Skip to main content


Millises keskkonnas sa veebi kaudu TV-d saad vaadata, sõltub sellest, millise TV-teenuse kasutaja sa oled.
Uus Elisa Elamus Liitumised alates 10.08.2021 Mine vaatama
Vanemad TV-teenused Liitumised enne 10.08.2021 Mine vaatama

You have your Proof of Concept — so what happens now?


Mailiis Ploomann, Elisa telekomiteenuste valdkonna juht
Mailiis Ploomann, Elisa telekomiteenuste valdkonna juht

You’re sitting in your office after a really successful meeting with your AI team. You have the results in your hand. Most of the assumptions that you made came true and the success criteria was met. You have the technology. It works. Now what?
You know the feeling?

Well I do. It’s not all bliss and joy to be honest J. And by now I also know that when it comes to AI projects, the real tricky part is actually not the magic your data scientists come up with — it’s what happens when you begin to scale their work and really implement all that awesomeness into your everyday life.

If you represent the business side of your company (and in order for all this to work, you should) then you are probably used to the idea of being the business problem owner for your “regular” IT projects.
And in that case it usually goes something like this: you explain your idea to your IT department, depending on the size of your company you either have a PO or IT analyst on the other side of that dialogue. They ask you a bunch of questions (plus of course you go through all the bureaucracy etc but since that is in no way relevant to my point, I will skip that part) and then you WAIT.

During that waiting period there is a lot going on in the IT dept but you are not needed and it is usually appreciated if you let them do their job in peace and not meddle too much.
Then at one moment you are asked to participate a demo and depending on what you see there, it is either the start of a happy ending or agony for both sides to redo some of the work or continue waiting. Anyhow, one of those demos will be a successful one and after a short while — you will have your solution. Coded, tested, documented, polished and working.
You thank the IT team, they move on to their next project and you move on with your business. 
It is expected that unless you need to change something — the solution you ordered will work the same way on day 1 and day 3790.

This process may vary in speed and steps but in essence it is the same in most companies. Business people make the order and use the result, while IT department does most of the heavy lifting in actually bringing the solution into life. So naturally — when you find yourself in that position we started this article with– you know there is (an AI) technology that works, you might feel tempted to follow the beaten path. Make an order and let the IT guys do their thing. Well. To make it short. That doesn’t work.

We have already discussed that the process of defining your AI projects should have a different approach than your other projects and that there is a specific set of roles you need in your team in order to make it work — but there is one more aspect you have to consider. And that is your own mindset in what’s next to come.

It’s important to realize that your AI solution is based on data. And that data is gathered in the past. As long as we are dealing with narrow AI (and that will be the case for quite a long time) there is no way your AI product can learn new information unless your data science teams provides it with the relevant data set. So when your business rules change — you need to “tell” your AI product that as well.

Imagine a chatbot that is completely competent in answering the question “how long are you open on Wednesday?”. It has learned in the past that 7 PM is usually the answer human agents have given so that’s its answer as well.
Now what happens when you make a decision to be open for another hour? How will it know?
It doesn’t make sense (for business people at least) to start gathering the new training set and let the bot give out false answers until you have enough of new data to retrain the algorithm.

For human customer service agents, you probably have a procedure in place for such changes. You write it down somewhere and you let them know via their superiors or send them an e-mail about it. Well, you can’t inform your chatbot that way :)

Or let’s imagine a fault detection algorithm that is set up in a ceramic tile factory. Using computer vision and pattern detection it is trained to recognize that blues tiles must always be blue. If another color is visible — something is wrong.
Once your creative designers come up with a new style — a blue tile with a single red stripe — you need a process to “let your algorithm know” from the first new tile that this is OK now. The business side simply is not OK waiting for the first 10 000 samples to be gathered.

Now the fun part is — this is a whole new territory for everyone involved. The business side is of course not used to the idea of thinking this way. The data science team has now idea how the business people make those decisions and how they come into effect. The “regular IT” is not that much involved anymore and the team that brought your PoC to life (with flying colors) might be devastated because everything is falling into pieces, and sooner or later someone (that is in charge of the budget) will come to a conclusion that all that is just not worth it.
This AI thing is just too young a technology and we should wait for it to mature before we try again.

Well now, that’s just a stupid thought and if you want your company to be successful in the future as well — it would be wise to kill such an attitude as early on as possible.
AI projects are nothing like your regular IT experience has taught you. Most of the rules and roles just don’t work the same way here. And it’s up to you — business owner- to make it work.

I am sure that there are companies that have their business “set” for a reasonable amount of time, maybe telecommunication is an extreme example in that sense but that means there is all the more reason for you to profit from our learning curve. 
We currently have 3 launched (in the scaled/ business as usual state) AI projects, 2 in the PoC phases and 3 in the preparation pipeline. Most of them in different AI branches (NLP, biometric facial recognition, deep learning predictive models, recommendation systems etc); the oldest of them being our chatbot Annika.
To give you some insight into the world that is beyond data science — I will use her (yes, it’s definitely a her but that’s another story) as an example for a companywide integration and really making your AI product work for you.

Annika is a customer service chatbot that has one aim — to resolve customer contacts. Not just provide customers with answers but to guide them towards a solution. That means, every single label it recognizes is divided into several different intents and (sometimes quite complicated) answer flows are built into its knowledge base.

Those answer flows look like mindmaps and in order to keep Annika up to date with everything that is going on in the real business world — every single node in that answer flow has a specific person responsible for it.
It actually goes a little further than that.
We have built a system where every single intent is actually owned by a specific product manager and it’s his / her job to check, verify and sign (in our case in every 3 months) that this answer flow is still up to date. We have also implemented a procedure for those “node owners” or intent owners to let the ML team know if they want to change anything in advance.
And since it’s not our first rodeo in product management — we have also implemented several RPA checks to make sure we catch every deviation as soon as possible.
To back this all up, we have revised and rethought most of our incentive systems as well.

While all this might seem natural and “nothing new” to several of you — I encourage you to be honest with yourself. Is it really working like that IRL in your company? Or are you like the most of us in the situation when you always have those customer service agents that have printed out something 3 years ago and still follow what’s written there today? Or just forget to check the CRM and give your customer the answer that is “usually true”? Well, I am quite confident that there is more of the latter out there and it is perfectly normal! Because fixing it means more rules and less creativity — and that’s just not something people appreciate.

Now (ro)bots are a different kind of workforce. They love routine and rules and repetitive tasks. And they are happy to follow them day by day. 
Still, in order to make them really WORK for you, make you more productive etc — it is YOUR job to put all the necessary processes in place. 
Not the data scientist’s, not the project manager’s but yours.

AI projects will never be like the IT projects you have been used to. They will not look the same nor feel the same. And it is really important to acknowledge that from the beginning.
Your project will probably not fail because you have poor data scientists (ok that might also be the case, but then it will be an easy one to fix as well). It is much more likely that they will fail when you as a business owner is not prepared to be the one who makes it work.

I personally know nothing about programming. Like really zero. From my perspective it goes into the category of doctors’ handwriting — undoubtedly a lot of intelligence there, but how they understand it is a mystery. This is probably due to some awful experiences back in 2000 sth when in the university we were forced to create some nonsense with VBA (I remember I had to make a green frog shaped button that would calculate the capacity of wood for some container). This was so absurd and far from reality that I quickly decided that I actually have other talents in this life and never revisited programming in that sense. 
But that green frog also gave me a valuable trait for the future — I know how to surround myself with the most talented people from different fields of life. And furthermore — I know how create an atmosphere where they can all exploit their individual talents.

And as it was my gut feeling back then that I will never be a programmer — I know for sure that IT and programming in general are entering a new era. An era that intervenes several disciplines together and it is up to the business side to be in the center of all that!
AI and data science will make your business flourish, but only when you are willing to lead the way. All the way!

The story was initially published in Estonian in Ärileht 11.08.2019

Seotud märksõnad