AI technology: When to build, when to buy

Artificial intelligence is fast becoming a business imperative. Whether it’s to improve efficiency, find new business opportunities, or keep up with — or get ahead of — the competition, companies across all industries are exploring the business benefits of AI, with AI adoption tripling in the past year alone.

For some companies, that means building AI systems from scratch. But finding the right talent is difficult and expensive, and, at 85 percent, AI projects have a high likelihood of failure, according to Gartner. Even when a project works, a commercial vendor might soon come out with something better, at a lower cost, with regular upgrades, more integrations, and a more intuitive UI. Or, your DIY AI sweat equity might be rendered superfluous when the new AI capability you’re working on is included as a free feature or upgrade to a platform your company already uses.

[ Cut through the hype with our practical guide to machine learning in business and find out the 10 signs you’re ready for AI — but might not succeed. | Get the latest insights with our CIO Daily newsletter. ]

Using a commercial product, on the other hand, can facilitate rapid experimentation with many different AI technologies, and minimal investment. And to succeed with AI, volume is important, says Rob Thomas, IBM’s general manager of IBM Data and Watson AI.

“I encourage clients to do 100 AI pilots,” Thomas says. “Not one, not two, but 100. Half of them won’t work, but the half that work can really pay off.”

There are commercial tools on the market already that are lightweight and take just a couple of weeks of investment, he says. In addition, the idea of embedded AI — AI that comes built into other platforms and systems — is taking off, he says.

Here’s what to take into consideration when deciding whether to turn to off-the-shelf AI solutions or to build your own.

Ensuring smooth travels

The Greater Toronto Airports Authority sees nearly 50 million passengers a year come through Toronto’s Pearson Airport, the largest and busiest in Canada. But of the 49,000 people that work at the airport, only 1,600 are focused on management, operational, and technology jobs.

One of those jobs is to take care of ticketing kiosks, used by passengers for faster check-ins. When a kiosk break down, or run out of paper, it sends an alert. Unfortunately, it took an average of an hour and 12 minutes from the time the alert was received to when it was back up and running again.

That was a problem, says John Thompson, the airport authority’s associate director of information services delivery. Knowing when a machine will break down, or run out of paper or ink, is not a simple calculation. Some machines are busier than others, or get different levels of use at different times of the day. So, the airport decided to look into intelligent analytics for a fix.

“With predictive analytics, we know when a machine is going to run out of paper, so we get there faster — even before the paper is out,” he says.

But the airport didn’t have the resources to build its own tools from scratch, he says. “We don’t build anything ourselves anymore. I don’t think anyone thanks you for developing your own software.”

Last year, the airport went live with the Symphony SummitAI cloud-based system for its internal IT support tickets, which took a couple of months to implement and configure for the airport’s workflow.

Thompson says that starting out with a small project proved to be a good way to get into AI.

“If you try to execute it as a big bang, it usually doesn’t work well,” he says. “Take it slow. Try to do one aspect of the technology at a time. My motto is, go slow to go fast.”

The rise of off-the-shelf AI

The Greater Toronto Airports Authority’s experience is representative of most companies’ forays into AI, as a lot of companies are opting to buy instead of build, says Gartner analyst Svetlana Sicular. “It is becoming very clear that do-it-yourself will not work. It is hard to find skills, and attrition is very high because everyone is hunting the same people.”

Meanwhile, platform vendors are increasingly embedding AI into their systems, making the technology available at a push of a button. Plus, vendors who build AI tools into their platforms already have access to an extremely large pool of very well organized training data.

Salesforce, for example, has a vast universe of labeled and categorized information that it can analyze for trends and patterns, and then make the most common or most requested analytics available to its customers.

Vendors also benefit from extreme economies of scale. They can afford to hire very specialized talent to develop and improve their AI models.

But it’s the data that’s really key, says Sicular.

“In the long run, the data is the biggest part of machine learning for companies,” she says. “That’s why companies like Google are so successful. They understand how to get data for machine learning and how to interpret it.”

Individual companies, by comparison, are limited to the data they collect themselves, or to what training data sets are available for purchase.

Commercial tools also offer other advantages, says Steve Herrod, managing director at General Catalyst Partners, a San Francisco-based venture capital firm, and former CTO at VMware. For example, it’s easier to find employees who are familiar with commercial tools than those who can use home-grown systems. Plus, vendors also offer free or low-cost training on their platforms.

“It’s always preferable to use off-the-shelf offerings when they’re available and sufficient for the job,” he says. “With each passing month, we’ll have more and more capable off-the-shelf AI software, leaving the build-your-own needs to a smaller and far more niche space.”

When to build your own

Buying off-the-shelf AI tools can be quick and convenient, but there are times when a company has no choice but to build from scratch. That includes cases when the data, or the models, are extremely sensitive or proprietary, or when commercial tools are simply not available.

For example, EnergySavvy, a software company focusing on the utility industry, has built proprietary algorithms to analyze utility customers’ energy use patterns.

“We have been servicing utility customers for close to ten years now, and that’s given us a deep understanding of how their programs are run, how they segment their data, and how to get insights from it,” says Kalpana Narayanaswamy, the company’s vice president of engineering.

Solving their problems requires an understanding of the inner workings of utility companies, she says. “And you have to know how to apply it. That’s where the magic is.”

To do that, the company has built a data science organization with a strong focus on industry expertise. The analytics platform itself is built on top of open source technologies, she says. As a result, EnergySavvy is able to go beyond the basic insights, basic targeting and basic customization available elsewhere, she says.

The AI component is also core to the company’s business growth, and is a key differentiator.

In general, when a company’s AI technology is a differentiator, it is hard to do that with a commodity, off-the-shelf system.

That was the case for Dialpad, a San Francisco-based provider of enterprise VoIP services. The company built its VoiceAI system from scratch, says Dan O’Connell, the company’s chief strategy officer, even though commercial speech recognition and natural language processing systems were available.

“We’d be using an API, which would be slower, less accurate and more expensive,” he says.

Dialpad wound up hiring its own natural language processing and speech recognition experts and data scientists, he says. “And a few people with PhDs in computational neuroscience.”

By building from scratch, the company was also able to offer unique features. In addition to real-time call transcription, for example, it also has live coaching, sentiment analysis and live smart notes and action items.

“It gives teams an otherwise untapped opportunity to take a scientific approach to understanding and acting on conversations,” he says.

Caveat emptor

Not every company needs to build its own AI technology, says Brandon Ebken, CTO at Insight, a Tempe-based technology consulting and system integration firm. “But the closer you are to your core business processes, with the potential to transform the customer experience, the more likely there is a need to do some type of customization,” he says.

And the sooner you get started, the more competitive advantage you can reap, he says. “I would not recommend that anybody sit and wait. AI is here for today. It’s no longer science fiction. Nearly all of our customers have some type of AI initiative. Digital transformation is being driven by AI — the customer who waits is going to be passed by their competitors, or miss out on some tremendous opportunity.”

Another use case that may require home-grown solutions is where there are privacy considerations, such as in the heavily regulated financial and health care industries.

For example, many commercial translation engines require data to be uploaded to the cloud, says Lauren Neal, principal at Booz Allen Hamilton. But government users, and those in regulated industries, want to be sure their data is secure. “They’d rather have an on-prem solution, to lock it down and make it safe. But there aren’t a lot of commercially available AI tools to set up for that mode,” she says.

It’s a sign of the infancy of this space, she says.

Other enterprises are also concerned about vendor lock-in, she adds. That’s especially true for companies that use built-in AI tools from platform vendors. “There’s a possibility that there’s less flexibility for them,” she says.

And because the technology is changing so quickly, being locked in to one vendors version of AI can become a disadvantage.

Best of both worlds

For most companies, build vs. buy isn’t an either-or decision. Instead, they buy when they can, and build when they can’t.

“We’ve certainly gone both directions,” says Reuben Firmin, CTO at ExecVision, a company that offers AI-assisted sales coaching for enterprise customers, including Workday, Ondeck Capital, and Paycor.

“We’ve used off-the-shelf tools for sentiment analysis, and built our own for speaker separation,” he says. “Where there exist off-the-shelf libraries, we didn’t reinvent the wheel.”

It’s important to do that research at the start of the project to avoid wasting effort, he says. “You can undertake a machine learning project and find out six months in that there’s a cheaper and more generally available way to tackle the same project.”

In particular, he warned, companies should be careful not to ignore traditional options. “Statistics can suffice for many projects,” he says. “And a lot of engineers are attracted to deep learning, but it’s not needed for every project that falls under the category of AI.”

When commercial tools are available, companies may need to do custom integrations, or add specialized code or configurations.

“There isn’t any solution off the shelf that solves all your AI needs and business challenges,” says Herb Hogue, senior vice president at PCM, an El Segundo, Calif.-based technology consulting company. “Most have a core framework, applications or SaaS products, but you still need to modify, enhance, or conform it to your existing business. That’s what we’ve seen.”

Most of the major cloud providers, including Amazon, Google, Microsoft and IBM, have platforms that make it easier to build customized machine learning models and AI systems, says Brian Atkiss, director of analytics at Anexinet, a Philadelphia-based digital solutions provider.

They also offer ready-to-go components and APIs for common functions, such as natural language processing, speech recognition, optical character recognition and chatbots.

Many companies have enough internal data for use cases, he says. “And the models will be more accurate using completely customized datasets for each use case versus generic and broadly available data used from off the shelf tools.”

Using a platform with a lot of built-in functionality allows the company’s development team to focus on the business process and user experience, says Richard Salinas, managing director for business automation at Sparkhound, a Houston-based digital advisory services firm.

“There is a common misconception that building an application from the ground up means ultimate flexibility,” he says.

Starting with a pre-built foundation gets you to market faster, he says. “And with the additional benefit of future-proofing the application by keeping the underlying technology decoupled and modular.”

More on AI and machine learning:

• 9 ways to avoid falling prey to AI washing
• 5 methods to adopt responsible generative AI practice at work
• Reskilling IT for the AI era
• How to know when AI is the right solution
• 9 IT projects primed for machine learning
• The future of ERP is AI
• 5 artificial intelligence trends that will dominate 2018
• 6 secrets of successful chatbot strategies
• How AI is revolutionizing business productivity
• Machine learning success stories: An inside look
• 9 machine learning myths
• 10 signs you’re ready for AI — but might not succeed