by Mary Branscombe

9 AI project mistakes to avoid

Dec 18, 201810 mins
Artificial IntelligenceEnterprise ApplicationsIT Strategy

From building isolated proofs of concept to not defining how to measure success, a wide array of gotchas can derail your AI project's prospects for delivering business value.

virtual brain / digital mind / artificial intelligence / machine learning / neural network
Credit: MetamorWorks / Getty Images

Business enthusiasm for AI continues unabated. IDC’s latest predictions say worldwide business spending on cognitive and AI systems — from chatbots to deep learning, plus the infrastructure to power them — will more than triple from the $24 billion forecast for this year to $77.6 billion in 2022.

More demonstrably, AI has gone from early adopters to mainstream business use cases, with a wide array of organizations across almost every industry exploring pilot projects and putting AI to work in production. But that doesn’t mean it’s foolproof to implement. If you don’t want to waste the money you’re going to spend on AI, here are some common mistakes to avoid.

[ 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. ]

Biting off more than you can chew

“Don’t try to boil the ocean on day one,” Lance Olsen, director in Microsoft’s Cloud AI team, tells You can’t transform your entire business decision-making process with AI overnight, so it’s best to start small and take evolutionary steps as you gain expertise.

Look for the low-hanging fruit. You need to develop a process for experimenting and validating the results of experiments before you tackle the most important systems. “Don’t necessarily go for your biggest investment right off the bat,” he warns.

Building isolated proof-of-concept systems

Building a one-off AI system that doesn’t help you create an overall process to do AI, and isn’t part of your existing data pipeline, won’t move you very far forward. You need to create a sustainable AI asset with each individual project. Here, sustainable means a system that generates enough ROI that you will keep investing in it to develop and scale it out further. Each time you do that, you help create an AI capability for the whole business, rather than just a new tool for one specific team.

Build on the business analytics you already do and turn those historical systems into predictions. “Start by making investments in optimization that use your existing pipeline and build on things you’re already doing,” says Olsen. Then you can move on to more revolutionary projects that make bigger changes to the way your processes work.

Starting without the right technology infrastructure

You need to be investing in both core and more advanced digital technologies before you start on AI, according to a recent McKinsey report. Companies that already have expertise in cloud computing, mobile and web development, big data and analytics are three times more likely to adopt AI tools. Three quarters of organizations adopting AI said they depended on what they learned from building existing digital capabilities. Or to put it another way: If your business isn’t ready to take advantage of cloud and data analytics, you’re not ready for AI either.

Starting without data

The vast majority of AI systems — certainly the ones that enterprises can build for themselves — are machine learning systems, and machine learning needs data. As Microsoft Corporate Vice President Julia White put it at the company’s recent AI in Business event, “Where’s my new bot? Well, what’s your bot going to learn from?” In fact, without good data, AI will hurt rather than help you, because you will have more confidence in something there’s no actual evidence for.

Moreover, if you only have the same public data as your competitors, you’re only going to get the same insights as your competitors, so you need to work with your organization’s own unique data. And that data is going to need cleaning and normalizing and preparing, assuming you’re even collecting the right data already.

Don’t underestimate the investment required; collecting and cleaning data typically makes up around 80 percent of a data scientist’s work. By starting with the data you already use for business intelligence and analytics, it’s also easier to make sure your AI system will support key business processes, making it far more likely to be useful. That should also help you define the tools and process for data preparation that you can use with data you’re not already using.

Not specifying how to assess and measure success

Data science is science. You need to have a hypothesis for what will improve business decisions, sales, customer support or whatever else you want to do with AI, and you have to test that in action and evaluate the results.

That means planning how to measure the success of a project — both in terms of adoption and outcomes. That can translate to aligning projects to employee deadlines, like the 90-day outlook for sales and marketing teams or the hourly quotas in a contact center. It also means having a control group that isn’t using the new system, which can seem counterintuitive if you’re investing a lot of money in developing it. You need to ensure that people are making data-driven decisions rather than relying on intuition; if they routinely ignore data, then having AI tools present it to them isn’t going to help. You also need to decide in advance what success will look like, because that’s the hypothesis you’re testing. Do you want more customer orders or larger orders? Do you want fewer customer support calls or faster time to resolution for the customers who do call?

Starting without knowing what problems AI can help you with

The problem with the term ‘artificial intelligence’ is that it can make it sound like anything is possible. The industry has made significant advances in the past few years, but you still need to know what AI can actually deliver and how it will integrate into your existing systems and business processes. Then you need to know what problems your organization has that AI could help you with. You can’t just adopt AI because you’ve read that all the other companies are.

“Executives need to consider two things before turning to AI,” Jacob Davis, senior director of analytics services at Cheetah Digital, tells “First, what are we actually trying to solve? How might we solve this problem right now and with the data on hand? If you can’t come up with something, even if it’s theoretical, within the realm of possibility for your current state, AI won’t help you. And second: Am I considering AI because of all the hype I’m hearing about it? You have to really evaluate your desire for these types of solutions because otherwise, you may invest a lot of money into something that won’t add real value.”

Starting without the right people in the right place

You’re going to need data science expertise and if you don’t have a dedicated data science team, that expertise will often be built up in the IT team. Wherever it is, it’s important not to keep it isolated in a single center of excellence. A recent Ovum study of Global 2000 organizations with AI projects in production done for data science software vendor Dataiku shows that to make projects succeed, those experts need to be involved with the business team whose problem they’re solving, as well as the project management and development team delivering the project. There are also cultural nuances in local business units that a central team may miss out on.

“Time and time again, we see teams in companies around the world and across industries that aren’t able to get their data efforts off the ground because they have no way for these people in different geographies — much less different types of people with different skills — to work together,” said Dataiku CEO Florian Douetteau. If you can’t get data science experts permanently based in key locations, use collaboration and knowledge transfer from central experts to build up local data science skills.

Trying to build your own AI capability for everything

While the much-publicized problems with IBM Watson underline that even pre-built AI services take time and expertise to integrate with your own systems and processes and have to be evaluated carefully, few businesses will have the expertise to build everything from scratch. AI tools are increasingly built into SaaS offerings such as Salesforce, Dynamics, and Adobe Marketing Cloud, although they may well be addons you have to pay extra for. There are cloud machine learning services from Azure, AWS and Google that offer either specific “cognitive services” such as machine vision and speech recognition that you can customize and build into your own tools and services, or galleries of common solutions that you can adapt for your own needs. Leverage these to get started quickly and then consider which other models and tools you need to build from scratch as workers get more comfortable with the productivity advantages AI can bring.

Expecting AI to do away with people

Like automation, AI will give you the most performance and productivity improvements when humans and AI systems work in collaboration. A recent study in the Harvard Business Review showed performance improvements of four to seven times as organizations adopted more and more human-machine collaboration. To get that collaboration, business teams need to be involved in assessing what AI systems will actually do for them. AI tools that provide recommendations, multiple options, decision support and escalation to experts for difficult cases are more useful than those that give simple yes/no answers without the involvement of any humans.

More on AI and machine learning: