The Investment AI win probably is most familiar to those conducting IT projects in Legal and Compliance: You identify an opportunity, invest time and money, and after some time you get the reward. Examples of Investment AI wins are:
- Setting up an information retrieval system for internal documents and then benefitting from better availability of information
- Automating compliance reports, using real-time data, to reduce manual effort
- Introducing intelligent document filing and sorting, cutting away administrative efforts
Now AI has some peculiarities that need to be taken into account when setting up implementation or development projects. Mainly these are:
It’s the data, stupid!
Contrasting common belief, AI is not based on magic. It is based on data. If you are running a larger AI project in your company, usually at the very core you are adapting an existing AI to your organisation’s data.
So it is okay to be impressed by shiny AI demoes, but soon after, you should ask yourself questions like: Which data could we use to adapt the AI to our use case? Does it require additional annotation? Is the data organised and available from one source?
Often your data also contains exceptions that you do not want an AI to learn from. Or it contains biases you do not want to see reproduced. Or your data structure is non-trivial, like it contains document drafts that later were dismissed, and that you want to filter out. In short: Your data usually needs to be cleaned and transformed before you can make good use of it. After this is done you need to ensure to maintain high standards of data quality.
As a consequence, and speaking from experience, you may overlook the need for data engineering and data engineers to ensure data quality. These people lay the base, and usually are not the ones you talk with when planning your project. So make sure you do not miss this need!
Is this a problem worth solving?
Assume you see an AI solution that is state-of-the-art, surprisingly cheap, works reliably, and even solves tasks you never thought a computer could solve – should you implement this solution as fast as you can? Well, not necessarily. The main question is whether it improves how work is done in your company, and this is surprisingly hard to predict.
The best you can do is to test it in practice as early as you can. So instead of trying to perfect your project planning and identify edge cases and further opportunities in advance, rather go small: What is the main problem you think this tool can solve? How can you test this assumption in practice with minimal effort? As a result, you may start testing in just some practice areas. Or just one feature of the AI product. In some cases, you might even ask students to manually do the task you intend to automate by the AI, to test with minimal commitment how such an automation would affect the overall process in practice.
AI is evolving
One factor that adds complexity to larger AI projects: AI is developing super fast. So fast that a better and cheaper solution might be available when your project is finished. At the same time, we do not really know how exactly AI will evolve: Which problems will soon be solvable by AI, and which ones will remain to be hard?
To make smart decisions on which Investment AI wins to aim for, you need to take this into account. There is no clear rule, but for orientation: Structuring data and improving its quality almost always is worth the effort, as it is required for any future AI project. Optimising for your use case based on existing technology often also is highly-rewarding, but the risk of being overrun by AI developments exists.
And in doubt: Go with it. Even if your solution is outdated soon after implementation, you will have learned a lot on AI and on your use case. This knowledge will help you achieve superior results in any subsequent project.
To wrap it up, Investment AI wins need careful handling of data. It helps to experiment early on to evaluate usefulness in practice. Keeping an eye on latest AI developments can be helpful, but in doubt, start experimenting.
