At Rise of AI 2019 in Berlin, Christian Kirschniak, a partner at PricewaterhouseCoopers who previously spent years working with AI at IBM, shares his experience taking AI from concept to real enterprise outcomes. He describes a typical pattern: a business person and a data scientist prototype an idea such as a recommendation engine, then struggle to operationalize it once IT, platform choice and data governance questions arise, often producing what he calls "POC hell," dozens of proofs of concept costing hundreds of thousands of dollars without production use. He argues the biggest recurring problem is sourcing the right data inside the enterprise, and that people, culture and governance matter as much as technology and data, including an underappreciated shortage of data engineers relative to data scientists. He highlights process-first thinking in European, brick-and-mortar companies as a barrier to treating data as an asset, and stresses accepting failure and selecting ideas by defined value rather than novelty. He summarizes these interlocking factors as an "AI molecule" and argues organizations must think holistically and involve everyone to succeed.
Key takeaways
- Kirschniak says AI projects commonly stall in what he calls POC hell, dozens of proofs of concept costing hundreds of thousands of dollars with nothing reaching production.
- He identifies sourcing the right data inside the enterprise as the single biggest recurring problem behind failed multi-million-dollar AI projects.
- He argues data engineers are undervalued relative to data scientists, citing a 20-to-1 ratio of applications received at PwC.
- He says European, process-first companies often treat data as a byproduct rather than as a core enterprise asset, which he sees as a barrier to AI adoption.
- He frames technology, data, people, culture, process and governance as an interlocking "AI molecule" that must all be addressed for AI initiatives to succeed.
Detailed notes
From a prototype idea to POC hell
Kirschniak describes the common starting pattern for AI ventures he has seen: a business person, such as a marketing manager, and a data scientist brainstorm an idea, for example customer segmentation or a recommendation engine, and the data scientist collects data, cleans and engineers it, trains a model and produces a result. Once the business side is satisfied, the team approaches IT to operationalize the model, triggering debate over which platform to use, whether a cloud provider such as AWS, Google, Microsoft or IBM, and uncertainty over whether non-standard tools are covered by company software licenses or what they do with the data. He says this often leads to what he calls POC hell, where companies produce twenty or thirty proofs of concept, sometimes mislabeled as pilots or minimum viable products, and spend hundreds of thousands of dollars without anything actually implemented in the enterprise.
Data quality, sourcing and compliance
Kirschniak identifies data as the next major hurdle: knowing which data was used, respecting compliance such as GDPR's restrictions on personal data, and navigating legal complexity that increases once datasets are combined. He says data quality has been a central industry problem for fifteen to twenty years, and that the single biggest problem he has seen is providing the right data sources to AI projects, citing multi-million-dollar projects that failed because data assumed to exist somewhere in the company could not actually be discovered or made usable.
People: sponsors, data scientists and data engineers
Kirschniak returns to his composite story of the marketing manager and data scientist, noting that the manager's sponsor had promised the board dramatic results, so a real 10% sales increase from the model felt disappointing next to inflated expectations shaped by Silicon Valley stories, such as predicting Oscar winners from open data. He argues that while data scientists are often called the sexiest job on earth, data engineers, who make models actually work inside enterprise systems, are undervalued and in short supply: at PwC, he says he receives about 20 applications for data scientist roles for every one application for a data engineer role, despite data engineers being essential to operationalizing AI.
Culture: process-first thinking versus data as an asset
Kirschniak argues that culture, not just technology, determines whether organizations can adapt to a shift he says will affect close to half of job descriptions. He says most brick-and-mortar companies he works with, particularly in Germany and Europe, think process-first, optimizing workflows and systems such as ERP, and treat data merely as a byproduct of process rather than as a core asset alongside production facilities, people and financial capital. He argues that without treating data as an asset, embracing AI becomes difficult, since data is the fuel AI applications run on.
Accepting failure and governing ideas by value
Kirschniak says accepting failure as a team, not just an individual, is rare in the companies he works with, contrasting this with entrepreneurial cultures like Silicon Valley, where many startups fail for every Facebook or Google that succeeds. On governance, he stresses selecting ideas wisely: enterprises rarely lack ideas, but struggle to turn them into reality, so each idea needs a defined value, not necessarily financial, set at the outset so outcomes can later be measured against it; he cautions against pursuing AI projects simply because they seem interesting.
The AI molecule and closing learnings
Kirschniak summarizes technology, data, people, culture, process and governance as interlocking parts of what he calls the AI molecule, arguing that neglecting any one of these in parallel makes an AI initiative likely to fail. He concludes that organizations massively fail to adopt AI for these combined reasons, that success requires holistic thinking rather than treating AI as only a technology, culture or governance problem, and that everyone in an organization needs to be genuinely involved, which he compares to being the pig rather than the chicken in a breakfast of bacon and eggs.
Quotes
What you get then is the so-called POC hell.
Christian Kirschniak
The biggest problems I've seen so far is to provide the right data sources to the data projects and AI projects in the enterprises.
Christian Kirschniak
Questions answered in this talk
What is "POC hell" in AI projects?
According to Christian Kirschniak at Rise of AI 2019, POC hell is when companies build twenty or thirty proofs of concept for AI ideas, sometimes mislabeled as pilots, spending hundreds of thousands of dollars without ever implementing a working model in production.
Why are data engineers important to operationalizing AI?
Christian Kirschniak explained at Rise of AI 2019 that data engineers, who integrate models into enterprise systems, are far scarcer than data scientists, citing a roughly 20-to-1 ratio of job applications at PwC, even though they are essential to making AI models actually work in production.
What does Christian Kirschniak mean by treating data as an enterprise asset?
At Rise of AI 2019, Kirschniak argued that many process-first, brick-and-mortar companies, especially in Germany and Europe, treat data only as a byproduct of business processes rather than as a core asset alongside production facilities, people and capital, which he says makes it harder to embrace AI.