Smart riskieren statt dumm riskieren: warum Wachstum Mut braucht | Umair Zaffar

Legacy Migration AI: Legacy Systems to AI Systems Without a Big Bang

Most AI projects in banks, insurers and public authorities do not fail because of the model. They fail because of the mainframe, the decades-old monolith and the databases the model is supposed to draw its data from. Legacy Migration AI, to me, means this: rebuilding legacy systems step by step so that AI services can work on them under control, without a big bang and without internal audit and supervisors losing the overview. This page describes how we do it at SIFAMO.

I was an IT architect long before I became an AI consultant. For more than ten years I have built large platforms for banks, insurers and the public sector, and in almost every one of those organisations the core process runs on a system that is older than most members of the AI team. That is not a flaw. These systems are the memory of the company: contracts, bookings, claims files, pension accounts. Anyone who wants to introduce AI has to work with them, not against them.

Why AI projects fail on integration, not on the model

The pilot almost always works. A business unit exports a few thousand cases as CSV, a team builds a prototype in a few weeks, the board is enthusiastic. Then the system is supposed to go into production, and suddenly it needs live data from the core system, write access to the file, permissions per user and a log that internal audit can read. This is where most projects die, and they die quietly: they stay pilots.

  • Data access: the core system releases data in a nightly batch or not at all. There is no API, and direct access to the tables is forbidden for good reason.
  • Data semantics: fields are called KZ_STAT_3 and mean different things depending on the client and the year. The knowledge about that lives in heads and in COBOL comments, not in documentation.
  • Permissions: the monolith knows roles, clients and the four-eyes principle. A language model with full access to the document store knows none of that.
  • Evidence: the old process has an audit-proof journal. The pilot has a log file on a developer laptop.

That is why Legacy Migration AI is first of all an architecture task. The model is the smallest and most easily replaceable part. Whoever solves integration, data and control can swap the model every few months. Whoever does not solve them has nothing in production, even with the best model.

Legacy Migration AI: five patterns that work in regulated companies

We tear nothing down. A core system that has managed contracts correctly for decades is a strength, not a problem. The question is how to build a layer around that system on which AI services can work safely. Five patterns have proven themselves for us, in exactly this order.

1. Strangler fig around the mainframe or monolith

A facade goes in front of the legacy system. Every new access runs through this facade, never directly against the tables. Business functions move one at a time behind the facade into new services, and the monolith shrinks from the outside in. Nothing is switched off before the replacement has proven under load and under review that it is equivalent. For AI this is decisive, because AI services need exactly this facade: a defined place where they read and return their proposals.

2. API and event layer first

Before the first AI service is built, the legacy system gets two things: an API for the business objects (contract, claim, customer, account) and an event stream that reports when something changes: claim created, document received, status changed. On a mainframe this can be done through change data capture, through exports at short intervals or through adapters on existing transactions. That it happens is not negotiable. Without events, AI always works on last night’s state.

3. Data products with lineage

From the raw data of the legacy systems we build data products: data sets with business names, documentation, versions and an owner from the business unit. Every field carries its origin with it, down to the source table and the transformation. This lineage is not bureaucracy. It is what an auditor wants to see when they ask what the model learned from.

4. Pseudonymised data zones: red, gray, green

Personal data never leaves the red zone in clear text. The gray zone receives pseudonymised data: names, addresses and identifiers are replaced with IDs, and the mapping table stays in-house, readable only by defined roles. The green zone holds data without personal reference, such as product rules, tariff tables and procedural instructions. Models and retrieval work in gray and green. The legacy system stays red. This zone model answers the data protection question architecturally rather than by policy.

5. AI services behind the policy gateway

Only now do the models come in. Every AI service, whether document extraction, decision support or customer dialogue, calls models exclusively through the policy gateway of our governance layer. The gateway knows the registry, checks scope and data zone, pseudonymises, writes the evidence log and escalates to a human when the signals demand it. For retrieval, access rights are enforced at query time, with the rights of the user, not those of the system. Agents get their own credentials with the least rights needed. How this looks in detail is in my post on the governance layer.

Legacy Migration AI in practice: a claims process leaves the monolith

Take a claims process at an insurer, the kind we have seen in similar form several times. Starting point: claims arrive by post, e-mail and portal. Claims handlers enter them in the policy administration system, check cover and plausibility, request documents, decide on payment or rejection and book. Everything in one system, journal included, interfaces only in the nightly run. The path to an AI-enabled system runs in four stages, and none of them is a big bang.

  • Stage 1, facade and events: the policy system gets an API for claim, contract and document plus an event stream for status changes and incoming mail. Nothing changes for the claims handlers. The gain: a current, pseudonymised picture of every claim outside the monolith.
  • Stage 2, document processing: an AI service reads incoming documents, classifies them, extracts fields and places a structured proposal in the file. The claims handler confirms or corrects. Every correction goes into the evidence log and the evaluation set. Limited risk class, because nothing is decided here.
  • Stage 3, decision support: cover check against the tariff, plausibility against the claims history, hints at known fraud patterns, always as a proposal with reasoning and sources. The human decides, the policy system books. Registry record with scope, evaluation set, tiered oversight and expiry.
  • Stage 4, customer dialogue: a RAG service answers status questions from the file, with access rights enforced at query time, pseudonymised, and abstaining when no source supports the answer. It makes no contract changes and no payment promises.

At the end of these four stages the monolith is still the system of record. But around it works a layer that understands documents, makes proposals and talks to customers, fully registered and logged. Whether the monolith is then scaled back further is an economic decision, no longer a prerequisite.

What stays, what moves, what never moves without controls

The most important decision in any migration is not the technology but the cut: which responsibility stays in the legacy system, which moves into AI services, and which may only move under conditions.

AreaWhere it goesWhy
System of record: contracts, accounts, bookings, journalStays in the legacy systemCorrect, audited, tamper-proof. If it is replaced at all, that happens after the migration, not during it.
Document processing: incoming mail, classification, extractionMoves into AI servicesHigh manual effort, easy to measure, the human confirms. The best first target.
Decision support: cover, plausibility, prioritisationMoves as a proposal with reasoningThe model reasons, the human decides, the evidence log records both.
Customer dialogue: status, documents, explanationsMoves into RAG services with rights checksAccess only to the own file, abstention without a source, no commitments.
Decisions about people: benefits, credit, hiring, pricingNever without registry, tiered oversight and kill switchHigh-risk under the EU AI Act. Automation only when reasoning, bias check and the ability to intervene are verifiable.

The role of the governance layer during migration

Many organisations build the governance after the migration is done. That is the expensive order. During the migration the governance layer is the tool that holds the old and the new world together.

  • Registry from day one: the existing rules engine in the monolith gets a record too, with owner, purpose and scope. The AI service later joins as another registered service, not as a foreign body.
  • Gateway as the only door: no service talks directly to a model, and none talks directly to the tables of the legacy system. That prevents the shadow solution that appears somewhere in every migration.
  • Tiered oversight per stage: document extraction with sampling, decision proposals with full review, customer dialogue with monitoring. The tier is in the registry and changes only with a new evaluation.
  • Kill switch with the business owner: the business unit can switch off any new service without a deployment. The process then runs as before, because the legacy system never went away. That is the real reason the strangler pattern works in regulated companies.
  • Quality gates before every stage: evaluation set from real, pseudonymised cases, bias check, load test, sign-off by business and risk. No passed gate, no route.

Testing and audit continuity: the same evidence log before and after

The question CIOs and heads of internal audit ask me most often: how do we prove that the process after the migration is at least as good as before? The answer has three parts.

First, shadow mode. Every new AI service initially runs in parallel. It produces proposals nobody sees while the old process decides. After a defined period, business and risk compare the results on the entire volume of that period, not on a hand-picked sample. Only on equivalence or demonstrated improvement does the service move to the stage where a human confirms.

Second, an evaluation set that belongs to the business unit. Historical, pseudonymised cases with known outcomes, versioned, with edge cases and known wrong decisions. Every new model version and every new prompt version runs against it before it gets a route.

Third, and this is what most overlook: one evidence log for old and new. We write the decisions of the existing process into the evidence log in the same format as later those of the AI services: input, data reference, applied rule or model version, result, reviewer action. An auditor sees one continuous history across the migration date and can place a case from before next to one from after without asking the engineer. That is audit continuity, and it costs almost nothing if it is planned early.

Timeline and team shape

A realistic timeline has four sections. The assessment takes weeks, not months: system landscape, data flows, permissions, regulatory classification. Facade, events and the first data products are the work of the first months and run entirely without AI. The first productive AI service, usually document processing, follows and should be running within the first year with evidence log and a passed gate. Decision support and customer dialogue come after that, each stage with its own gate. Anyone who promises faster is skipping the facade or the governance, and both come back to bite at the first audit.

The team is smaller than many expect, but staffed differently from a classic AI team:

  • One enterprise architect responsible for the target picture, the cuts and the order. That is the role I take in projects myself.
  • The people who know the legacy system. They are the bottleneck and the most important resource. Anyone who lets them go in the course of the migration loses knowledge that no model can reconstruct.
  • Data and AI engineers for events, data products, lineage, services, retrieval and evaluation.
  • A business owner from the business unit who signs registry records and holds the kill switch.
  • Risk, compliance, data protection and internal audit from the first week, not as a sign-off at the end.

Typical mistakes we see in migrations

  • Big bang: replacing the core system and introducing AI in one programme. Two risks multiply, and the journal breaks exactly where the auditor looks.
  • CSV as data strategy: exports from the legacy system that models learn from, whose origin nobody can prove later.
  • Model straight onto the database: a language model with read access to the production tables because it is faster. No zone, no gateway, no log.
  • Starting with decisions about people because that is where the largest savings are. It is also the largest regulatory risk and the wrong first step.
  • Boundaries in the prompt instead of in the gateway and the network policy. A prompt is an instruction, not a boundary.

Further reading

  • AI Governance: registry, logging, escalation and what supervisors and internal audit want to see.
  • EU AI Act: risk classes, obligations and deadlines that determine the order of the migration.
  • AI Consultant: how a project with us runs from assessment to productive system.

About Umair Zaffar

I am founder and CEO of sifamo GmbH, an AI consultancy for regulated industries with offices in Berlin and Frankfurt am Main. Founded in 2019, bootstrapped, 28 consultants, architects and developers in 2026. I come from enterprise architecture and built large platforms for banks, insurers and the public sector for more than ten years before AI in production became our focus. Our clients include Commerzbank, VW Bank, Bosch, DHL, Zurich, VGH and the German federal administration. Since 2023 we have rebuilt SIFAMO itself around AI, from recruiting to an ERP with an AI board. How that came about is on my founder page.

Frequently asked questions

Do we have to replace the mainframe before AI is possible?

No. In most cases it is the other way round: the mainframe stays the system of record, and the AI services work on it through facade, events and data products. Whether and when the legacy system is scaled back is decided afterwards, economically and without time pressure.

How long does a legacy migration to AI take?

The assessment takes weeks; the first productive AI service with full governance should be running within the first year. The further stages follow at their own pace, each with its own gate. What matters is not the total duration but that every stage is productive, reviewed and reversible on its own.

Which data may the model see?

No personal data in clear text. Models and retrieval work in the gray zone with pseudonymised IDs and in the green zone with data that has no personal reference. The mapping stays in-house.

What does the EU AI Act mean for the migration?

It determines the order. Document processing and status information are usually limited risk and suitable as an entry point. Anything that co-decides on people’s benefits, credit or contracts is high-risk and needs registry, logging, human oversight and documentation before it goes into production. The obligations in detail are on my page on the EU AI Act.

You have a core system nobody wants to touch and an AI strategy that depends on exactly that? Reach me via the contact page, by e-mail at contact@sifamo.com or directly on LinkedIn.