AI projects that make sense — businesswise!
Mailiis Ploomann, Head of Telecom Services in Elisa Estonia
AI has been a hot topic for quite some time, Estonian Government has set a goal to have at least 50 different AI solutions in public service by the end of 2020 (starting from 5 as of today), enterprises all over the world are realizing that next level effectiveness can only be achieved through innovative methods and AI is often one of the options on the table — but why do these projects succeed so rarely? This has been the burning question in our heads in Elisa since 2017 when we first started our own AI journey — and I believe we have finally close to an answer, after a long and winding (but really fun) learning curve.
There actually IS a way to build sustainable AI projects that are not just innovative for PR statements but make your business stronger, more resilient and ready for the future, but in order to achieve all that — you need to go through a certain set of actions and be really honest when answering those questions!
What is the business problem you are trying to solve?
It all starts with that. Forget ML (machine learning), AI, neural networks, data engineering and all the other fancy terms that don’t make any sense to you at the beginning. It doesn’t matter at this point. Not the slightest.
What matters is your business strategy and what are the key obstacles that keep you from achieving excellent results? No AI can provide a solution to a problem
my business is not profitable, it is you who have to figure it out and then come up with an actual problem that needs to be solved.
If we think about an AI model as means to automate problem solving, then what we really need is a process. Does your problem look like a process? In order for an AI model to succeed, it should.
It should have a clear beginning and ending with a logical path between those two. Someone should be able to describe how the decisions are made (what are the
rules) during that path and on what grounds (based on which data) are decisions made? Who are the involved parties and what is their input and output?
In a broad scale we could say that most of the tasks we as human beings do, could be divided into two categories. Tasks that require high cognitive skills with the ability to transfer knowledge between domains (and though it sounds really sophisticated, it is actually something a 2-year-old can master quite easily J ) and tasks that we perform on “autopilot” because they are repetitive, routine and quite honestly just dull in most cases.
As human beings are emotional creatures, we tend to get bored and tired when performing the latter tasks, which also makes us error –prone and that is the ideal place for an AI model to step in.
An AI model does not get bored, need a vacation or even mandatory lunch hours, It can keep going 24/7 with the same performance level and often even outperform humans in that sense — so if your business problem contains of a process that has repetitive, routine tasks with clear rules — you have most probably found a process that should be automated.
On the other hand, this is not a decision you should make lightly. Just because something can be automated, does not mean it should be automated. A good example can be found in the line of chatbots — probably most broadly used AI models today. Just because you have (human) customer service agents chatting with your customers today does not necessarily mean you should try to automate it at any cost. It all comes down to the nature of your chats. If most of the conversations tend to look the same, with customers having similar questions and your agents giving copy-paste answers, then a chatbot is definitely what you need. But if your products are mostly tailor made and rather complicated, it probably means customers could have a whole area of different questions and answering them requires an agent to check and look up several different data sources and derive an answer from those, which might be a simple task for a human being — but rocket science for a bot.
And that in essence leads us to the next relevant question:
Will you ever get a return of investment — does it makes sense to automate this process using AI?
In order to answer that question, we should hop back to the first question (what is the business problem we are trying to solve) and define some key metrics that we want to start improving.
Innovative and new technologies are expensive, there is no doubt about that but using them before competitors might just give you the competitive edge you have been looking for — so it is often a risk you should consider taking. Nevertheless, it only makes sense if you know how to measure success, which is why you should define your key metrics at the very beginning.
For instance — a chatbot could serve several different business purposes. You could aim for a faster service time or 24/7 service hours. Or maybe you just want to relocate your human agents from mundane and repetitive tasks to more productive (and emotionally intelligent) ones.
The first two usually mean you want to improve customer service SLA and satisfaction rates, the most popular metric there is often NPS (net promoter score) which means you can now calculate how much a potential rise in customer satisfaction / NPS score is worth to you and then set the goal for your AI model based on that.
When your problem is that human labor is just too expensive for some simple tasks, then you can calculate ROI based on the cost reduction you get when automating the jobs /relocating your people to do more productive tasks.
Both approaches are totally fine — but you should pick one to be your key metric (the true north) and measure the success of the project based on that.
Now that you have defined a process to be automated and concluded that it indeed makes sense to automate it, you should look for a partner who can help you understand whether this process can be automated by an AI model.
Data science is quite a specific skill and one of the stupidest thing you could do, is go exploring it on your own with no-one to guide you along the way. Reading blogs and listening to podcasts is definitely eye-opening but a smart person wouldn’t experiment with heart surgery based on something he/ she read somewhere, even if it sounds fascinating. At least it would be a lot cheaper to learn from other people-s mistakes in the case of AI models, so find yourself a vendor or service provider that you feel good with.
Now when choosing that partner, there are some key criteria to keep in mind.
You should look for someone who actually
knows what data science is (not just someone who is really eager to find out with you) and at the same time is sincerely interested in your business goals.
Good data science is a mixture of mathematics, statistics and IT with a hint of biochemistry and understanding of business processes. Should anyone claim otherwise — it is better to end the relationship at that very moment. Since data science is a global industry you can choose a service provider that suits your needs the best from basically anywhere in the wolrd.
In Estonia, there are already several world-class players like MindTitan, Alphablues or Feelingstream and since it is quite a young field, you should keep your eyes and heart open for newcomers as well.
Once you have chosen a partner that feels right to you, it is time to put together a team that will make it all happen.
From that point on, you should bear in mind that it will not be yet another IT project, where your job (as a business side representative) ends after putting together the request for the IT guys and starts again after the development is finished and live in production. On the contrary.
A successful AI project always starts with a thoroughly described business problem and a problem owner from the business side who is able and willing to support the project all the way through.
While data engineering and the actual building of the model should be left to the data science partner, interpreting the results and deciding the paths to take / reject the ones that don’t make any sense is 100% your job. It is after all your business and only you can know what are the key components you need to improve your chosen metrics.
And since data scientists are, well scientists (or at least engineers in their hearts) and business people are usually not — it is wise to have one more role in your team. A team leader or a project manager who can act as a translator for both parties. Ideally a person who speaks “IT and engineer” fluently but who’s mindset belongs to the business side. It would be his/her job to keep this project productive and make sure all parties understand what is expected of them.
It would also be that person who should start building the inner processes to keep the model productive after it is built and in line with the ever changing business processes.
There is no doubt — the era of business requests and IT developments as two different and separated tasks is ending. Future is full of interconnected and interdisciplinary projects and it is our job to evolve with that. Only then we will be able to build AI projects that really make sense, business wise!
Story was originally written for and published in Estonian leading leadership magazine Director (march 2019)