Artificial intelligence is high on the agenda in many companies. A current survey of marketing decision-makers ranks the topic among the most important strategic fields. At the same time, the same surveys report a high failure rate: around 85 per cent of projects do not reach productive regular operation, but remain stuck in pilot status or are quietly discontinued.
The reason is remarkable. As a rule, it is not the models that fail, nor the interfaces. What is missing is processes, clear responsibilities and a reliable data basis. Put another way: the problem is organisational, not technical.
The pilot as a dead end
A typical course looks like this: a business department tests a tool, for example for drafting text, classifying enquiries or evaluating form data. The test works, the results are convincing on a small scale. Then the real work begins – and that is precisely where the project often ends.
Because in the pilot, much can remain improvised. Data is compiled by hand, results are checked manually, errors are corrected informally. That is not enough for regular operation. There, defined input data, documented responsibilities, approval paths and an answer to the question of who intervenes when the system outputs something incorrect are needed.
Anyone who asks these questions only after the pilot asks them too late. The effort of subsequently fitting a working experiment into an organisation often exceeds the effort of the experiment itself – and that is precisely where budgets run dry.
Data basis before model selection
A recurring pattern: companies discuss providers and models intensively before it has been clarified which data is even available in a usable form. Product information sits in three systems, customer data is maintained to varying standards, content in the content management system – i.e. in the editorial backend of the website – has grown over years and is inconsistently structured.
A language model does not compensate for such gaps. It reproduces them. Anyone planning, for example, an AI-supported search or an automated product description should first check whether fields are filled consistently, whether redundancies exist and whether there is a leading source for each data item. This preparatory work seems unspectacular, but it decides whether a project holds up.
What distinguishes teams that make it into operation
Several characteristics that separate successful undertakings from failed ones can be derived from these observations:
- A concrete process instead of a technology. The starting point is a nameable task with measurable effort – not the wish to use AI.
- Named responsibility. There is a person responsible for output quality, data maintenance and further development. Not a committee, not a department.
- Defined termination criteria. Before the start it is clear what success will be measured by and when an undertaking will be ended. This prevents pilots that continue for years without ever becoming productive.
- Involvement of the specialist users. Those who execute the process daily know the exceptions. It is these exceptions that automated procedures fail on.
- Control points in the workflow. For critical outputs, human approval remains planned in – documented, not improvised.
Keep expectations realistic
Some of the discontinuations are due to excessive expectations. If a project is started with the prospect of comprehensive automation, every partial relief looks like a failure. If, on the other hand, a limited, clearly defined improvement is aimed at from the outset, the benefit can be demonstrated – and that forms the basis for the next step.
A scope that remains manageable is therefore sensible: one process, one data source, one user group. Only when this segment runs stably in operation is expansion worthwhile. The reverse path – a broad platform decision without a proven use case – regularly leads to systems that nobody uses.
Governance is not an afterthought
As soon as AI productively generates content or prepares decisions, questions arise about traceability, data protection and liability. Anyone who processes customer data must know where this data flows and how long it is stored. Anyone who publishes texts needs editorial control. These points belong in project planning, not in later remediation.
The organisational framework is therefore not a brake block, but the prerequisite for a pilot even being allowed to transition into operation.
Classification for web and software projects
For web and process projects, this means above all one thing: the effort shifts from model selection towards the preparation of data, interfaces and workflows. A cleanly structured content management system, well-maintained product data and documented approval paths are the actual foundation of any AI function on a website or in the back office.
Anyone who creates this foundation can later add or swap out AI building blocks without having to set up the project from scratch – and thus avoids ending up in the group of projects that remain stuck in pilot status.