When it comes to the future of doing business, AI and machine learning are the hottest topics around. For all the discussion of AI’s promise, though, something is still lacking — straightforward information about how to make AI work for you. If you approach your solution in the wrong way your results won’t make sense, but if you go about it right you can build a performance system that moves your business forward. We’ve guided many organizations in their successful adoption of AI-based solutions — here is some practical advice about how to approach the challenge.
We’re talking about business process automation
AI solutions often involve taking repetitive, time-consuming tasks and developing methods to complete them more efficiently. This means that often, AI is a form of Business Process Automation (BPA), except the process you are automating is informed human judgment. BPA usually addresses predictable, highly repetitive tasks, but advances in computing have dramatically expanded the scope of what can be automated — including things like automatic processing of texts and images. Still, thinking about AI solutions in this way can help you structure your adoption approach on more familiar ground.
Work on automatic text processing began half a century ago. When the US Postal Service started using Optical Character Recognition (OCR) to help sort the mail, it was traditional BPA through and through — machines were programmed to recognize characters in portions of typed address labels, and letters were routed as a result. It was the dawn of modern AI, though, that really kicked things into high gear. Now we can build systems that are able to do really impressive things like understanding the meaning and sentiment of texts, and recognizing objects in images.
Long before the advent of modern AI, the US Postal Service was using a form of text processing to help automate mail delivery. Image source: this amazing video from 1970.
Remember, though, your AI system isn’t doing anything mystical. It’s all about learning to recognize patterns in data, and using those patterns to automatically process other data — much faster, and at much greater scale. The key difference between AI-powered automation and BPA more broadly is that AI is trained, instead of built.
Learning takes time. Be incremental.
You probably wouldn’t hire an inexperienced person and expect them to do a perfect job right away. Nor would you probably assign them the most complicated task first, when you could start them off with something more straightforward. Successful onboarding requires adequate training and reasonable expectations. Machine learning is learning too — your AI is only as good as the data it’s been trained on — so this wisdom still applies.
Getting the most out of your data is a process of incremental improvement. Image source: toptal.com
When we think of automation, we often think of fast, hyper-accurate processes that work with the flip of a switch. It’s understandable, then, that people approach AI with similar expectations — especially when it comes to familiar tasks like understanding text and images. Remember, though, that these are complex tasks and there’s a learning curve with your AI solution too. The best results can take time — but that doesn’t mean you can’t get value out of your system now.
Instead of expecting total automation out of the gate, consider introducing AI to your process incrementally. By zeroing in on your AI system’s most confident predictions, which will be the most accurate, you can quickly automate away parts of your work with confidence. Doing this also helps you measure just how much work you can automate, and what is best left to humans. For example, while almost all US mail is now automatically sorted using OCR, humans are still on hand to deal with particularly messy writing and incomplete addresses.
Adjusting your process structure may also help you benefit from AI. I’ve worked with many organizations that wanted to classify their data in very fine-grained ways, but often, they got more out of AI by focusing on coarser-grained questions like separating actionable and unactionable complaints, or identifying whether tweets containing a keyword were relevant to their brand. When the Postal Service began automatically sorting letters, they started with ZIP codes instead of whole addresses. Start simple. Partial automation can still lead to big gains.
There are tradeoffs
Assembling a component involves the same steps whether it’s done by human or machine, and the machine can do it faster. However, using AI to automate tasks like text classification or image analysis isn’t quite the same situation. Humans have minds, whereas machines just look for patterns in features they extract from the data.
What that means is that machines can behave differently from their human counterparts in how they classify things. Often, this improves results but sometimes it can also lead to errors that a human wouldn’t make. Don’t let that erode your confidence in AI — at least not necessarily. Your system probably also got some things right that humans would have missed. Remember that learning is an incremental process, and by focusing on your system’s high-confidence predictions, you can probably still automate a portion of your data processing with human-level accuracy, and with the right data your model will improve over time.
Adopting AI also, almost certainly, means a degree of behavior change. Roles may shift, processes will be altered, and your results might look a bit different. Organizations are often reluctant to change behavior, because there’s a break in continuity and an up-front cost of adapting. But it’s worth clearing this hurdle because you will adjust more quickly than you think, and the positive benefits of automating your labor-intensive process will more than offset the cost of switching. In fact, it will probably even open up opportunities that didn’t exist before.
Adopting new technology involves some tradeoffs, but the benefits likely beat out the costs.
In evaluating a human employee, it might make sense to spot-check their work — select a handful of recent judgments they made, and see if they look solid. Based on this, you can make pretty reliable inferences about how good the rest of their work will be. While this might be a useful strategy in evaluating human employees, there are reasons to be careful using it when evaluating the performance of your AI system.
First, because your AI can operate at a much greater scale than a human, your small handful of examples is less representative. Second, because machines and humans reason differently, you should evaluate them differently. Third and most importantly, too much spot-checking can cause you to fixate on certain isolated errors. I’ve seen organizations that were new to AI lose faith unnecessarily because they encountered one or two “inexplicable” errors, even though their system overall was performing fairly well.
Rather, consider evaluating your classifier at a larger scale. One option is a technique known as cross-validation; another is to create a separate, large test set of items you hold out for evaluating your model’s predictions. Are the high-confidence predictions accurate? Does the model’s performance improve over time as you collect more training data? Does it excel or struggle with certain types of examples? Answering these questions using a larger dataset will help you evaluate your AI system in a much more useful way.
Feature importance for a sample model in the Appen platform. The terms in green predict negative sentiment about airline experiences, while the terms in gray predict the opposite. The presence of some odd items reflects the limited amount of training data in the model.
Additionally, ask why your AI system made the predictions it did. As we mentioned before, AI works by taking an input and extracting features from it (like word sequences in text or patterns of color in images) and using those features to identify newly-encountered items. By understanding what leads your model to make the predictions it does, the model becomes less of a black box and you may even identify opportunities for improvement. If you really want to dig deep, this paper is probably the best discussion of the topic I’ve ever read.
Maintain and improve your models over time
Accurate data classification is often a moving target. Your data changes over time, and so do the questions you are trying to answer. A news story that was relevant to your company a year ago may no longer be important today, and what was once a low-priority customer support issue may now be critical. To capture these changes, and maintain the accuracy of your AI system, you need to commit to ongoing evaluation and maintenance.
The most important ongoing maintenance task is keeping your training data up-to-date. You’ll want to provide your classifier with fresh data to make sure that new trends are accurately captured and available to the model. You’ll also want to make sure that the labels being applied to your data are consistent, and that when things change, out-of-date items in the training data are either updated or removed so that you don’t confuse your model.
Appen’s human-in-the-loop workflow supports three of the most important concepts discussed here: focus on your high-confidence predictions; combine human and machine intelligence to get the best results, and iterate the development of your AI model.
A key component to this process is human-in-the-loop learning, a process of model training where the model asks for human input on just those items it’s not confident about — meaning that human involvement becomes as efficient as possible, and the model gets exactly the data it needs. Human-in-the-loop learning drives Appen’s data annotation platform, which we built specifically to make getting started with enterprise-grade AI solutions as efficient as possible. By combining a machine learning environment with a world-class data enrichment platform, Appen makes it easy to combine the best aspects of human and machine intelligence at scale.