Generative AI in the Enterprise: A Review After Three Years

What has proven its worth are applications with a fixed place in the workflow, clear ownership in the business and a metric that shows their benefit. Three years after generative AI entered companies at scale in 2023, a clear pattern is visible. The difference between effective and ineffective initiatives lies less in the chosen model than in how the organization takes the application over. This review sorts the evidence and experience into criteria that senior management and heads of department can use for investment decisions.
How widespread is generative AI in German companies?
AI has reached the breadth of the German economy: according to official statistics, around 26 percent of companies with ten or more employees used AI in 2025, rising to 57 percent among companies with 250 or more employees. A year earlier the overall share stood at around 20 percent. Most of the growth reflects language models and assistants that employees can use without a dedicated development project.
Adoption, however, is not yet a result. Broad executive surveys paint the same picture: usage is high, while only a minority reports a measurable effect on earnings. That minority redesigned its workflows and built the application into the daily work of a business unit. The review after three years therefore starts with the question of what these companies did differently.
What has worked with generative AI?
Three patterns have proven themselves: narrowly defined tasks within the workflow, access to verified internal knowledge and support in software development. Where a language model takes on a clearly described activity, such as summarizing long documents or drafting replies, the benefit can be measured. Handling time per case and rework rate serve as metrics. The customer service field study also shows that experienced staff gain least and newcomers most, because the assistant brings the knowledge of the best performers into the workflow.
The second pattern concerns access to existing knowledge. Assistants that draw on verified internal documents shorten searches in functions with an extensive body of rules. Here the benefit stands or falls with the quality of the sources. Connect outdated manuals and you receive outdated answers.
The third pattern lies in software development, where assistants shorten routine work. The effect on a team's delivery performance only shows once review and release keep pace. Otherwise the bottleneck shifts to another point in the process. The article Produktivitätsgewinne durch KI covers the evidence and limits of these gains.
All three patterns share one feature: before launch, the organization decided who owns the result and how it is measured.
What has not worked with generative AI?
Three expectations have not been met: licenses for everyone as a strategy, pilots as proof and full automation without a review step. The broad roll-out of licenses has generated usage. Only a minority of companies achieves a measurable effect on earnings, namely those that redesigned their workflows. Without a described use case, employees use assistants for isolated tasks whose effect never reaches the team. The minutes saved spread across the day and appear in no metric.
Pilots follow a similar pattern. A pilot proves that something works technically. Proof that it works in the daily routine of the business only comes with the transfer into operations, with clear responsibilities and practiced routines. Why this transfer often fails to happen and how it succeeds is described in Vom Pilot zur Skalierung.
The third expectation concerned automating entire processes without human review. Where errors become expensive, a review step remains necessary. The organization designs it to take place within the lead time the process allows. A review step that slows the process down gets bypassed.
All three findings concern the organization of the users more than the technology of the models. Accordingly, the lever lies in senior management decisions on ownership and measurement, and on the sequence of initiatives. The article AI management: what executive leadership must steer describes how this steering is built as a whole.
Four questions for operational readiness
From a delivery perspective, the experience condenses into four review questions. An application counts as ready for operations when all four are answered with yes.
- Workflow: Is it defined at which point in the process the application acts and what happens before and after?
- Ownership: Does a named business unit own result and quality, and does IT secure operations?
- Measurement: Is the before-and-after metric fixed before introduction?
- Enablement: Are users trained, and is there a point of contact for questions?
The questions apply equally to every application, whether a sales assistant or document review in finance. An initiative that leaves a question open belongs on the program's work list and is not yet ready for roll-out.
When does generative AI pay off for a company?
Generative AI pays off when the measured time saved per case, multiplied by the number of cases and the cost rate, exceeds the running costs of the application. This calculation only works with a baseline captured before launch.
The fifth step decides credibility. Time saved without a defined use appears in no income statement. The article ROI von KI-Projekten messen shows how to calculate the value contribution in detail. For company-wide assistants the same calculation applies per role. For roles with a high share of writing and research it usually works out, for roles that are mainly operational it does not.
Maturity levels and decisions
A simple staging is sufficient for steering an AI portfolio. Every application sits at exactly one stage, and the move to the next is tied to conditions.
- Trial: A team tests feasibility with real data.
- Pilot: A limited user group works in the real workflow, and the metric is measured.
- Operations: Responsibilities and support are established, users are trained.
- Scaling: Further units adopt the application with an adapted workflow.
Each quarter, one decision is made per application: expansion to the next stage or termination. Remaining in pilot status without a deadline is not an option, because it ties up capacity and delivers no result.
Niehues Consulting as a partner for the transfer into operations
Niehues Consulting supports organizations along the Niehues Delivery Model with its phases Understand, Build and Introduce. In Understand, business analysis and the business case are established. They determine which applications receive priority in which order. In Build, the central review question is whether the prototype proves its benefit in the real workflow with real data. In Introduce, two questions decide success: who owns operations from day one, and which metric shows after 90 days that the solution is landing? Niehues Consulting transfers the solution into operations with internal or external teams and introduces it with clear routines and responsibilities, supplemented by user enablement.
Marin Niehues comes from the delivery side: he builds organizations that deliver and brings AI into operations. His projects range from venture-backed scale-ups to DAX 40 groups, with assignments in 20+ countries on six continents. The free initial consultation takes 30 minutes and sorts your company's applications against the four review questions: request a consultation.
Frequently asked questions
What has worked with generative AI after three years?
Applications with a clearly described task in the workflow have proven themselves, such as summaries and draft replies. Access to verified internal knowledge comes on top. A named owner and a metric fixed in advance are decisive.
How many companies use generative AI?
In Germany, official statistics show that around 26 percent of companies with ten or more employees used AI in 2025, and 57 percent of companies with 250 or more employees. Growth over the previous year stems mainly from language models and assistants. The larger the company, the higher the share.
Is generative AI worth it for businesses?
Yes, for applications with high case volumes and measurable time saved per case. The calculation works out when the extrapolated time saved exceeds the running costs and the freed time has a defined use. Without a baseline the benefit remains an assertion.
Is Microsoft Copilot worth it for every employee?
Rarely for all employees at once, frequently for roles with a high share of writing and research. Companies therefore start with clearly defined user groups, measure the effect per role and extend licenses afterwards. The prerequisites are covered in the article on introducing Microsoft Copilot.
How do you measure the ROI of generative AI?
Through one metric per application, fixed before launch, such as handling time per case or rework rate. Usage figures such as the number of active users show adoption but do not evidence value.
Why does generative AI not show up in the financial results?
Because the time saved is dispersed and has no defined use. Minutes saved by individual employees only appear in the income statement once a workflow is redesigned and the freed capacity is deployed deliberately. Companies with a measurable earnings effect did exactly that.
Why do so many AI applications get stuck in pilot?
Transfer into operations is a task in its own right that requires its own ownership and its own budget. User enablement comes on top. Without this mandate, the initiative ends with the prototype.
Are AI agents the next step after generative AI?
AI agents build on generative AI and take on multi-step workflows with their own sequence of actions. For companies the same prerequisites apply as before, only with greater weight: a clear workflow, a named owner and a review step where errors become expensive.
Who should own generative AI in the organization?
A named business unit as owner for each application. Senior management steers the portfolio through maturity stage and metric, and IT provides the platform and operations.
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 consultation30 minutes · free · no obligation



