At the Rise of AI Summit 2020 in Berlin, Andreas Liebl, Managing Director of UnternehmerTUM and the appliedAI Initiative, presents a maturity model for how companies scale artificial intelligence and explains why most struggle to get value from it. He cites an MIT and BCG study finding that 90% of companies have invested in AI but only 40% report any business gains. Liebl argues the core difficulty is that AI differs structurally from traditional software: it depends on data rather than code, on training rather than development, and on ongoing evaluation rather than one-time verification. He walks through appliedAI's four-stage maturity model, from experimenter through practitioner, professional, to shaper, and uses a computer-vision quality-control example to show how requirements change as a single use case scales into an enterprise-wide, board-level topic. He presents appliedAI's newly published AI Journey Map, a tool for structuring the challenges companies face at each stage, and argues that scaling AI ultimately depends less on technology than on organizational ambition and board-level commitment.
Key takeaways
- Liebl cites an MIT/BCG study finding 90% of companies have invested in AI but only 40% report any business gains.
- He argues AI differs structurally from software because it depends on data not code, on training not development, and on evaluation not one-time verification.
- AppliedAI's maturity model has four stages: experimenter, practitioner, professional and shaper, and most German companies remain at the first two.
- A computer-vision quality-control use case needs progressively more infrastructure, governance and liability handling as it scales from proof of concept to enterprise-wide use.
- Liebl says about 90% of AI proofs of concept should be built with startups or outside providers rather than developed in-house.
Detailed notes
A four-stage AI maturity model
Liebl describes appliedAI's mission, as a non-profit tied to UnternehmerTUM, as helping companies and the whole country accelerate toward AI maturity. He lays out four stages: experimenter (running proofs of concept), practitioner (applying AI in two to three use cases but with low efficiency and replicability), professional (scaling AI across the organization to extract real value), and shaper (changing the industry's business models by becoming an AI-first company). He says most German companies, per surveys such as Bitkom's, sit at the not-started or experimental level, with very few reaching shaper status, which he frames as a problem for Germany and Europe's ambition to compete globally against the US and China.
Why scaling AI is structurally different from scaling software
Citing the MIT/BCG finding that only 40% of companies that invested in AI report business gains, Liebl identifies three structural reasons AI is harder to scale than conventional software: AI depends on data, not code, requiring reproducibility and auditability of data rather than just of programs; it relies on training rather than development, since uncertainty lies in finding a solution rather than in changing user requirements; and it requires ongoing evaluation and retraining rather than one-time verification, because of model drift. Citing a point he attributes to Andrew Ng, he argues that adding deep learning to an existing company does not, by itself, make it an AI company, any more than a shopping-mall website makes a retailer an internet company.
The strategy house and an AI-in-production example
Liebl structures appliedAI's approach as a 'house' covering vision and ambition, use cases, organizational enabling factors, and execution, applied differently at each maturity level. He illustrates this with a computer-vision quality-control case: at the proof-of-concept stage, the requirement is simply data and a working model to separate good and defective parts. At the practitioner stage, the challenge becomes system integration with cloud or on-premise infrastructure, a process for handling suspicious parts flagged by the model (including manual inspection and a feedback loop for retraining), and questions of liability if a defective part reaches a customer. At the professional stage, the same technology must be extended across many use cases (picking, logistics, robot supervision), which raises new issues of data pipelines, privacy and anonymization, bias, cybersecurity, auditability, benchmarking and system stability.
Five challenges from practitioner to professional
Liebl lists five challenges companies face moving from practitioner to professional AI use: rethinking AI business models rather than only products and processes; human-centered AI, since many companies have stopped building prototypes (some have 200-300) and now need to focus on process change, incentives and upskilling; moving from single use cases to use-case clusters and value clusters, because organizations can only manage a limited number of topics at once; building data pipelines; and moving from individual AI projects to an industrialization framework that scales AI reliably. He says appliedAI has published an AI Journey Map (available at appliedai.de) that structures these challenges across all four maturity levels to give companies a roadmap.
Ambition and board-level commitment
Liebl argues AI is ultimately a board-level topic, not one to be delegated only to experts, because the board must define the company's ambition for AI. AppliedAI published a report, 'Artificial Intelligence for Boards,' with input from more than 20 C-level executives on what starting to use AI meant for their companies. He cites a Chinese bank aiming for zero human operations as an example of an ambitious industry target, while noting that staying at the practitioner level is a legitimate choice for companies that do not aim to be global AI leaders.
Q&A highlights
Asked whether the scale of the AI journey map might discourage executives, Liebl says it reflects real, collected challenges from many companies and is meant to be realistic rather than encouraging or discouraging. Asked where the biggest problems occur on the journey, he points to the ambition/commitment gap at leadership level and to the lack of a working industrialization framework outside major tech platforms as appliedAI's main focus for the following year. On deciding which proofs of concept to pursue, he advises that in about 90% of cases companies should work with startups or other providers rather than build internally, reserving in-house development for the few use cases that touch core company knowledge. Asked about policy, he argues government programs focus too much on the earliest maturity levels and should instead help companies and startups reach the professional and shaper levels to build global AI champions.
Quotes
Our mission as applied AI, as a kind of a non-profit organization, is to help the whole country to get to an AI age by accelerating the journey to AI maturity.
Andreas Liebl
AI is not something that is done by the experts, but it's something that needs to be done at a board level.
Andreas Liebl
Questions answered in this talk
What is appliedAI's four-stage AI maturity model?
According to Andreas Liebl at Rise of AI Summit 2020, the four stages are experimenter (proofs of concept), practitioner (a few applied use cases), professional (AI scaled across the organization), and shaper (an AI-first company that changes its industry's business models).
Why do most companies fail to get business value from AI investment?
According to Andreas Liebl at Rise of AI Summit 2020, citing an MIT and BCG study, 90% of companies have invested in AI but only 40% report business gains, largely because AI requires treating data, training and evaluation differently from traditional software development.
Should companies build AI proofs of concept in-house or with outside partners?
According to Andreas Liebl at Rise of AI Summit 2020, appliedAI advises that in around 90% of cases companies should work with startups or other providers rather than build AI capability internally, reserving in-house development for use cases touching core company knowledge.