A guide with a working document
Nine chapters for orientation and a fact sheet summarizing the results of a project. Not an AI textbook, not a maturity model, and not a recommendation for a specific product.
The AI Navigator is loading. If this message remains visible, the JavaScript did not run. Please paste the full content into an HTML widget.
Skip to contentThe NIRO AI Navigator is a tool for specific AI projects in industrial companies. It organizes the questions that need to be answered before, during, and after a pilot project, and records the answers for each project in one place.
Nine chapters for orientation and a fact sheet summarizing the results of a project. Not an AI textbook, not a maturity model, and not a recommendation for a specific product.
For anyone who is responsible for, evaluates, or approves an AI project. Technical terms are explained where they are needed; the most important ones are also listed in the glossary in Chapter 09.
A completed project profile form with objectives, key metrics, baseline values, data access, roles, checkpoints, pilot decision, and operational responsibility. Available as a handout for download or printing.
An anonymized example from the program, condensed into three steps. It shows what kind of responses the Navigator requires before a technology is selected.
Customer specifications are read manually and transferred to an internal requirements document. The process is time-consuming, and omitted requirements lead to follow-up questions.
Processing time per document and the correction rate are measured in the current process. Only then is a comparison made to determine whether an AI-supported workflow offers an advantage over a simpler solution.
The pilot is limited to a single document type. The success threshold, critical errors, checkpoints, fallback procedures, and decision deadline are established before the start.
The NIRO KI Navigator was developed through collaboration with companies in the NIRO network. Details on how the program unfolded and what came of it can be found at the bottom of this page.
There’s no set starting point. Choose the entry point that fits your current situation. The NIRO AI Navigator doesn’t prescribe a fixed order. It guides you to the questions that should be addressed next.
The NIRO KI Navigator is not intended to replace a technical, legal, or operational assessment of specific cases. It is a document based on experience and insights gained from real-world projects within the NIRO network.
Identify typical weaknesses in the process, use real-world examples, and describe an initial candidate without giving away the solution.
Find your bearings → Getting Started 02Check the impact, requirements, data access, and the simplest adequate solution before selecting a technology.
Evaluate the idea → Introduction 03Make sure to thoroughly prepare the pilot, roles, checkpoints, approvals, and subsequent operations.
Ensure successful implementation →Regardless of how you get started, four tasks consistently arise in these projects. The chapter references indicate where they are addressed in the NIRO AI Navigator.
The page is divided into three phases. Chapters 01 through 03 are sufficient for an initial overview. A complete run-through with a specific project takes about 45 minutes and can be paused at any time.
The project profile brings together all the information about an AI use case in one place: objectives, data, roles, decisions, and open questions. Use it to structure your project and discuss the next steps with your team.
Your entries will not be saved or transmitted to NIRO or any other parties. They apply only to the currently open page.
Three anonymized examples from the program show just how brief a useful early draft can be. A fully completed example follows below.
The process follows three steps: clarifying the subject matter, assessing feasibility, and embedding the application within the process. In addition, there are six dimensions through which each project is examined. The third step does not end with implementation but transitions into ongoing operations.
Select a step to see which questions are answered there.
The six dimensions are not a maturity model. Each project is examined from all six perspectives; the depth to which a dimension needs to be addressed depends on the use case and its current phase. Select a field to view the guiding questions.
These statements indicate that a project has not yet been sufficiently clarified. Select a signal to see the appropriate response.
The examples draw on the projects covered in the program. They serve as inspiration and illustrate typical scenarios, but do not constitute a recommendation for any specific technology. This chapter does not include any profile fields; the practical work begins in Chapter 04.
Not every problem is an AI problem. The suitability check assesses the nature of the task ( ), the impact matrix identifies the target metric, and problem typology leads to the appropriate solution family ( ). This section fills out sections B and C of the profile.
The first two questions determine whether an improvement is worthwhile. The next two questions indicate whether a meaningful test is already possible. Even so, an attractive project may not yet be ready for a pilot.
A project can only be evaluated if it has an impact on a measurable target metric. Select an impact category to view the typical key metrics.
First, the desired task is described. Then, off-the-shelf software, rules, traditional methods, machine learning, optimization, and generative AI are compared as possible solutions.
These terms are often used as successive stages of maturity. In fact, refers to different forms of control. The key factor is who determines the next relevant step during execution and which actions the system is permitted to trigger.
| Criterion | Rule / Standard Function | Automation | AI-powered workflow | AI agent |
|---|---|---|---|---|
| Who determines the procedure? | The function or rule is fully defined | The sequence of steps is fully specified | The sequence is predetermined; individual steps use AI for evaluation or generation | The goal and constraints are specified; the system selects individual next actions at runtime |
| Scope of action | None outside the rules | None outside the defined process | Substantive flexibility in the AI step, but no freely chosen process path | Limited flexibility in the selection and order of permitted actions |
| System Permissions | Read or perform a clearly defined function | Defined actions in connected systems | Defined actions based on AI results and fixed checkpoints | Use multiple tools or systems within explicit permissions |
| Suitable when … | the business logic can be clearly defined | a stable, recurring process is to be automated | unstructured information is processed within a known workflow | the path to the goal varies depending on the situation and cannot be fully predetermined |
| Example | Approval Once a Specified Threshold Is Reached | Create an order according to defined rules | Classify a customer email and generate a draft response for review | Investigate the issue, gather information from multiple sources, and propose the next course of action |
| Need for Oversight | Check Rules and Exceptions | Monitor processes, error cases, and interfaces | Define AI results, uncertainty, and human checkpoints | Additionally, ensure action permissions, scope, logging, termination, and fallback options |
The more selection and action rights a system is granted at runtime, the more important it becomes to implement restrictions, logging, monitoring, and fallback options.
First, determine the type of data, then its usability. Both apply to the individual use case, not to the company as a whole. This section fills out Block D of the profile.
Different decision-making problems require different types of data. Select all types that are relevant to the project. Multiple selections are allowed.
Five criteria, each with four levels. For each row, select the level that best describes the current status at . The evaluation distinguishes between accessibility, content suitability, and the prerequisites for stable operation at . The deciding factor is the weakest criterion, not the average.
AI projects rarely fail because of the model itself. They fail because no one is in charge or because the results aren’t used in everyday work. This aspect determines whether the application is actually used in everyday work.
A role describes a responsibility, not necessarily a specific position. A single person may hold multiple roles , or these roles may be filled by external support. Before the pilot begins, no key responsibility should remain unclear.
Acceptance arises from the interplay of recognizable benefits, comprehensibility, trust, and effective integration into the workflow. A lack of transparency can make usage particularly difficult, especially when making decisions with significant consequences.
The result is applied in the actual workflow and improves the desired metric.
The department can validate a result and explain any deviations.
Clear presentation, visible uncertainty, explainable influencing factors.
Implementations often do not proceed in a linear fashion. Uncertainty, questions, criticism, and refinements should be factored into the plan, without dismissing legitimate objections as mere resistance .
Involve stakeholders and affected parties while the process and control points are still malleable.
Avoid abstract AI promises; instead, explain specific tasks, expected improvements, and known error cases.
Evaluate criticism, workarounds, and corrections as indications of process, quality, or trust issues.
Don’t just provide one-time training; instead, update knowledge as new features, error patterns, and responsibilities emerge.
A working prototype is not yet an application. Value is only created when the result is integrated into data flows, workflows, responsibilities, and control points . This section covers sections F and G of the profile.
Select a step to view the related design questions.
The decision-making level determines how much control, documentation, and coordination are required. It is one of the first specifications to be defined, not one of the last.
The preliminary review does not determine whether a project is permissible. It identifies which risks, agencies, and in-depth reviews may be relevant prior to pilot or production use.
This classification is intended to provide structure and does not replace a case-by-case review of legal, data protection, or employee participation matters.
Not a weighted score, but a transparent decision: Are the problem and benefits relevant, is the next learning step justifiable, and are the risks manageable?
Most guides end with the implementation phase. Within the company, however, that’s where the longest phase begins: The application must continue to run even when no one is watching it, when source systems change, and when the person who built it has moved on to other tasks.
The transition from project mode to regular operation is a distinct step and not a side effect of the rollout. Without clearly defined responsibilities, the application will fail at the first sign of a malfunction.
An often-overlooked metric. A technically flawless solution can be bypassed in everyday use without this being immediately apparent in the model or system quality.
Two separate questions: Is the result getting worse? And has the target value changed compared to the baseline? Without the second measurement, the benefit remains merely an assertion.
An ERP update, a modified form, new master data logic, a different model, an adjusted prompt, or a changed platform can affect the application without being noticed.
Continues the preliminary review from Section 07 in the operational phase. Without a history, it is impossible to reconstruct, after an error, which status led to which decision.
Project costs are one-time expenses, while operating costs are ongoing: licenses, usage fees, computing power, database maintenance, and the labor hours required for testing and support.
Many applications depend on a single person. If that person leaves the company or the service provider changes, the solution can no longer be maintained, even if it continues to run.
A fixed date with pre-agreed criteria. Shutting down is a legitimate outcome: an application with no impact wastes attention and undermines trust in future projects.
Key metrics from the first project rarely remain. What remains are the reusable elements: clarified data access, defined roles, verified approval workflows, and an evaluation metric.
The answers summarize insights from the workshops and projects and must be applied to the respective use case.
Only the terms that are actually needed to work with the Navigator.
This anonymized example walks through the case from the overview across all six stages. It is intentionally concrete, without resorting to fictitious project results or metrics.
Customer specifications are read manually and transferred to an internal requirements document. This process is time-consuming; omitted requirements lead to follow-up inquiries.
The goals are to reduce processing time and ensure a more traceable transfer of requirements. Prior to the pilot, processing time per document and the correction rate in the current process will be measured.
An AI-supported workflow extracts requirements according to a defined process. Subject-matter validation remains the responsibility of humans; an autonomous agent is not required.
Representative specifications and the internal target format are available. Terminology, document variants, usage rights, and access to legacy documents are clarified in advance.
The pilot is limited to a single document type. Success thresholds, critical omissions, a human checkpoint, a fallback plan, and a review date are defined before the start.
For a rollout, operational responsibility, versioning, usage rate, ongoing costs, and triggers for reassessment are documented.
The workspace is initially closed to keep the Navigator page uncluttered. The field labels correspond to the instructions in sections 02 through 08. Clicking a field label in the body text takes you directly to the corresponding field. Your entries apply only to the currently open page. They will be lost when you reload or close the page.
Ready to get started?
Open the complete online profile or start by downloading the comprehensive case study as a starting point.
One row per action. The “Goal” column is key: it explains why this step comes next. Observations and obstacles also belong here. These are the details that can be shared with the group later.
| Date | Context / Action | Objective of this measure | Challenge / Observation | Dimension | Status | Row |
|---|
The NIRO AI Navigator brings together the insights gained from a collaborative development and learning process within the NIRO network. It makes the key questions from completed corporate projects available for use in future AI projects.
The starting point was the question of how companies can identify suitable AI projects, evaluate them realistically, and integrate them with benefits, data, organization, and operations from the very beginning.
Together with companies in the NIRO network, we addressed and evaluated specific starting points, application ideas, obstacles, and potential solutions.
The NIRO KI Navigator serves as a guide and assessment tool. The profile document records the results, decisions, and outstanding issues of a specific project as an ongoing working document.
The content summarizes practical experiences from the program. It does not constitute a representative impact study or a technical, legal, or operational review of individual cases.
The content is derived from the projects of the participating companies. Your feedback on obstacles, solutions, and missing questions helps refine the final version and continue the work with the NIRO AI Navigator within the companies.
Where did things get stuck? Data access, interfaces, approvals, acceptance, costs—categorized into one of the six dimensions.
Even small ones: a phrasing that convinced management, a way to get through an approval process, a usable data extract.
Who needed to be involved, in what order, and with how much advance notice? This kind of experience is hard to find in books but invaluable in everyday practice.
What questions arise in your company that aren’t answered here? This helps identify where the NIRO AI Navigator has its limitations.