Given these numbers, businesses are highly motivated to turn these projects into successful outcomes or risk falling behind the competition. In my experience, data science projects fail not because of poor technology or lack of available data. Nor is it because of a lack of talent or technical skills. It's because of the last mile problem. Data scientists are struggling to deploy their models into business processes and applications where they are used by the business.

Let’s talk about the typical process and then focus on where the breakdown typically occurs. A project begins with a business problem to solve. Generally, there are four main areas businesses want to improve upon with AI and/or data-driven analytics: profits, cost, customer experience and risk.

Organizations believe the answers to these challenges can be found in their mountains of data. They charge data scientists with unlocking the answers contained within the data via pattern detection, correlations, predictive signals, etc. The data scientist then builds the model. But what approach should be taken and what algorithm should be used? Are the correlations they’ve discovered relevant? Often, the data scientist is addressing these questions in a vacuum, away from the individuals who understand the business and who will need to trust the model to make decisions. And this is where the process begins to break down.

Data science is a team sport. Organizations can't bolt data science onto existing processes and expect success. Successful data science projects marry technical excellence with soft skills, such as collaboration and transparency, to form trust among decision makers and business users through the ability to communicate results in simple, yet powerful ways founded in a deep understanding of the business problem. As such, multiple stakeholders must be involved from the start. By incorporating collaboration throughout the process, better results can be achieved.

Once the problem is clearly identified by all stakeholders and the data scientist understands the "why," it is time to start building the right model -- one the business user can understand. That’s why it is important that data scientists collaborate with the business stakeholder to bring together technical know-how and domain expertise. Model creation is an iterative process where the business user can engage with the model and run what-if comparisons and simulations, validating the value and providing feedback along the way.

The model output must be put into business context and make sense to a business user if it is to be trusted and used. The business user also needs to employ that model in a system they understand, can operate and have access to -- often with visual aids and dashboards. It calls to question why we are seeing companies like Looker and Tableau being acquired for billions by bigger technology companies looking to shore up their data and analytics offerings.

With this context and understanding, the business user can then help validate the outputs and make sure they reflect the real-world situations and nuances that the domain expert innately understands. This collaborative process also helps the data scientist know how to tweak the model so it becomes smarter, more accurate and more effective. Over time, the business analyst needs to measure the outcomes and see how the predictions are trending. The iterative process between data scientist and business user continues throughout its use in order to improve results.

It’s important to stress that these aren’t technology projects that should be run by the IT department, though CIOs and IT staff will need to be involved in managing the technical aspects that underpin these projects. It’s also not about digging into data to find interesting nuggets. The business -- and all of its stakeholders -- must understand what the objective is and that data science is a way to predict outcomes and support decision making.

AI and machine learning can be very useful at crunching large volumes of complex data to find patterns and information that a human cannot. But AI only works from the perspective of the data and the function it was programmed to do. While AI excels at sifting data and reporting findings, humans are still needed to calibrate the output and make the final judgment call. I’ll end with an example that my colleague likes to use to illustrate how this might work.

Let’s say a 35-year-old, single male living in New York City making $125,000 per year receives an accidental deposit of $5,000 in his checking account. Do you predict he’ll return the money or keep it? What if I add that he’s lived in the same apartment for more than 10 years, has minimal debt and pays all of his bills on time? He seems like an upstanding citizen and, after synthesizing dozens of additional data points beyond what we mortals can comprehend, an AI system might predict that he would flag the bank’s error.

We could rely on the system to make that decision, but what if I told you that it was George Costanza from Seinfeld? If you’re a fan of the comedy, then you know the answer: Of course, George won’t return the money! As you can see, the model may inform the decision, but it shouldn’t make the decision.

Data science projects require significant investments, and businesses should expect tangible results. The investment in creating successful operational processes to drive end-to-end collaboration among all stakeholders is what will separate the winners from the losers.

source:https://www.forbes.com/sites/forbestechcouncil/2019/07/23/solving-the-last-mile-problem-for-data-science-project-success/#5c48981e5493