AI is still new to enterprises. As with any new technology, companies will face many challenges as they attempt to integrate AI into their existing operations and workflows. Yet, just because AI adoption can be difficult doesn’t mean it has to be. There are a number of common mistakes that businesses make with AI both on an organizational and technical level that can be avoided so that incorporating AI is far less of a headache. Here are five best practices that help you avoid the mistakes people most frequently run into when implementing AI.
1. Identify a high-value problem the business is looking to solve
To bring true value to an organization, when AI is first being implemented, it should be used to address a specific problem. This is true regardless of what type of team is working on the issue.
We generally see three different ways that data science teams are deployed:
- A centralized model under a chief data officer or lead analyst where there is then visibility across the whole organization.
- A diffused model in which each business unit has its own dedicated group of data scientists that are there to augment that work of that unit and that unit alone.
- A hybrid model, in which a centralized group establishes a competency center in which they build platform and reusable assets that are then dispersed to data science teams within individual business units.
There are obviously significant differences to all three of these approaches, but all are most successful with AI when they’re solving that’s been deemed important within the business unit or the business as a whole. With AI changing the art of the possible, it can be tempting to rationalize trying out new technology without tying it to business value. As Whit Andrews recently wrote in a Gartner report, “Rather than their experimental or “cool” value, AI projects and pilots should earn their priority based on the needs of organizations considering them.”
Some examples of the types of questions companies could use AI to answer:
- How do we improve uplift with our take rates on advertisements on our mobile channel?
- How can we counter disruptive competitors attempting to capture our customer base?
- How do we optimize spend in a particular area of the organization?
Regardless of the question itself, it must be a targeted problem on which there is universal agreement that it merits solving.
2. Ensure you have sufficient data
AI is now more practical and powerful than ever before because companies have so much data at their fingertips. From IoT to big data generally, AI works best when it is fed by copious amounts of data. This type of data reserve is what allows the algorithms to detect the strongest patterns and correlations, and therefore offer the most useful insights.
But just because there is more data in general doesn’t mean that companies always have enough data to answer the specific business question they want addressed. For AI to be effective, it must have access to a wide array of data, both structured and unstructured, that lives within the enterprise but also may be blended with commercially available data, or data generated from social and mobile. Deep learning requires a depth of data; a general rule of thumb is to have at least 10,000 rows of data before putting it into an AI learning model.
This diversity and volume of data ensures that the results AI pushes out are credible and not just aberrations or overfitted results from tiny data sets.
One cautionary note: you might have the volume of data you need but face more work than expected to get it in a usable form to support AI initiatives. Forrester’s Michele Goetz recently wrote that “firms taking the AI leap typically hit immediate data roadblocks as they attempted to bring all of the relevant data together and curate it to ensure quality.”
3. Define the outcome – What does success look like?
It’s generally good business practice to never start a new initiative without first being clear on the outcome you hope to achieve. This is especially true for AI. Too often, we’ve seen companies embark upon AI without knowing the end result they’re hoping for. You want to ensure you have a proof of concept or proof of validation prior to starting an AI-based project so you can demonstrate the feasibility of the work, and be able to quantify the results once everything is completed.
Outcomes can range from general to highly specific. You could be trying to augment a team’s decision making or trying to improve an operational process. You could be trying to replace parts of that process to drive automation. Whatever the goal, having a broader strategy in place so that the AI analytics and techniques can be applied effectively is crucial.
4. Measure your results
Once you’ve identified your problem and have the strategy in place to begin using AI, you also want to ensure you have the right instrumentation to measure the results. This can be an analytic platform or something less concrete, but the only way to make AI success repeatable and applicable to the highest value use cases, is by establishing this type of baseline measurements. It’s the proof points for everyone in the business that AI is actually working.
5. Have a way to operationalize recommendations
Too often, people start out looking at the feasibility of applying techniques like LSTMs or RNNs but don’t pay attention to the larger operational context. What are the business processes? Who are the people involved? Are you trying to augment their decision making? Are you trying to replace aspects of that operational process and drive more automation? Are you in a highly regulated environment and you need to really focus on the explainability or interpretability of these models because there’s an audit and compliance aspect to them?
It’s important to think about the people side of AI, the impact on processes, and see how it will work in practice for the results you’re trying to drive. How will you manage change in these areas?
The practices just described can move AI from the realm of a science experiment that meets feasibility criteria to a technology that serves as a force multiplier for your team’s efforts and a competitive advantage for your business.
Atif Kureishy – VP, Global Emerging Practices | AI & Deep Learning at Think Big Analytics, a business outcome-led global analytics consultancy.
Based in San Diego, Atif specializes in enabling clients across all major industry verticals, through strategic partnerships to deliver complex analytical solutions built on machine and deep learning. His teams are trusted advisors to the world’s most innovative companies to develop next-generation capabilities for strategic data-driven outcomes in areas of artificial intelligence, deep learning & data science.
Atif has more than 18 years in strategic and technology consulting working with senior executive clients. During this time, he has both written extensively and advised organizations on numerous topics; ranging from improving the digital customer experience, multi-national data analytics programs for smarter cities, cyber network defense for critical infrastructure protection, financial crime analytics for tracking illicit funds flow, and the use of smart data to enable analytic-driven value generation for energy & natural resource operational efficiencies.