How do you simplify complex data systems?

Simplifying a data landscape is mostly subtraction: inventory what exists, consolidate the redundant parts, give every dataset an owner, and standardize how data moves. New platforms come last, if at all. Uliasti designs data architecture for Swiss enterprises, including work for SIX and Aargauische Kantonalbank, and this is the sequence that works in practice.

Complexity is accumulated decisions

No one designs a complicated data landscape. It accumulates: a tool per department, a pipeline per project, a report per manager who once asked for one. Each decision was locally reasonable. The sum is a system where nobody can say with confidence where a number came from, and where every change breaks something two teams away.

That matters because the cause tells you the cure. If complexity came from unmanaged accumulation, the fix is management and subtraction, not another platform on top of the pile.

Step 1: Inventory what actually exists

Before touching anything, map it: systems holding data, pipelines moving it, reports consuming it, and the people who actually use each one. Include the unofficial layer, the Excel exports and Access databases that run more of the business than anyone admits.

Two findings appear in nearly every inventory we run. A meaningful share of pipelines and reports have no remaining consumer; they run because nobody dared to stop them. And several systems hold overlapping copies of the same domain, each slightly different, which is where the "which number is right" meetings come from.

Step 2: Consolidate before you modernize

The instinct is to migrate everything onto something new. Resist it. First retire what the inventory exposed:

• Switch off pipelines and reports with no consumers, after a notice period. Deletion is the cheapest modernization there is.
• Merge redundant copies of the same data behind a single source, and repoint the consumers.
• Only what survives this pruning deserves to be migrated anywhere.

Consolidating first regularly cuts the scope of any later platform work by a third or more, which is a third of the cost you never spend.

Step 3: Give every dataset an owner

Data without an owner degrades. Every core dataset needs one named person who answers for its quality, its definitions and who may access it. This is not a governance committee or a two-year framework; it is a list of names next to a list of datasets, published where everyone can see it. Lightweight ownership catches most problems that heavyweight governance programs are built to catch, at a fraction of the ceremony.

Step 4: Standardize how data moves

Most landscapes have five ways of moving data between systems, built by five generations of contractors. Pick one orchestration approach and one pipeline pattern, apply them to everything you touch from now on, and migrate old flows opportunistically when they next need changes. Every pipeline should be observable: when it ran, what it moved, whether it succeeded. Silent failure is how data quality problems age into data quality disasters.

When a new platform is actually justified

After the pruning, some situations genuinely call for new infrastructure: data volumes or workloads the current stack cannot carry, regulatory requirements the old systems cannot meet, or analytical ambitions, including AI, that need data the current landscape cannot serve in usable form. Then build deliberately; that is our enterprise data architecture and scalable data platforms work. A platform bought to fix disorder just gives the disorder a bigger house.

If AI is the driver, note that the preparation is the same: the access, quality and ownership work above is exactly what integrating AI into existing systems requires.

Frequently asked questions

Do we need a data warehouse or a lakehouse?

Decide after the inventory and consolidation, not before. Many SMEs discover that one governed warehouse covers their real needs, and some need none at all once redundant systems are merged. The architecture should follow the surviving workloads, not the other way around.

Can we simplify without stopping the business?

Yes, and it is the only way that works. Everything is incremental: retire one pipeline, merge one redundant store, assign one owner at a time. There is no cutover moment. Big-bang data migrations are how companies turn a complexity problem into an availability problem.

How long does it take to see results?

The inventory takes two to four weeks and pays for itself immediately in retired pipelines and reports. Visible simplification, meaning fewer systems and one agreed source for core numbers, typically lands within three to six months. Full consolidation of a grown enterprise landscape is a six-to-eighteen-month program done incrementally.

Does simplification require new headcount?

Usually not. Ownership is a role attached to existing people close to the data, not a new department. Where firms bring in outside help, including ours, it is for the inventory, the architecture decisions and the consolidation engineering, all of which are bounded engagements rather than permanent additions.