Joydeep Dam, Director of Algorithm Development & AI at BridgeI2I Corporation, talks about the distinguishing features of an AI-enabled Digital Enterprise. He defines a digital enterprise as one with a well-characterized way of capturing data, plus a smart way of utilizing it, so that the result is a good way of basing decisions on it. The typical case is that a particular firm might have a lot of data, which then results in a lot of fancy dashboards or reports being created (Joydeep mentions a bank which had a database that created about 2000 [two thousand] different reports every single week.) But the question is, do the insights derived from that data actually flow into decision making?
He then goes on to describe the basics of the 'AI wrapper' one could put around the data generator/report generator stage:
- A Watchtower - a component that (figuratively) looks over data, identifies anomalies in it, and detects errant or deviant behavior from it or in it. A Watchtower could be conceived as a Digital Dashboard with an Intelligence Module overlaid, which one could perhaps call 'DDI'. Or one could have an 'SPI' (Signal Pool Intelligence) Watchtower, which might perform roughly the same function (but with signals of a more elemental form, for example, the outputs of an IoT array). Or one could have a BI Watchtower, i.e., a Watchtower with Business Intelligence, on which a structure is added that extracts insights or adds graphics and visuals, or even creates entire presentations based on the input,
- A Recommender - which, based on data input of various kinds (e.g., browsing, personal, credit or purchase history of an individual or firm) creates a set of customized recommendations for either the sales or the purchase side of any business transaction. This could be based on a ranking such as 'most often first' or 'most recent first' or, if likelihood can somehow be estimated, 'most likely first'.
- An Optimizer - a module which runs either a mathematical or heuristics-based optimization routine under a set of given constraints, and objective function(s), and provides a unique answer if available, but could also yield a set of sensitivity analyses.
- A Converser - (The word in this form is mainly used in French, as an intransitive verb, similar to the English 'to converse'). However Joydeep uses it in the sense of 'that which converses or that which is conversed with'. This module has the function of turning the analysis into a form that the user can 'consume' - that is, it converses with the user. This could take many forms, starting with just more intelligent graphics or video, to conversational 'chat-bots' (which may analyze text and/or speech to generate probability distributions of word occurrence, and then, based on semantic and/or grammatical rules, output text and/or speech).
These four elements can be taken as the basic elements which, if present in a given software product or service, can be talked about as representing an 'AI-enabled Digital System'.
A question asked by a member of the audience after the talk deals with the issue of Data Paucity. He asks, what do you do if the AI algorithm needs a lot of data but you only have a little of it? This is particularly the case for smaller companies and early-stage startups, but even otherwise, is a very good quesition. The answer is, the solution depends on the situation. One strategy is to wait until there is more data, or try to acquire it from elsewhere if it already exists, or to acquire (or occasionally, even generate) what might be called 'proxy data' that can 'stand in' (as an alternative which may have much the same statistical characteristics). Many AI algorithms can 'learn continuously', i.e., you can train them with the data you have, and then continue training when more data becomes available.
