Delivery & Steering

Managing AI projects: How they differ from classic IT projects

Marin Niehues17 November 202511 min read

Man working at a pinboard with sticky notes

AI projects are managed through a measured business metric with a quality threshold, because their outcome only becomes clear in the real workflow. A classic program delivers what the requirements say. An AI program delivers results that are good enough in the real workflow at a defined level of reliability. Anyone who steers both in the same way sees green lights in the status report while the business unit does not use the result.

This article describes the four control variables that change compared with IT, places the familiar process models in context and sets out the criteria for continuing or stopping an initiative. How the steering of individual projects feeds into the steering of the whole portfolio is described in the article AI management: what executive leadership must steer.

What is the difference between an AI project and an IT project?

An AI project differs from an IT project in that its results vary in quality and can only be assessed on real cases. A classic IT project, whether delivered in waterfall or agile fashion, tests against defined requirements: a function either works as specified or it does not. Progress can be read from completed work packages or accepted features.

For AI applications, this logic holds for only part of the work. A language model or a forecasting method produces results whose quality varies with the data and the context of the task. The same input can produce different answers from a language model. Whether a solution is good enough is decided on real cases with real employees. The value can therefore only be demonstrated in stages, and steering has to recognize those stages.

Timing adds a further difference. In classic IT, a shortfall usually shows up in testing. In AI, it often only appears once the business unit works with the solution. Anyone who reports only development milestones until then is steering without the decisive information. Why so many pilots therefore never reach operation is covered in Why AI pilots do not make it into operation.

How do you manage an AI project?

You manage an AI project through four variables: a business metric as the goal, quality as a measurement series, data access as a delivery with a deadline, and operation as part of the project scope. Each of these changes compared with IT.

The first variable is the goal. An IT project defines a functional scope; an AI project defines a business metric, such as handling time per case or the share of results released without rework. This metric sits in the business case and is agreed with the business unit before the first prototype, together with a measured baseline. How a robust business case follows from it is described in Building a robust AI business case.

The second variable is quality as a measurement series. Before development starts, the team defines a set of representative test cases and a threshold above which a result is usable in the workflow. Every version is measured against this set, so that progress exists as a number.

The third variable is data access. In AI, data is a delivery with a deadline and an owner. A delayed export or unresolved access rights stop a project more effectively than any technical problem, which is why they belong on the critical path as dependencies.

The fourth variable is operation. An AI application keeps changing after launch: models are switched, inputs shift, and the cost per use fluctuates. Monitoring and cost control, along with responsibility for quality, are therefore part of the project scope and are planned before rollout.

Which process model suits AI projects?

AI projects are best served by an iterative process model with fixed decision points that explicitly covers data work and operation. Three models shape practice.

The best known is CRISP-DM, the Cross-Industry Standard Process for Data Mining. An industry consortium developed it in the late 1990s, and version 1.0 appeared in 2000. The model describes six phases: business understanding, data understanding, data preparation, modeling, evaluation and deployment. Its strength is that it starts from the business goal and builds in feedback between phases. Its limit lies at the end: ongoing operation with monitoring and further development appears only as a plan. Researchers have therefore extended the model for machine learning with quality assurance and monitoring, known as CRISP-ML(Q).

Scrum suits the development work in AI projects with one adjustment. The Scrum Guide provides for sprints of one month or less with a sprint goal. For user interface and integration, this works unchanged. For model work and data preparation, the outcome of a sprint is less predictable, because only the attempt shows whether a quality threshold can be reached. A proven approach is to take such work into the sprint as a time-boxed experiment with a predefined metric. The sprint goal is then to answer a question with measured values.

For steering by the executive team and the business unit, a three-phase cycle run per application is sufficient, and it embeds the specialists' models. In the Understand phase, the workflow is analyzed, the target metric is set and the business case is calculated. In the Build phase, a prototype proves its value in the real workflow; here the team works according to CRISP-DM or in sprints. In the Roll-out phase, the solution is transferred into operation with clear responsibilities until the business unit adopts it.

When should you stop an AI project?

Stop an AI project as soon as a predefined criterion is missed at a decision point and no measurement suggests improvement. Stopping is an explicit part of steering. An initiative ended early frees budget for applications that promise greater value.

The criteria are set down in writing before the start. Negotiating them only in the steering committee prolongs initiatives whose assumptions have long since failed to hold. How such decisions are distributed across a portfolio, and how applications reach regular operation, is shown in From pilot to scale.

Roles and decision paths

In classic IT projects, steering often sits with IT, and the business unit appears mainly as the source of requirements and the recipient. In AI initiatives, the business unit decides success, and it therefore fills the role of product owner. This person owns the target metric, prioritizes the scope and decides whether quality is sufficient for the workflow. The delivery manager steers timelines and dependencies across the initiative and keeps the business unit and the technical team connected with operations.

Decision paths are fixed in advance: who confirms the quality threshold, who releases budget, who stops. Without this, every question goes to a steering committee and waits for the next meeting. With it, the named person decides at the moment the measurements are available.

Planning duration and budget realistically

An AI project reaches operation in three to six months when data access is resolved, and in six to twelve months when deep integration into legacy systems is required. The binding plan comes at the end of the Understand phase, once target metric and data access are clear.

For the budget, a rule of thumb from delivery practice applies: rollout and operation cost at least as much over an application's lifetime as its development. A budget that covers only the prototype funds half the initiative. Where external support is added, it is typically billed in the German consulting market at around 1,000 to 2,500 euros per consulting day, depending on seniority.

Niehues Consulting as a partner for steering AI projects

Niehues Consulting in Düsseldorf supports AI initiatives from the delivery perspective. Founder Marin Niehues is a management consultant who builds organizations that deliver and brings AI into operation, in projects ranging from VC-funded scale-ups to DAX 40 groups and in 20+ countries on six continents, with his home base in the DACH region.

The approach follows the Niehues Delivery Model with the phases Understand, Build and Roll out. Three of its review questions matter most when steering AI projects. In Understand: what is the baseline, and how do we measure success? In Build: does the prototype prove its value in the real workflow with real data? In Roll out: who owns operation from day one?

AI management produces the business analysis with requirements and the business case, along with the sequence of applications. AI prototypes prove the value in the real workflow. Operationalisation and rollout take them into operation with internal or external teams. In delivery management, Marin Niehues takes on steering as delivery manager, project manager or product owner and designs roles and decision paths. The entry point is a free 30-minute initial consultation that places the target metric and data situation of your application in context: Contact.

Frequently asked questions

What is the difference between an AI project and an IT project?

An IT project delivers a predefined functional scope whose fulfillment can be tested. An AI project delivers results whose quality varies and only shows on real cases. An AI project therefore centers on a business metric with a quality threshold, and progress is measured instead of estimated.

Who should lead an AI project?

Business responsibility lies with the product owner in the business unit, who owns the target metric and adoption of the solution. A delivery manager or project manager steers timelines and interfaces. Both roles need clear decision rights, including the right to stop.

How long does an AI project take to reach production?

An AI project reaches production in three to six months when data access is resolved, and in six to twelve months with deep integration into legacy systems. The binding timeline comes at the end of the Understand phase. The prototype is built in short cycles so that quality can be measured early.

Is Scrum suitable for AI projects?

Scrum is suitable for AI projects when model work and data preparation are planned as time-boxed experiments with a predefined metric. For user interface and integration, Scrum works unchanged. The sprint goal for model work is to answer a question with measured values.

What is CRISP-DM?

CRISP-DM is a cross-industry process model for data analysis and machine learning, published as version 1.0 in 2000. It describes six phases from business understanding to deployment, run iteratively. For ongoing operation, the CRISP-ML(Q) extension adds quality assurance and monitoring.

When should an AI project be stopped?

An AI project should be stopped when a predefined criterion is missed at a decision point. Typical criteria are unresolved data access after eight weeks and a missed quality threshold after three iterations. A business case that no longer adds up with measured figures counts as well.

How much does an AI project cost?

The costs comprise development and integration, licenses and usage fees, rollout as well as operation over the lifetime. Rollout and operation cost at least as much as development. External support is typically billed in the German market at around 1,000 to 2,500 euros per consulting day.

Why do AI projects fail?

AI projects mostly fail on organization and less often on technology. Common causes are a missing target metric with a baseline as well as unresolved data access. Often there is also no business unit that owns the value, and no one has budgeted for operation.

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