Erfahrungen herausragender globaler Unternehmensführer bei der Einführung von Daten und Künstlicher Intelligenz

Start with the smallest unit of the "data value" you can imagine. That is a decision. Someone is about to perform an action, and a little better information would make that choice a little better. That is the entire atomic value.

Every data platform, every governance network, and every artificial intelligence program is just a mechanism developed to make this moment more frequent, more cost - effective, and more reliable.

Listen directly to the experiences of leaders who are building and expanding data and artificial intelligence capabilities in some of the world's most prominent organizations. These organizations include Dasa, one of the largest companies in diagnostic medicine and clinical analysis in Latin America; United Talent Agency, one of the most influential companies in the global entertainment industry; Kotak, one of the largest investment fund companies in India; and even the White House.

Point 1: The quality of a decision depends on its context.

Before considering a platform or governance, ask yourself: What makes one decision better than another?

Nehhaa Purohit, the Chief Digital Officer of the United Talent Agency, has shared some interesting experiences because she has paid a high price for it.

Although her dashboard showed that everything was okay, with an availability of 99.99% and a latency of less than 50 milliseconds, in fact, the revenue per inference decreased by 2% per month over a six - month period. A loss of $14 million went unnoticed because the intelligent system was working perfectly, but the assessment was wrong.

Availability measures whether the pipeline is unobstructed, not whether the data transmitted in the pipeline is still valid. Purohit calls this gap "context debt": The data is still technically valid, but semantically outdated and gradually losing its validity, and all monitoring devices still show "okay".

A decision is "good" because the data still matches reality at the time someone acts: This is more difficult to ensure than availability, and almost no one measures it.

Point 2: A good decision has no cumulative effect, but repeatable decisions do.

A well - thought - out decision is nice. But if you want to repeat this decision within an entire team or a quarter, you will encounter the real problem that every leader has to deal with: The trade - off between speed and trust.

Animesh Kumar, the Chief Technology Officer of The Modern Data Company, says that the usual interpretation of this phenomenon is a false conflict. Corporate governance is classified as "defense", which means it is at the top of the cut list when the company needs money for its development.

This is completely wrong because unregulated data slows down the growth rate. No one trusts data that is not traceable, so they verify it manually again, which adds a hidden cost layer to every subsequent decision. The solution is to consider data governance as a standard attribute of the platform, rather than an additional payable project.

Mohamed Amin, the Vice President for digital transformation at Hosta, has the same problem, namely decision - making inertia: the delay between insight and action. His solution is clarity, prioritization, and accountability: Replace raw indicators with signals that can be directly applied to the business, and detailed reports with a few key performance indicators (KPIs) that can drive decisions, and assign a responsible person to each indicator.

Different words, the same equation: Repeatability means removing the resistance for the next decision, not just providing information for the current decision.

Some people need faster answers than the official channels can provide, so they develop their own solutions. Kumar's solution is to break the problem down into structure, behavior, and incentive mechanisms, because if the official channel is slower than the workaround solution, the workaround solution will always win.

Point 3: A repeatable decision only makes sense if it affects the key performance indicators that the Chief Financial Officer is interested in.

Even a well - thought - out and repeatable decision is not a strategy, because senior management doesn't care about repeatability itself, but only about profit and loss.

Dia Adams, the Chief Data and AI Officer of Datafolx AI and former corporate data strategist at the White House, comes to a clear conclusion:

Stop talking about technical indicators and start talking about business results.

Instead of saying: "We have built a data lake", you should say: "This has achieved $X in revenue through high personalization and reduced customer churn by Y%." She believes that operating income should be adjusted, because EBIT (Earnings Before Interest and Taxes) includes both revenue growth and cost reduction, and the board already trusts this number.

Jon Cooke, the Chief Executive Officer of Nebulyx AI, answers this question with a defensive and offensive - analogy. Defense refers to the essential governance work in regulated industries: A versioned, citation - specific knowledge base means that auditors can know with a question: "Why did you make this decision in March?"

The offensive strategy is a combination of different use cases, each of which is linked to a directly measurable indicator. These use cases build on each other, because clean data can eliminate quality problems, which speeds up the offensive and avoids the slowdown of the next use case.

Justin York, the head of Rubicon, questions the entire premise. When he was asked which element of the profit and loss statement should be adjusted, he answered: Take the credit carefully, even if you can't trace the exact cause. When sales increase, is it really the credit of the data strategy or does someone still have to "make the sale"?

Normally, it is difficult to really evaluate the return individually because data is involved in all areas, and the greater trust risk is that leaders over - price "solutions" that collapse in the face of actual workload and difficulties. Adams and Cooke have built a trustworthy financial history; York avoids over - exaggeration. Both approaches are correct, the key is whether the numbers are traceable.

Point 4: Nothing can be scaled if no one uses it.

Even a well - founded, repeatable decision linked to profit and loss can fail in the last mile: If non - technicians have to act based on the available insights, the decision fails.

Gabriel Vernalha Ribeiro, who is responsible for data governance and artificial intelligence at Dasa, believes that the problem should start at the data generation stage. His solution is process improvement: All projects must have a data collection plan and success indicators from the start, otherwise they will not be approved.

Structured data from the start "is no longer a maintenance cost, but an always - available asset". His principle for the last mile is: Provide insights in the tools that people already use, and replace the diagrams to be interpreted with simple "what to do next" recommendations.

Cooke's "decision capsule functions", developed independently by UTA's Purohit, go one step further: They display a single user interface in the workflow environment that shows the recommended action, the confidence interval, and the main drivers.

Purohit's version gracefully reverts when the context is outdated. It hands the decision - making power back to humans instead of confidently giving a wrong answer. This design choice has avoided an estimated loss of $8 million.

Ribeiro's final statement sums it up: A culture cannot be forced, it must be nurtured. A team really takes responsibility only when bad data causes obvious losses to the team and is linked to the team's own key performance indicators (KPIs).

Point 5: The system can only be maintained if managers think the same way.

The most surprising episode in this series has nothing to do with data at all, but with whom you hire to operate the system.

Purohit's method of selecting candidates is the clearest implementation of first - principles thinking. She divides candidates into pattern - recognition experts (good in stable states) and first - principles thinkers (good at breaking down highly unclear problems from the ground up).

Her test: Diagnose a system where the dashboard shows everything is okay, but the revenue per inference is silently decreasing. Candidates who only focus on latency and error rate are eliminated. They seek certainty in probability questions. Those who try to fix the semantic deviation pass the test. One manages the system, the other creates value.

Cooke describes the same change in terms of artificial intelligence agents: The Chief Data Officer is no longer the librarian who organizes known knowledge, but the architect of a limited decision - making space. In these areas, the agent acts under intelligent constraints.

Adams calls this a "business architect" instead of a "data installer". The terms are different, but the intention is the same: The task is no longer to keep the pipelines unobstructed, but to ensure the accuracy of the assessment when decisions are made based on the available data (whether from humans or systems).

This article is from the WeChat account "Data Driven Intelligence" (ID: Data_0101), author: Xiaoxiao. Published by 36Kr with permission.