One of the biggest issues when applying AI to business is that most executives have a misconception of what AI can and can’t do.
It’s simple to speak about AI on an abstract level, but when it comes to actually integrating deep technologies into existing workflows and processes, it’s easy to get lost.
In this blog we cover four important action points to consider when determining your businesses’ AI needs and how to move from ideation to action.
Many companies are eager to embrace the idea of employing AI, but fail to pinpoint clear objectives for the use of these technologies. Only having an abstract view of what AI can do is a surefire way to be disappointed.
It’s important to recognise that there is no single AI technology that will be the end-all and be-all to solve your company’s inefficiencies and will surely not reinvent your company’s strategic goals. One new technology shouldn’t overhaul all of your enterprise’s major functions or direction.
On the contrary, AI is most effective when it’s dealing with very well-defined, narrow circumstances. When considering what you want new technologies to achieve, you must narrow your scope and identify the specific business objective that you want achieved.
For example, cutting document processing time by 50%, or decreasing insurance claim processing time from 1 week to 1 day, are good examples of a clearly defined objective. Increasing sales or hiring better qualified staff are far too general.
A good way to think of this is that the current machine learning and deep learning systems can perform only one task very well. For this reason, it’s best to think of what tasks AI can take over, as opposed to jobs or activities. For example, AI can perform customer profile analysis and recommend financial products or services, whereas it can’t serve a customer entirely on its own.
Source: The AI Republic: Building the Nexus Between Humans and Intelligent Automation (Tse, Esposito and Goh, 2019)
The best way to use AI currently is to reduce costly, mundane and labour-intensive tasks. It is far easier to see a direct return on investment and shows quicker results than employing AI with the purpose of expanding revenue. This low-hanging fruit also results in a quick win, helping your case for expanding AI capabilities to more areas of your business.
Weigh the risk
One misconception about AI involves the idea that implementing this technology into business means letting machines make decisions for us. Although some decisions will be handed over to machines, it doesn’t mean all of them will be.
If an AI model proves its accuracy and the decision being made is a relatively minor one, you can feel comfortable letting it continue autonomously. If a decision has major consequences, however, then you will likely want humans to be involved in the process alongside AI. Human-in-the-Loop (HITL) and Expert-in-the-Loop (EITL) give you this option, as they are branches of AI that leverage both humans and machines to create machine learning models.
Many law firms are now adopting AI to work on standardised contracts. Instead of a lawyer spending hours redrafting contracts with many customisations, machine learning can be used to streamline this process. The issue here is that if a machine makes even a small mistake, it could cost the firm hundreds of thousands, maybe even millions, in litigation. So even with a very high accuracy rate, humans still need to check the final product of the machines before passing it to clients.
Get your “last mile” right
No matter how much of a job is automated, there will always be at least a small portion of that job that needs to be overseen and managed by humans. When we talk about the “last mile”, this is what we are referring to.
Why is the “last mile” so important?
- Firstly, it’s imperative to create a foolproof system. Although the goal of AI is to eventually give an output of 100% accuracy, but it’s simply not there yet. For this reason, businesses still need workflows that are designed around human intervention.
- The second reason is that while there are many jobs that can be automated, it’s in our best interest to leave a number of them to humans. For example, the tourism and hospitality industry’s success is largely dependent on the customer experience linked to human interaction. Ever considered asking a robot for a room upgrade? It likely wouldn’t be the experience you’re looking for.
- Lastly, some tasks are simply impossible for machines to take over. For example, after getting your boarding pass from a machine at the airport, flyers still have to interact with airport personnel to check in their luggage. In more modern airports, this task is being done by flyers themselves, who have to print out their own luggage tags, place them on their bags, and drop off their luggage on conveyor belts.
Less is more
Historically, lots of emphasis has been placed on acquiring data. Most companies believed they needed massive amounts of data to start any AI project, which in many cases may be true if you want to train a model or machine. However, the idea that more data is better is not always the case. How much data one needs depends on what is trying to be achieved.
So how much data do you need?
There are three things to consider when trying to figure out what type of data you need and how much of it is necessary for what you want to accomplish.
- Minimum Algorithmic Performance (MAP)
This is similar to the idea of a minimum viable product (MVP), where businesses create a very basic version of a new product or solution in order to test and validate before moving forward.
When creating a MAP, focus is on demonstrating the performance of the model’s critical duties. The reason for this is to understand the absolute minimum level of accuracy needed to achieve the goal at hand (MAP threshold).
- Performance threshold
There are times when perfect accuracy is not needed to solve every problem, and there are times when problems can be far too complex for a machine to solve with perfect accuracy. In this case, your objectives will not be reached no matter how much data you have acquired.
This can be seen using the example of drivers’ licenses in the UK. Unlike in the US, UK drivers licenses are all standardised. The same piece of data is always in the same spot on the license, regardless of region or person. This allows a machine to easily pull data from them without needing a massive data requirement. It is quite easy in this case to achieve near perfect accuracy, achieving the performance threshold needed.
- Stability threshold
The amount of data you have is not as important as the data’s relevance to the problem you want to solve.
AI models decay over time due to the evolution of data sets, which are continually becoming outdated. If you’re an e-commerce business who sells sports caps, you will have accumulated a large amount of data over the years on those items. If you decide to expand or pivot your business to selling trainers, the data sets you have will no longer be relevant. In this case, quality trumps quantity.
It’s easy to become overwhelmed by the sheer amount of AI technologies and solutions out there. A pitfall far too many enterprises fall into is thinking too big, overgeneralising the impact AI could have on their companies. Remember, AI is best used in very well-defined, specific circumstances and accompanied by human talent.
Are you struggling to develop an AI strategy for your business? Check out our game-changing framework AI Factory here or Get in touch and one of our consultants will reach out to you within 1 working day.