AI Strategy & Business Case

AI management: the guide for executive and divisional leadership

Marin Niehues2 November 202621 min read

Marin Niehues speaking with a microphone at a lectern next to a slide showing a feedback matrix.

AI management is the steering of all of a company's AI initiatives as a portfolio, from selection and the business case through to measured value in operation. This guide answers the questions executive and divisional leaders ask at the outset. For each answer it points to the articles that go deeper. This article provides an overview and does not constitute legal advice.

The starting point is much the same in most companies. AI has arrived in day-to-day work, through assistants and through features in existing software, and often in a first set of applications of their own. Official statistics for Germany show that roughly one in four companies with ten or more employees uses AI, and more than half of companies with 250 or more employees. Using AI is not yet the same as managing it. The difference shows in a simple test. For every AI application, executive leadership can name the result it changes and the person accountable for it. Add to that the value it has delivered so far. If any of these is missing, steering is missing. The sections below build that steering in the order decision-makers need it.

What is AI management and what does it include?

AI management covers five tasks of company leadership:

  • setting objectives
  • assigning accountability
  • selecting applications and putting them in sequence
  • defining the framework for law and risk
  • steering the portfolio with metrics

Building individual applications is part of it, but that is the specialists' job. Management decides which applications get built at all and with what budget. It also defines the criteria by which initiatives are continued or stopped.

Distinguishing the neighbouring terms brings clarity to the discussion:

  • The AI strategy sets out where the company wants to go with AI and which business results should change. It is written once and reviewed annually.
  • AI governance covers committees and policies together with approvals. It is the regulatory frame within AI management.
  • AI implementation takes a single application from prototype to regular operation. It is the execution layer covered in depth by the guide to AI implementation.
  • AI management connects all three into an ongoing process with a fixed rhythm.

How the individual areas of control are built is described in AI management: what executive leadership must steer.

A four-stage maturity model has proven useful for assessing where you stand. It measures steering and deliberately ignores the number of tools.

Anyone running pilots in several units without assessing them together is at stage 2. The step to stage 3 has the greatest effect, because it shifts budget from initiatives without an owner to initiatives with proven value. How an assessment of your own position works is set out in the article on AI maturity.

Who is accountable for AI in the company?

Overall accountability for AI sits with executive leadership, which assigns it by name to one of its members. The value of an individual application is owned by the divisional head in whose process it operates, and its operation by a named team. This dual assignment is the single most important decision in AI management, because it settles who answers for the result in the end.

Three failure patterns arise in practice. If AI sits solely with IT, the result is technically clean solutions that no business unit asks for. If it sits solely with an innovation team, the result is prototypes with no route into operation. If it sits with nobody, uncontrolled use of freely available tools grows. All three can be avoided with a clear allocation of roles:

  • Executive leadership: sets objectives and budget, decides on continuation and closure in the portfolio review.
  • Accountable executive or Chief AI Officer: runs the portfolio, owns governance and reporting.
  • Divisional head with product owner: owns the value of each application, prioritizes requirements and secures usage in the process.
  • Operations team: secures stable operation and further development after go-live.

Whether a dedicated Chief AI Officer makes sense depends on the size of the portfolio. From around ten parallel initiatives across several units, the coordination effort justifies a dedicated role; below that, a named executive with a small central team is sufficient. The mandate and staffing of the role, together with its metrics, are covered in the article on the role of the Chief AI Officer.

The second structural question concerns the organizational model. A central team pools scarce specialists and suits the start. A center of excellence sets standards and enables the units that own their applications. A decentralized model suits companies with autonomous business units and mature teams. The criteria for choosing and the signs that a model no longer fits are covered in AI organization: central team, center of excellence or decentralized.

Strategy and roadmap set the direction

An AI strategy answers three questions in a fixed order. First: which business results should change? Next: in which processes does that change happen? Last: which capabilities does the company need for it? A strategy that starts with technologies misses the first question and produces solutions in search of a problem. How a strategy with objectives and priorities becomes a roadmap is shown in the article on AI strategy for companies.

The scope of the strategy follows the size of the company. For a mid-sized firm with one line of business, two to three pages with objectives and first fields of application, each with an owner, are enough. A group with several business units also needs a split between group-wide platforms and the units' own applications.

The strategy produces the roadmap for the first twelve months. It plans in quarters and connects three strands:

  • the first application with measurable value in operation
  • building governance and capabilities
  • the technical foundation

A proven rule is to bring two to four applications into operation in the first year and to work on no more than six in parallel. More initiatives spread scarce business-unit capacity across too many places. The detail is set out in the AI roadmap for the first twelve months.

A one-day workshop works well as the entry point for executive leadership. It creates a shared understanding and prepares the first decisions, above all the objectives and the candidates for the first application. The format and outcome are described in the article on the AI workshop for executive leadership.

Which AI applications should a company implement first?

A company should first implement the applications that promise high value in a frequent process and draw on available data. Each also needs a business unit with spare capacity to take it on. If one of these features is missing, the application moves down the sequence, however convincing the technology looks.

Selection starts with a potential analysis. Over four weeks it maps the processes with high volume and recurring decisions. It ends with a prioritized list and the business case for the first initiatives. The approach week by week is described in the article on the AI potential analysis.

Four criteria have proven themselves for assessment, each scored on a scale of one to five:

  • Value: how strongly does the application change a result metric, such as cycle time or cost per case?
  • Feasibility: is the data available in usable form, and are the interfaces to existing systems accessible?
  • Risk: how serious are the legal consequences of an error, and what classification follows under the AI Act?
  • Ownership: is there a divisional head who owns the value, and are there users with time for the rollout?

The fourth criterion is the one most often overlooked. An application with high value and no owner remains a prototype. How the criteria can be combined in a weighted process is shown in the article on prioritizing AI use cases.

Does AI pay off for the company?

AI pays off when its value in the process exceeds the total cost of the application over its planned life, and that only shows in measurements from real operation. Research confirms productivity effects under real conditions: a field study of more than 5,000 customer service agents found that an AI assistant raised issues resolved per hour by around 14 percent, and by around 34 percent for less experienced agents. Figures like these prove the potential. They prove the value in your own company only once a baseline has been measured and usage in the process is secured. What further studies show on productivity, and why the effects often fail to reach the financial results, is set out in the article on productivity gains from AI.

A robust business case follows a fixed five-step calculation:

  1. Define the result metric: for example handling time per case in minutes.
  2. Measure the baseline: over at least four weeks in the real process.
  3. Calculate gross value: time saved per case, multiplied by cases per year and the hourly cost rate.
  4. Weight by usage rate: gross value only counts for the share of cases actually handled with the application.
  5. Deduct total cost: development and integration, licenses and usage fees, rollout with training, and operation with quality assurance over the life of the application.

An example shows the effect of the fourth step. If an application saves eight minutes on 60,000 cases a year at an hourly cost rate of 60 euros, the gross value is 480,000 euros. At a usage rate of 50 percent, 240,000 euros remain. The usage rate therefore decides the result more than the technical quality does. Time freed up only becomes money once the organization uses it, for example to handle more cases without additional staff. The full structure with cost items and assumptions is described in AI business case: calculating value and cost robustly, and measurement after launch in measuring the ROI of AI projects.

For the budget, a rule of thumb from delivery practice applies: over the life of an application, rollout and operation cost at least as much as its development. A budget that only covers the prototype funds half the initiative. How to build an AI budget with cost items and a reserve is shown in planning an AI budget. Where external support is added, it is typically billed in the German consulting market at around 1,000 to 2,500 euros per consultant day, depending on seniority. Billing models and cost drivers are covered in the article on the cost of AI consulting.

Governance, law and risk belong in one framework

Effective AI governance consists of a small committee with a few clear rules and an approval process aligned with the risk of each application. A committee of five to seven members decides faster than a large round. Alongside executive leadership and the business units, IT and the legal function belong on it, as does data protection. An application with low risk needs a simple approval; an application affecting staff or customers needs an in-depth review. How committees and policies fit together with clear responsibilities is described in building AI governance.

The legal framework in the EU is set by the AI Act. It sorts AI systems into four risk classes:

  • prohibited practices
  • high-risk systems with extensive obligations
  • systems with transparency obligations
  • minimal-risk systems without specific requirements

For most companies, three points matter. First, the regulation requires providers and deployers to ensure sufficient AI literacy among their staff. Second, the obligations follow the role, that is, whether a company provides a system itself or only deploys it. Third, certain fields of use fall into the high-risk category, including applications in recruitment and employment. The overview is in the article on the EU AI Act, and the high-risk category is explained in high-risk AI under the AI Act.

Beyond the AI Act, AI applications touch on data protection and copyright. Information security comes on top, and in companies with a works council, co-determination too. AI management ensures that these reviews come together in one process instead of holding up each application one after another. The framework for systematic assessment, from risk register to measure, is described in AI risk management: a framework for decision-makers.

For employees, an AI policy summarizes the rules on two to four pages. It names the approved tools and the data that may go into them. It also sets the review duties for outputs and names contact persons. A short policy gets read; a long one gets filed. Content and rollout are described in the article on the AI policy for employees.

ISO/IEC 42001 as a management system for AI

ISO/IEC 42001 is the international standard for AI management systems. It was published in 2023 and describes how an organization steers the development and use of AI, with policies and roles, with risk assessment and continual improvement. Its structure follows the harmonized structure of ISO management system standards that companies know from ISO 9001 and ISO/IEC 27001. An annex contains a catalog of 38 controls. The company selects the applicable ones on a risk basis and justifies the selection in a statement of applicability.

The standard is certifiable. What is certified is the organization's management system, not an individual model or application. There is no legal obligation to certify, and a certificate does not replace compliance with the AI Act. It does provide an audited structure around which those obligations can be organized. A clear rule applies to the decision. Certification pays off for providers of AI solutions and for companies whose customers or tenders demand proof. It also pays off where a certified ISO/IEC 27001 management system is already in place and can be extended. For everyone else, using the standard as the outline for their own governance is enough. Effort, benefits and the route to certification are described in the article on ISO/IEC 42001.

Technology decisions follow the use case

The choice of model and provider takes up the most time in committees and is at the same time the easiest to correct. Language models evolve quickly, and the differences in performance keep shifting. Companies therefore protect themselves with an architecture in which the model can be swapped, and avoid committing to one model for years. What decision-makers should know about the performance and cost of the large models is described in large language models compared.

The fundamental question is buy, adapt or build. For standard tasks such as drafting and summarizing, a bought solution is the right choice. For processes with their own data and logic, such as reviewing contracts against company-specific criteria, adapting a platform to your own data is the usual route. Building your own pays off where the application promises a competitive advantage that cannot be bought. The criteria in detail are covered in make or buy: developing, buying or adapting AI.

Five criteria count when selecting a provider:

  • performance in your own use case, tested with your own data
  • the terms on data processing and storage location
  • integration into the existing IT landscape
  • total cost over the term
  • the ability to switch provider later

A product demo is no substitute for a test with real cases. How procurement and the business unit select together is shown in selecting an AI provider. For companies with particular requirements on data sovereignty, the article on sovereign AI assesses European models and cloud offerings.

The portfolio is steered with metrics and rhythm

An AI portfolio is steered through metrics on three levels and a quarterly review. The first level measures usage per application, such as active users and the share of cases handled with AI support. The second measures impact on the agreed result metric. The third looks at the portfolio as a whole, above all the number of initiatives per approval gate and the time from idea to regular operation. The ratio of effort to realized value comes on top. Which selection belongs in reporting to executive leadership is described in the article on AI metrics for management reporting.

Every initiative passes through three approval gates with clear criteria. The prototype is approved when it shows a measured effect on the result metric in the real process with real data. The pilot is approved when the target group uses the application regularly without support from the project team. Regular operation begins when responsibilities are named and an operating budget is approved. From that point, an agreed procedure governs changes.

The portfolio review is where reports turn into decisions. It decides on continuation and scaling, and equally on closure. Closure in particular is a sign of good steering: an initiative whose assumptions are not confirmed in the prototype frees capacity for a better one. Whether AI belongs in the existing project portfolio or in a separate one, and how many initiatives should run at once, is covered in steering AI in the project portfolio.

Niehues Consulting as a partner for AI management

Niehues Consulting supports executive and divisional leadership in building and running this steering. The work ranges from business analysis and the business case to the selection of applications and portfolio steering. Marin Niehues is a management consultant with a delivery perspective: he builds organizations that deliver and brings AI into operation. His projects range from VC-funded scale-ups to DAX 40 groups, he has worked in 20+ countries on six continents, and his home base is the DACH region.

The approach follows the Niehues Delivery Model with the phases Understand, Build and Implement. For AI management, three of its review questions matter most. In Understand: which business result should change measurably? In Build: is the business case confirmed with real values? In Implement: which metric shows after 90 days that the solution is landing? Anyone who can answer these three questions for every initiative is steering a portfolio. Where structures and decision paths need to change, delivery management and organizational design complement the work, with Marin Niehues as delivery manager or product owner.

Criteria for selecting a consultancy in general are described in choosing an AI consultancy. The entry point with Niehues Consulting is the free 30-minute initial consultation via the contact form.

Frequently asked questions

What is ISO/IEC 42001 and do you need certification?

ISO/IEC 42001 is the international standard for AI management systems, published in 2023. There is no legal obligation to certify. Certification makes sense for providers of AI solutions and for companies whose customers demand proof. Organizations already certified to ISO/IEC 27001 can extend their management system with manageable effort.

What is an AI management system under ISO/IEC 42001?

An AI management system is the documented framework an organization uses to steer its use of AI, from the policy through roles and risk assessment to continual improvement. ISO/IEC 42001 sets out the requirements for it and contains a catalog of 38 controls from which the company selects the applicable ones. It is this system that is certified, not an individual application.

Which certifications should an AI vendor have?

An AI vendor should hold at least ISO/IEC 27001 certification for information security. ISO/IEC 42001 certification additionally demonstrates a management system for AI. For cloud services in regulated sectors and the public sector in Germany, an attestation against the BSI C5 criteria catalog is an established proof.

Who is responsible for AI in a company?

Overall accountability sits with executive leadership, which assigns it to one of its members or to a Chief AI Officer. The value of each application is owned by the divisional head in whose process it operates, and its operation by a named team. Both levels must be filled by name.

How do you get started with AI management?

The starting point is an inventory of all running initiatives and tools in use, followed by a potential analysis that assesses the candidates. This produces a sequence with approval gates and the business case for the first application. In parallel, executive leadership names the owners and puts a policy in force.

How long does it take for AI management to have an effect?

Steering with a portfolio, named owners and a policy is in place after three to six months. The first application with measured value in operation is realistic within twelve months, provided data is available and a business unit is ready to take it on. The full effect shows in the second year, when several applications run in regular operation.

How much does AI management cost?

The costs consist of the initiatives in the portfolio and the effort for steering itself. For the initiatives, rollout and operation cost at least as much as development over their life. External consulting is typically billed in the German market at around 1,000 to 2,500 euros per consultant day, depending on seniority.

Do mid-sized companies need AI management?

Yes, as soon as more than one application is in use or employees use AI tools in their daily work. In a mid-sized company, a lean form with one accountable executive and a short policy is sufficient. Add to that a list of initiatives with business cases and a quarterly review.

What is the difference between AI management and AI governance?

AI governance is the regulatory frame of committees and policies together with approvals. AI management is the broader steering that, alongside governance, also covers objectives and selection, business cases and the metrics of the whole portfolio. Governance is therefore one component of AI management.

Let’s talk about your initiative.

In 30 minutes we clarify where your organization stands and which step will pay off most now.

Request a consultation

30 minutes · free · no obligation