Last year, IBM announced a Data Science Elite team whose only job is to help big enterprise companies push their first AI models into production.

Now, more than a year after the program’s launch, Rob Thomas, the IBM executive overseeing the AI SWAT team, reports that it has been a “huge success.” The team has increased from 30 data scientists to 100, and there are plans to grow significantly next year. “We hire them wherever we can, actually,” Thomas said, noting that these data scientists operate all over the world. (Thomas told VentureBeat about this elite team as part of a wider-ranging interview.)

Companies as diverse as Harley Davidson, Lufthansa, Experian, Sprint, Carrefour, and Siemens used the team for a necessary kickstart on AI projects. And the best part: It’s all for free — or at least there are no contractual obligations to pay. IBM is betting the companies will like the products enough, and get so excited by the opportunities, that they’ll retain IBM to build out these projects — paying real money for IBM’s cloud, mainframe, or other services.

In all, the team has done 130 of these engagements, at 115 companies. The team completes these projects rapidly — in four to six weeks. That’s considerably faster than the “up to 12 weeks” timeline the group initially envisioned, partly in order to reduce cost, but also because IBM realized it could get results sooner. And those results are in the form of an AI model deployed as an API — allowing a company to integrate the AI into any of its applications.

“Within that time frame, you’re going to have a great AI success story and a model in production,” says Thomas, who is general manager of data and Watson AI at IBM. “These are some of the best data scientists in the world.”

How it works

Here’s the step-by-step recipe, according to Seth Dobrin, VP of IBM Data and AI, chief data officer of IBM Cloud, and a key architect of the team.

• CTO conversation: First, IBM’s data team strikes up a conversation with a company’s chief technology officer (CTO) to identify a real problem they’re trying to solve and agree on a scope for application.
• Design workshop: The elite team then starts a design-thinking workshop around AI. This is preferably for two days but could be just one. This is where the team meets with the company’s subject-matter experts to discuss the value of the business decisions the AI project will drive. It also ensures there is adequate available and quality data. A plan is drawn up, usually including three sprints.
• Sprints. This is where the heads-down coding happens. In a first sprint, engineers pull data and engineer features. In a second sprint, they develop the features and do a first pass at a model. The third sprint tunes that model and makes sure it delivers what the client needs. The group then deploys the model as an API. In most cases, IBM stops short of integrating these APIs with a company’s applications, though it did in a few cases integrate the API in Watson Assistant, IBM’s white-label virtual assistant.
• Visualization: Throughout this process, IBM also invites in a different team — data visualization experts — to design what the output of the project needs to look like so that it’s not just an API, but provides context for business users for how the AI can be used.
• Extension: Finally, after the previous steps kickstart the project, a client can take over or engage one of IBM’s services to take over, or there’s a third option, called a Build. This is a co-investment, where IBM agrees to cover half the cost for a 12-week extension of the project.

Examples of success

The team has helped companies in just about every industry, says Dobrin, from sports scheduling to call center optimization, fraud detection, and image recognition.

Dobrin says that almost all of these were considered successes — at least from the client perspective — because of the extensive up-front definition of the problem and the resulting recipe for deploying AI that customers walk away with.

Banking: A good example is Nedbank, a South African bank that wanted a way to predict which ATMs were about to fail and which ones to repair first. When IBM was finished, Dobrin said, Nedbank executives “felt like for the first time ever, they really knew how to implement machine learning and AI in the enterprise in a meaningful and valuable way.” The bank could then take the blueprint of what IBM showed it and replicate it in other areas of the business, he said.

Trading: Another example is JPMorgan Chase, which has advanced knowledge of AI and needed help using tools the bank already had, but in a different way: leveraging GPUs for a scoring engine. IBM helped with that project, which aimed to prevent the bank’s traders from making trades not recommended by their sophisticated models, said Dobrin.

Creative agency: Wunderman Thompson, the WPP-owned creative agency, had a challenge leveraging its data and 17,000 features to customize marketing campaigns for customers. It had been forced to use the same process template over and over. The agency finished an eight-week engagement with IBM’s team and “literally the next day turned it into a product they were selling to their customers,” said Dobrin. Now each client project is customized with data science, he said.

Lessons learned

IBM and its clients have learned some valuable lessons along the way, Dobrin says:

Importance of subject matter experts: In the Nedbank case, IBM dutifully delivered a model that would schedule ATM repairs based on which ATMs were most valuable. That direction came from the three Nedbank managers working with IBM. But then other bank groups argued that another metric was far more important to optimize. Traffic in Johannesburg was so bad that paying truck drivers cost the bank far more than anything that could be gained from servicing the most valuable ATMs first. Concludes Dobrin: “You really need the people who are doing the work in the room, and not, you know, second or third line managers.”

The simplest solution is best: In the Nedbank case, the bank’s rules-based fraud detection system was creating huge volumes of false positives. The bank consequently halted transactions that were otherwise legitimate, upsetting customers. When Nedbank executives asked IBM to replace the bank’s fraud-detection engine with AI, Dobrin said his team wondered aloud why they would replace something that was working. “Why wouldn’t we just focus on reducing your false positives?,” he said. IBM’s project reduced false positives by orders of magnitude, Dobrin says. “If there’s something that’s working and just needs to be improved, just start where you have it.”

Bring in visualization early: IBM learned to bring in data visualization experts earlier to build out an interface that end users could understand. Data science projects sometimes don’t do a good job of explaining why they’re useful, Dobrin says, and so it’s important to integrate them into an application or dashboard. When the elite team started, it hacked these dashboards together at the end of a project, since IBM had few data visualization experts on hand. But IBM learned to bring them in earlier. In one case, James Fisher, an oil and gas company, found the “beautiful” visualization tools the IBM team created to predict failures in transmission lines one of the more valuable parts of the eight-week engagement, according to Dobrin.

Need for a design-thinking workshop: IBM identified the need for an overall sync on a project’s design, beyond visualization, including what data is needed and where it resides. Before introducing the workshops, IBM would go into projects with clients, trusting them when they said they had their data ready. But often they didn’t. Worse, clients couldn’t ensure the quality of their data, and sometimes it wasn’t properly labeled. Twelve-week engagements would extend into 24-week engagements. “We had to put the data in order first, when really the client should have been doing that.”

AI essentials: Python, Spark, Tensorflow, and Kubernetes

Through the elite team deployments, Dobrin has come to see that companies working with AI need to master some essential tools and frameworks.

First, clients working with IBM’s team use three key tools: Watson Studio, which allows companies to build and train AI and machine learning models and prepare and analyze data in a hybrid cloud; Watson Machine Learning, the deployment engine for Watson Studio; and OpenScale, which adapts and governs the AI models over time.

But beyond that, successful teams need to be Python experts and know how to leverage market-leading tools like Spark and Kubernetes to accelerate and distribute processing of data with containers, Dobrin said. Distribution is important because successful companies need to be able to deploy on every cloud. “They’re on IBM, Google, Azure, Amazon, Salesforce, Workday, Concur,” said Dobrin. “And if it’s not containerized in some way, it’s just going to be untenable.” Clients also need a framework like Tensorflow if they want to do deep learning. Over time, IBM’s elite team has expanded projects to include Watson Discovery, a tool for AI feature identification. Moreover, several companies working with the team inserted their model outputs into Watson Assistant, Dobrin said.

Expansion

The Data Science Elite group is planning to grow next year, and is expanding the model into other areas, including data ops. “We’re proposing to grow the team fairly significantly. I don’t know the exact number just yet. But there is a plan to grow,” said Dobrin.