This has been tried before, and the track record is… mixed. Netscape tried to rewrite their browser from scratch in the late 90s. The rewrite took years, the product was still half-baked on release, and the company essentially destroyed itself in the process. Microsoft spent 12 years and up to 100 developers trying to build a new OS from scratch (Midori) before quietly shelving the whole thing.
But there are successes too. In 2012, Basecamp rewrote their product from scratch, launched it alongside the existing version, and doubled new signups. Microsoft built VS Code as a greenfield project — no backward compatibility promises — and it became the most popular code editor in the world. The pattern in the successes is consistent: they ran the new thing alongside the old, treated it as a fundamentally new product rather than a replica, and didn’t try to do a hard cutover overnight.
Why AI changes the math
With AI, this is likely to become feasible for entire companies, not just products. AI can process enormous volumes of information in ways that weren’t possible before — mapping org structures, workflows, activity drivers, commercial environments. Feed all of an enterprise’s data into an AI system, have it iteratively work through the data to understand how the organisation actually operates, and then redesign the org, systems and processes for the new world. Once the plan is documented, structured, and agreed to by the executive team and board, the AI plans the implementation. Systems build, testing, go-live. Legacy bank becomes AI-first neo bank.
The biggest obstacle to this kind of transformation in the past has been the change implementation and cost of running two organisations in parallel. It’s nearly impossible to turn a 10,000-person legacy bank into a 5,000-person neo bank overnight. And building the 5,000-person neo bank alongside the legacy bank is prohibitively expensive. But if a bank can be run by a few hundred people, armed with AI… maybe it’s possible? The successful rewrites suggest the “run both in parallel” model works. AI might just make the parallel version cheap enough to actually pull off.
The missing piece
Of course all of this will only come to pass (without a violent revolution, of course) if we build Government v2 and Tax System v2 as well, such that the rewards of these kinds of gains in productivity are better shared. Here’s hoping...