AI without the hype: How companies achieve measurable results instead of pilot projects

Hardly any topic is being discussed as intensively in companies right now as artificial intelligence. Many projects fail not because of the technology, but because of decisions made long before the first pilot. Leonard Püttmann, Solutions Architect for AI at accompio, explains from his daily project work what is important from the very beginning.

The most important points briefly

• AI projects most often fail because companies start with the wrong premise, often the technology, rather than with specific business challenges.
• According to Leonard Püttmann's estimate, data issues underlie at least 80 percent of cases where scaling fails.
• The first measurable results from AI projects are often visible within 90 days.
• Relevant KPIs for AI projects are: Accuracy rate, hallucination rate, and output latency.
• Many corporate documents are created for humans and cannot be directly processed by AI systems. Therefore, data quality must be checked and, if necessary, restructured before technology is deployed.
• A modular system architecture prevents new AI models from endangering existing processes.
• Acceptance within professional fields is a key factor for the successful implementation of AI.
• A realistic roadmap prevents knee-jerk reactions and creates the foundation for sustainable competitive advantages.

Why technology is the wrong starting point

According to Leonard Püttmann, the most common misconception in AI projects is that companies make the technology the starting point. AI is not an end in itself, but a lever for existing processes, data and people within the company.

The right start begins with a question that must be answered independently of AI:

What are the biggest pain points in the company?

Depending on the business unit, this answer will vary. Only once this has been clarified will the question follow as to which technologies can be sensibly used and where their use provides added value.

This order is crucial. Companies that invest in technology first and then look for suitable use cases run the risk of ending up with exciting but commercially ineffective projects after three months. Failed AI initiatives feed critical voices within the company and make later projects significantly more difficult.

How to build a robust AI strategy

In the interview, Leonard Püttmann describes three consecutive phases:

Phase 1: Define the problem

Regardless of technology, identify what the biggest problem in the company is. It is important to, different perspectives to get hold of, because different people in the company give different answers.

Phase 2: Examine technological possibilities

What systems and technologies are available? How can they be applied to the defined problem? Here, the aim is to, existing system landscapes to understand and examine where new components can be sensibly added.

Phase 3: Rollout

Not as a one-off rollout, but designed so that the project will go over entire organisation scalable. This requires an architecture from the outset that creates blueprints, so Recurring patterns identifies and transfers it to other parts of the company.

The difference compared to isolated pilot projects: those who plan from the outset where the project is to be scaled make better decisions regarding data, system architecture, and resources.

„You shouldn't let technology guide you first, but rather the problems within the company.“

— Leonard Püttmann, Solution Architect at accompio AI

How AI projects are measured

AI results are measurable. Leonard Püttmann names three KPIs that are used in AI projects at accompio:

Accuracy rate

Experts evaluate the AI's responses in defined scenarios. This value indicates whether the AI reliably delivers correct outputs.

Hallucination rate

For business applications, this value should be in the near zero range.

Output latency

How quickly does AI deliver usable results? Especially in areas that require quick answers, speed is a quality factor in itself.

The timeframe for initial results: within a maximum of 90 days, something tangible and measurable should be available. In practice, initial evaluable results are often available after just a few weeks. This allows for an early assessment of whether a project is heading in the right direction.

Why scaling frequently fails

Many AI projects never make it beyond the pilot phase. According to Leonard Püttmann, this is the case in at least 80 percent of instances due to Data problemsA small pilot can get by with little data, but as soon as the project is extended to other parts of the company, the data, databases or system interfaces are missing.

The solution begins before the first pilot: those who plan from the outset which parts of the company and which processes an AI project will later be rolled out into can clarify in good time which Data which are necessary for this and how the Access This is secured by. For this, you need a Close cooperation with the company's IT specialists who know the affected systems.

In-house development or external solution?

The accompio philosophy is clear.

A more sensible approach is to augment an existing system landscape with a new platform, designing the interfaces in such a way that systems can communicate seamlessly with each other. This also protects against vendor lock-in: companies that wish to change or supplement their AI provider risk no interruption to ongoing processes.

AI as a service This includes an option that provides access to various AI models and can be integrated into existing system landscapes.

Let's develop an AI strategy

accompio offers comprehensive solutions to help you harness the potential of AI. We support you right from the start.

Our services

Organisationalen Strukturen und Change Management

AI projects change roles, processes, and responsibilities. Leonard Püttmann recommends defining a person from the outset who will oversee the project. This AI owner doesn't need to know the technology inside out, but they do need to understand the organisation: Who are the relevant contacts? How do employees react to new technologies?

„You have to factor in that roles, processes and responsibilities, everything in the company will change a little with AI.“

— Leonard Püttmann, Solution Architect at accompio AI

For more agile organisations, a hybrid structure can be beneficial, with a central point of contact but also a „champion“ in each team who can answer AI-related questions internally.

Change management is not an optional add-on. Many employees have not yet worked with AI and have specific concerns or unrealistic expectations. Both can be addressed through early discussions and showcasing concrete results. Those who skip this phase risk that a technically successful solution will not be accepted internally and therefore not used.

Remaining competitive in the long term: Modular resilience

AI models are evolving rapidly. New architectures and possibilities are emerging at short intervals. Companies that have structured their systems in such a way that a change of model affects the entire process landscape will find themselves having to start from scratch after twelve months.

Leonard Püttmann is talking about modular resilience here: the system is designed in such a way that individual components can be replaced without affecting the rest of the processes and systems. This protects investments in existing infrastructure and allows new models to be integrated without starting from scratch.

FAQ: Frequently asked questions about implementing AI projects in the company

How long does it take for AI to produce its first results?

According to Leonard Püttmann, Solutions Architect for Artificial Intelligence at accompio, the first measurable results should be available within a maximum of 90 days. In many projects, the first evaluable data is available after just a few weeks. Complete quality is not yet to be expected after 90 days, but a clear directional statement can be made.

Where exactly should a company start when it comes to AI?

The first step is to identify the biggest problem within the company and to establish what data is available to solve it. Only then should the decision on which technology to use be made. This process prevents resources from being channelled into projects that are technically exciting but have no business impact.

What is the difference between a pilot project and a scalable AI deployment?

A pilot project tests a technology in isolation. Scalable AI projects are designed from the outset to produce blueprints – that is, recurring patterns that can be applied to other parts of the organisation. The difference lies not in the size of the project, but in the architectural decision made at the outset.

Unternehmen sollten KI selbst entwickeln, wenn sie über die nötigen internen Ressourcen und das Fachwissen verfügen, ein tiefes Verständnis für ihre spezifischen Geschäftsanforderungen haben und eine hochgradig maßgeschneiderte oder differenzierende KI-Lösung benötigen. Sie sollten externe Lösungen nutzen, wenn sie schnell eine Lösung implementieren müssen, wenn das interne Fachwissen fehlt, wenn die Kosten für die Eigenentwicklung zu hoch sind oder wenn sie auf bewährte, Standard-KI-Anwendungen zurückgreifen können.

Developing AI platforms in-house is not a sensible use of resources for most companies. It is more effective to supplement existing system landscapes with new platforms, designing interfaces in such a way that providers can be switched later without endangering ongoing processes.

Was ist die Rolle eines KI-Eigentümers in einem Unternehmen?

The AI Owner supports AI projects organisationally. This person does not need to be technically proficient in the technology, but must understand the organisation, coordinate stakeholders and take on change management tasks. In larger or more agile companies, this role can be supplemented by „Team Champions,“ who act as AI contact persons for each department.

About the author

Leonard is a solution architect in the field of AI, with a particular focus on intelligent knowledge management and the use of Large Language Models (LLMs).

LinkedIn