When an automation project fails, you almost always hear the same sentence: the tool is no good. In the rarest of cases is that true.
"The AI doesn't work." Usually that's not the case.
When an automation or digitalisation project doesn't deliver, the diagnosis comes quickly: the software doesn't fit. The AI produces nonsense. The system is no good.
In the rarest of cases is that true. What looks like a technology problem is almost always a data problem. The input data was incomplete, contradictory, or simply not there. And no system in the world can turn bad data into good results.
Computer science has had a term for this for decades: garbage in, garbage out. Put rubbish in, get rubbish out.
What has changed: in the past, an experienced person corrected that rubbish in their head, without anyone noticing. The clerk already knew which detail was missing from the order and asked about it. But as soon as you automate, that silent correction mechanism falls away. The error becomes visible. And expensive.
That is why every process in the IPO framework begins with the input. Not because it is the most exciting part, but because everything else builds on it.
Why input is especially hard in mid-sized companies
In a typical industrial company, data lives in many places at once. In the ERP system. In dozens of Excel files on various drives. In email inboxes. On paper. And increasingly in machine sensors delivering values by the second.
These sources grew over years, they were not planned. Each one serves its purpose. The problem arises in the interplay: none of these sources has priority.
tells the truth?
When the ERP shows a stock level of 40 and the Excel list says 35, nobody knows for certain what is correct. Both get used, depending on who happens to be looking. This is the normal state in mid-sized companies. And it is not a failing, but a logical consequence of organic growth. But it is the root of most process problems.
The six dimensions of data quality
Whether an input can be processed at all can be measured along six dimensions. They sound technical, but each has a very concrete consequence in everyday work.
Completeness
Are all mandatory fields filled? An order without a bill of materials cannot go into production, but is accepted anyway.
Correctness
Do the values match reality? A supplier that has not delivered for a year blocks purchasing.
Consistency
Do all systems say the same thing? When sales, warehouse and accounting differ, gut feeling decides instead of the data.
Timeliness
How old is the data? An outdated materials price list produces quotes that lose orders or burn margin.
Uniqueness
Are there duplicates? The same customer created three times distorts every analysis before it begins.
Availability
Is the data in the right place at the right time? Information that arrives too late might as well have been missing.
Anyone who wants to automate a process needs an answer for each of these six dimensions. Not for every field, but for those the output depends on.
From chaos to structure
The good news: you don't have to solve everything at once. But you need a target picture and the right order.
The first step is almost always data cleansing, the systematic removal of duplicates, filling gaps and standardising formats. Unspectacular, but indispensable. It is the work nobody wants to see, and it decides between success and failure.
The second is the principle of the single source of truth, one leading source. For every important piece of information, it must be defined which system tells the truth. Not three systems contradicting each other, but one source you trust, and everything else derives from it.
As data volumes grow and requirements rise, two concepts then come into play, each serving its own purpose:
What matters is the order: structure first, then the right tool. A data lake filled with unchecked data does not unfold its value, it becomes a data swamp. Whoever gets their data quality under control first makes the technology decision later from a position of clarity. And then data warehouse and data lake are powerful tools that create real value.
Where input really breaks: the interfaces
The most common source of error is not the individual data source. It is the handover between them.
Every time data flows from one system into another, from one department to the next, from one format into another, an interface is created. And every interface is a potential breaking point. Whoever wants to improve their input quality therefore looks at the interfaces first. Not at the sources themselves, but at the paths between them.
What this means for automation and AI
Here the circle closes. An automation is only as good as the input it receives. An AI run on contradictory data makes contradictory decisions, only faster and at a larger scale.
Clean input work is not a preliminary. It is the actual foundation.
Whoever invests here automates later on solid ground. Whoever skips it builds on sand. In the next article we turn to the processing sector: what happens to the input, where value is created, and why optimisation always comes before automation.