← All articlesModernization

Don't rewrite the old system. Put a door in front of it.

November 4, 2026 · 3 min read

Every company has that system nobody dares touch. The smart move is rarely to replace it. It is to wrap it, and let AI work through a clean, safe entrance.

  • AI Agents
  • Data & RAG
  • Modernization
  • Education
  • Trust & Safety
  • Coaching

The system everyone is afraid of

Every company of a certain age has one. It was written ten or fifteen years ago, by people who have since left. It runs the invoicing, or the inventory, or the policy records. It has no documentation worth the name, and its screens look like they belong in another era. Nobody dares change it, because the last time somebody did, month-end closing was delayed by a week.

Now the business wants to use AI: answer customers automatically, summarise records, flag anomalies. The first reaction in most meetings is to say that the old system has to be replaced first. A budget is drawn up, a two-year programme is proposed, and then the whole initiative quietly stalls, because replacing the thing everyone depends on is the riskiest project in the building.

The old system is not the obstacle to AI. The obstacle is the assumption that AI has to live inside it.

Wrap, don't replace

There is a middle path, and it is older than AI. Instead of changing the legacy system, you put a thin, well-defined layer in front of it. The layer exposes a handful of clean operations, such as 'look up customer', 'get open invoices' or 'create order', and translates each into whatever the old system understands: a database query, a file drop, a screen it can drive.

From the outside, the system now looks modern. New tools, dashboards and AI agents talk to the layer and never to the legacy code. Inside, nothing has changed, which means nothing can suddenly break. The layer also becomes the single place to enforce rules: who may call what, how often, and what gets logged.

  • The old system keeps doing the work it has done reliably for years.
  • New capabilities are built against the layer, in modern tools, at normal speed.
  • Every call passes through one choke point you can monitor, rate-limit and audit.

A path that is hard to regret

The strength of the wrapping approach is that each step has value on its own and none of them closes a door. You start by choosing one read-only operation, for instance looking up an order's status. It carries almost no risk, because it cannot change anything. Within a few weeks a customer-facing assistant can answer 'where is my order?' without anyone opening the old screens.

Next come the safe writes, such as creating a draft or logging a note, still behind approval. Over months, more of the daily traffic moves through the layer. Eventually someone asks which parts of the old system are still doing anything unique. That list, built from real usage, is exactly the specification you would need to retire or rebuild it, and it is something no workshop could have produced up front.

What AI actually adds

Once the entrance is clean, AI becomes straightforward to apply. An agent can read an email, work out that it is a request to change a delivery address, call the 'update order' operation, and confirm to the customer. A model can summarise a customer's ten-year history in three lines. A monitoring job can notice that today's order volume looks unlike any previous Tuesday.

None of these need the legacy system to understand AI. They only need a reliable way in and out. That is why the wrapper, not the model, is usually where most of the engineering effort and most of the value sit.

When a rewrite is the right call

Wrapping is not a permanent answer to everything. Some systems are genuinely at the end of the road: the hardware cannot be replaced, the vendor has gone, or the code cannot be changed without breaking itself. When the wrapper starts to need more workarounds than the system does work, it is time to rebuild.

The difference is in how you arrive there. Teams that wrapped first rebuild with a clear list of used features, working interfaces to keep stable, and real data to test against. Teams that start with a rewrite from a blank page often discover, two years in, that they have rebuilt something nobody needed and still missed the one report finance depends on.

Rewrites are not the enemy. Rewrites as the first step, before anyone has learned what the system really does, are.

Keep reading