Cisco has tasked a 27-year-old former founder with overhauling legacy code across the company with AI support.The parameters are unusual: unlimited AI tokens are available, meaning no fixed quota of requests and generated text that would limit the use of language models.The stated goal: 70 per cent of products are to be renewed by the end of 2027.

Nothing more about the project has been made public so far.That is precisely why it is interesting as a case study – less because of the figures, more because of the decision behind it: a corporation treats the modernisation of legacy systems not as a side task of ongoing development, but as a programme of its own with its own responsibility and its own budget.

Why legacy code is so persistent

Legacy code is software that runs in production but is maintained using outdated techniques, without sufficient tests or without up-to-date documentation.In the web and CMS landscapes of mid-sized companies, this typically looks like this: a website that has grown over more than ten years, several generations of extensions, custom modifications by service providers who left the project long ago, plus interfaces to ERP, CRM or shop systems that worked at some point and that nobody has touched since.

The problem is rarely the technology alone.It is the lack of knowledge about what the code actually does and why it was built that way.Every change is therefore risky, every risk leads to postponement, and every postponement increases the distance to current technology.After a few years, a complete rebuild appears to be the only option – with all the well-known drawbacks: a long standstill in further development, high one-off costs and the loss of functions that nobody remembers any more.

What AI can actually contribute here

The most productive use of language models with legacy systems is not in writing new code, but in understanding existing code.Models can go through large amounts of unknown code, make dependencies visible and summarise in plain language what task a module performs.That does not replace a business decision, but it considerably shortens the onboarding phase – precisely the part that traditionally makes modernisation projects expensive.

In practice, AI is mainly useful in legacy projects for:

  • Inventory: documentation of modules, interfaces and data flows that were never recorded in writing.
  • Test coverage: generating tests for code that was previously untested – as a safety net before anything is rebuilt at all.
  • Routine refactoring: recurring adjustments across many files, for example with version upgrades or the replacement of outdated functions.
  • Risk assessment: indications of dead code paths, duplicated logic or obvious security problems.

The crucial point: these tasks are verifiable.A generated test either runs or it does not.Documentation can be checked against actual behaviour.Where results can be verified automatically or quickly by hand, AI support delivers reliably – where they cannot, all that is produced is plausibly sounding text.

Unlimited tokens do not solve the problem on their own

What is special about the Cisco approach is the decoupling from usage costs.If every request to a model becomes noticeable in the budget, restraint sets in: you try less, generate fewer variants, skip the third pass.If this limit falls away, the way of working shifts towards experimentation.

But the bottleneck only moves.It is then no longer the tokens, but the review of the results.Every generated change must be read, tested and signed off.Without automated tests, clean version control and a working deployment pipeline – that is, a regulated way of bringing changes into live operation under control – more AI output simply produces more unreviewed changes.The foundations of modern software development are not replaced by AI, but become more important.

Transferring this to mid-sized projects

A corporate programme with a multi-year goal cannot be transferred one to one.The pattern behind it can.For grown TYPO3 installations, custom portals or older specialist applications, a step-by-step approach is advisable:

  1. Inventory before rebuild: record which parts of the system are actively used, which merely run along and which can be dropped without risk.Code that is not modernised is cheaper than modernised code that nobody needs.
  2. Build safeguards: tests for business-critical processes before the first rebuild begins.This is where AI support is particularly economical, because writing tests for legacy code is tedious and unpopular.
  3. Modernise in slices: replace individual areas while the overall system remains in production.In CMS projects, for example, extension by extension instead of a relaunch in one go.
  4. Record knowledge: the documentation created during modernisation belongs in the repository, not in a chat history.

The business advantage lies in spreading the risk.Each step can be assessed individually, aborted individually and already delivers benefit on its own – for example in the form of lower maintenance costs or a better security posture.A complete rebuild only delivers benefits at the end, and experience shows that the end keeps getting pushed back.

What can realistically be expected

Whether Cisco will reach the target by the end of 2027 is open, and a rate of 70 percent renewed products says little in itself about quality or operational stability.Reliable comparative figures on the productivity of AI in modernisation projects are hardly available to date.Anyone planning budgets should therefore work with measurable interim results instead of blanket savings promises.

A limited pilot makes sense: a clearly defined module, a defined period, a before-and-after measurement of effort, error rate and lead time.This provides a basis for decision-making that is based on your own code – and not on reports of experience from an entirely different order of magnitude.

For web and software projects, this shifts above all the cost-benefit calculation between rebuilding and further development: understanding and safeguarding existing systems becomes cheaper, the complete relaunch loses part of its justification.The prerequisite remains a proper technical basis of tests, version control and automated deployments – without it, AI merely accelerates the production of changes that no one can keep track of anymore.Anyone operating a grown TYPO3 or custom solution is currently best served by a small, measurable modernisation step instead of a big promise.

Sources