The news that artificial intelligence is poised to take over the world has been greatly exaggerated. Sensationalistic headlines and moonshot futurists might have you thinking that robots will be rendering us all obsolete in the next month!
The truth is more measured. Of course, AI has become a large part of powering our day-to-day lives, but you simply cannot use it to make your most meaningful business decisions. For that, you still need humans. Our cognition can account for nuance and context, those "known unknown" and "unknown known" factors that a machine learning algorithm isn't equipped to process.
That doesn’t mean you can’t harness AI’s incredible power for your business. As you design AI-augmented intelligence, practice the guiding principles below to ensure you’re staying within the boundaries of AI’s current capabilities and that you don’t fall victim to trying to do too much, too fast:
Start with known knowns
Let's take two questions, both of which can hypothetically be fed into a machine learning algorithm:
Imagine you are combining through security footage, trying to track down a red Ford F-150 truck with a dent on the right side of its bumper.
Here are the inputs that you would need to answer the question “Is this vehicle I’m looking for?”:
- Is the truck red?
- Is it a Ford F-150?
- Does it have a dent on the right side of its bumper?
These inputs require a simple binary Yes/No answer that is unequivocally true 100 per cent of the time. Within these parameters, a machine learning algorithm can achieve relative accuracy (defined as 80 per cent+ accuracy 80 per cent of the time).
Now, let's take the question "Can I leave the office to get to dinner by 6PM?"
Here are some inputs to consider:
- How long will it take to complete my to-do list before I leave work?
- What does my afternoon schedule look like?
- What will traffic be like?
These inputs may seem relatively simple, but there are “known unknowns” or “unknown unknowns” that impact the answer. AI can scrub your calendar and see that you have a half-hour check-in with your boss at 4PM, but it won’t know that the meeting usually runs over by an extra 15 minutes. The inputs have too many dependencies to reach a useful level of accuracy.
Here’s an illustrative example of how a machine learning algorithm can go awry: A large university hospital was using machine learning to determine the likelihood of pneumonia patients developing complications, with the objective of reserving beds for high-risk cases. Based on the data, the algorithm recommended sending asthmatic patients home, unaware that these patients, susceptible to complications, were sent straight to intensive care as part of hospital policy. As a result, they rarely developed problems.
The machine did what it was tasked to do — interpret the data. However, absent of context, it came to a conclusion with dangerous implications.
Frame your business problems for machine learning
Accounting for AI’s limitations on contextual understanding, here are some general guidelines to help you determine which of your business problems can benefit from the the incorporation of
On the surface, each of the questions above requires a simple black-and-white answer — Yes or No. However, machine learning is much better equipped to handle the first question (“Is this the truck I’m looking for?”), because there are no “gray areas” when it comes to processing inputs and an algorithm’s data-processing power outstrip ours by a significant factor. On the other hand, you’re much more likely to hit an accurate mark with the second question ("Can I leave the office to get to dinner by 6?”), because you can intuitively fold in factors like your boss’s long-windedness or your energy levels at the end of the day.
- AI is best suited towards “soft” goals, such as improvements to operational processes or customer service. AI-derived insights are unreliable for producing immediate tangible benefits (like higher profits), although those may certainly arise as a byproduct. Steering away from these kinds of goals as the primary objective manages expectations and maintains focus on projects AI can bring utility to.
- Use problems for which the solutions can afford margin of an error. As aptly demonstrated by the hospital example, there should be wiggle room for AI to be wrong and for humans to implement appropriate course correction.
- Aim for augmentation over automation. Higher specificity with your queries — “How many years is Employee X likely to stay with the company?” — can provide insight in addressing larger business problems, like “How do I retain talent?” The first has more “Known Known” inputs (i.e. the average duration of an employee's stay; turnover rates per department; NLP Analysis for exit interviews or employee feedback), but human-driven analysis can account for “big picture” factors, like the influence of corporate culture and outlier events, such as a recession.
Perfect all things data
A non-negotiable prerequisite for effective machine learning is data — lots of data, preferably clean, labelled, logically structured, and with high points of variability. This should be coupled with sound strategy, methodology, and governance.
Achieving this requires strong multidisciplinary cooperation. An “out-of-the-box” AI solution simply does not currently exist in the ways that we want it to or the ways in which some claim; all the capabilities and expected results of an AI algorithm are baked into the code, and requirements are strengthened by leveraging multiple points of expertise. A McKinsey study on companies that successfully leveraged AI reinforces the importance of a diverse, agile team, “made up of highly committed business representatives, analytics translators, user-experience design experts, data engineers, and data scientists…[This] mitigates the risk of creating another isolated silo (such as design, digital) as the company builds its analytics capabilities."
Consistent application of analytical methodologies will also strengthen your AI efforts. Use Incorporate proven methodologies into your use case and use your analysis to iterate more effective training inputs for AI models — this will ensure that your AI is evolving in tandem with your business needs and processes, and vice versa, that your strategy can adapt with advancements in AI.
AI can certainly have an outsized impact on your business, but only if you approach it with a deeper understanding of its capacities and limitations. Don’t fall into the trap of considering it a magic bullet. Understand how to structure around AI’s strengths and use it in collaboration with human logic, creativity, and analysis. Your results will be more grounded, useful, and actionable for your business.
Aimee Lessard leads Signafire’s global analytic efforts
Image Credit: PHOTOCREO Michal Bednarek / Shutterstock