Every enterprise has one: the system nobody fully understands, that everything depends on, whose original authors are long gone. The instinct to rewrite it is understandable and almost always wrong — big-bang rewrites are where budgets and careers go to die. The strangler-fig pattern, named for the tree that grows around its host, offers the alternative: wrap the legacy core, redirect capabilities one at a time, and let the old system keep running until, one day, nothing routes to it anymore.
The mechanics, concretely
- Facade first — put an API layer in front of the legacy system so every consumer talks to the facade, not the monolith. Nothing changes behaviorally; you've gained a control point.
- Slice by capability — pick one bounded function (invoice validation, address lookup, inventory checks) and rebuild it as a modern service behind the facade.
- Parallel-run — route traffic to both old and new implementations, compare outputs on real production data, and promote the new one only when the diff report is clean.
- Rollback armed — every cutover has a rehearsed one-step reversal. Confidence comes from the rollback you've tested, not the migration you hope works.
Where AI slots in
The facade seams are exactly where AI capabilities enter without destabilizing the core. Document intake that used to be manual keying becomes an extraction pipeline feeding the same legacy interface. Routing decisions hardcoded in COBOL-era rules get shadowed by a model that must agree with the rules for months before it's allowed to disagree. The legacy system never knows anything changed — it just receives cleaner inputs, faster.
Shadow mode is the discipline that makes AI-in-legacy safe: the model runs on real traffic, its outputs logged but unused, until the comparison data — not the demo — earns it authority.
Why this beats the rewrite
Each slice ships value independently and funds organizational patience for the next. Risk stays bounded: the worst case for any increment is reverting one facade route, not resurrecting a decommissioned system. And your team learns the modern stack on low-stakes slices before touching the terrifying ones. The rewrite promises a better future in eighteen months; the strangler-fig delivers a slightly better present every few weeks — and compounds.
The prerequisite nobody skips successfully
You cannot wrap what you haven't mapped. The first weeks are always archaeology: tracing data flows, finding the undocumented integrations (there are always undocumented integrations), and ranking dependencies by blast radius. It's unglamorous, and it's the difference between a controlled migration and a sequence of surprises.