At Rise of AI 2018 in Berlin, Ronny Fehling, Head of AI at BCG GAMMA, argues that companies unlock AI's business value by thinking big about the opportunity while starting with one to three concrete lighthouse projects. Drawing on client work rather than research, he contrasts rule-based expert systems with machine learning systems that derive rules from past outcomes, and stresses that most enterprise data is unstructured and still untapped. He walks through two real cases: a manufacturing client where natural language processing matched new quality issues against past written notes, and a general ledger project where clustering and NLP mapped hundreds of thousands of multi-language invoices to cost categories. Both cases reveal a common lesson: getting AI to produce actionable recommendations is only part of the challenge, since humans interacting with the system can introduce their own biases into it. Fehling closes by urging European companies to rethink entire value chains under AI rather than only automate existing processes, arguing Europe's manufacturing base combined with its data gives it a real advantage if used well.
Key takeaways
- Fehling argues successful enterprise AI projects start with a big vision but launch through one or two concrete lighthouse projects.
- At a manufacturing client, operators began agreeing with the AI's suggestions so often that its apparent accuracy reached 100%, revealing a feedback-bias problem.
- In the invoice-classification project, clustering reduced 800,000 invoices to about 40,000 groups, with 5,000 groups covering roughly 90% of total spend.
- Machine learning added only about 15% additional automation on top of the rule-based clustering step in the first version.
- Fehling contends most AI disruption destroys incumbents' existing business rather than improving it, favoring Europe's manufacturing and data.
Detailed notes
Thinking big while starting with lighthouse projects
Fehling frames the goal as unlocking not thousands but potentially a hundred million dollars of value from AI, while starting execution with one to three lighthouse projects that can build momentum and shift an organization's mindset toward data-driven thinking.
From engineering rules to data-driven learning loops
He contrasts traditional rule-based, engineering-built systems with machine learning systems that derive rules from past output data, and argues the goal should be actionable predictions that trigger a decision and generate more data, rather than dashboards alone. He notes that around 80% of enterprise data is unstructured, such as sensor data or documents, and remains largely untapped.
Manufacturing case: when the system trained the humans
In a client project, NLP matched new manufacturing quality issues against hundreds of thousands of past handwritten and machine-written notes, presenting operators with similar historical cases as decision support. The prediction accuracy started at 87% but climbed to 100% within months. Fehling says this revealed that, contrary to expectations, the machine had trained the humans: operators increasingly agreed with the system's suggestions even when wrong, feeding its own biased output back into the loop, which required adding guardrails.
Automating invoice classification at scale
For a general ledger project, the team needed to map spend to about 300 cost categories across roughly 800,000 multi-language invoices with a small team. By stripping invoices down to their most informative keywords and clustering similar entries, they reduced the data to about 40,000 groups; a Pareto analysis showed 5,000 of those groups covered close to 90% of total spend. Rule-based logic handled most of the remaining classification after human validation, with machine learning adding roughly 15% more automation on top.
Rethinking the value chain, not just automating it
Fehling closes with two points for succeeding with AI in Europe: organizations need data-driven thinking beyond isolated pilot functions, and AI disruption typically destroys incumbents' existing business rather than simply improving it, so companies should rethink their value chain rather than automate current processes. He points to Europe's strong manufacturing base combined with its data as a comparative advantage against the US, which outsourced much of its manufacturing.
Quotes
Machine learning is good to compare apples and oranges.
Ronny Fehling
Contrary to what we're thinking, that the humans actually train the machine, the machine actually trained the humans.
Ronny Fehling
Most of the AI disruption means destruction for the incumbent, and the customer, you are the incumbent.
Ronny Fehling
Questions answered in this talk
Why isn't a dashboard enough to get business value from AI, according to Ronny Fehling?
According to Ronny Fehling of BCG GAMMA at Rise of AI 2018, dashboards alone provide limited value; the goal should be actionable insights where the system predicts an outcome and prompts a decision, ideally generating more data for continuous learning.
What happened when a manufacturing AI system's prediction accuracy reached 100%?
Fehling describes a manufacturing quality-analysis case where operator agreement with the AI's suggestions pushed its apparent accuracy from 87% to 100%, which turned out to mean the humans were deferring to the system rather than the system genuinely improving, requiring the team to add guardrails.
How did BCG GAMMA classify hundreds of thousands of invoices with a small team?
Fehling explains the team reduced roughly 800,000 multi-language invoices to about 40,000 similarity clusters through NLP and keyword extraction, then found 5,000 of those clusters covered close to 90% of total spend, letting an eight-person team handle a workload competitors needed thousands of people for.