If-Then rules cannot exhaustively cover all possibilities that can be encountered in real world situations. One of the hallmarks of A.I is the ability to generalize across problem domains, not just individual problems like Vision or Natural language or Automated driving. Specifically, a true A.I system would not need a programmer to specify each and every problem that it might have to solve or even the specific conditions within each problem that it might encounter.

Consider this, the number and type of problems that potentially exist is virtually infinite. So, to be able to exhaustively catalog all the potential problems is virtually impossible. Furthermore, to be able to then write If-Then rules for all the problems and their individual conditions is a step beyond the impossible. Automated driving is a good example as a problem where there can be an infinite number of If-Then rules to write down if one were to go that route but you can always create a new situation that the programmer didn't think of like what should the system do if a meteor crashes in the middle of the road.

What A.I promises is automation of reasoning ability in domains and problem sets that are broad. In some very narrow fields where the possibilities are countable and finite and it can be taken for granted that the system exists in a pre-defined set of states, one could use If-Then rules to surpass human ability. For example, in Chess simple rule based systems could surpass human abilities a long time ago. But even in this well defined game, a rule based expert system worked well until a new system (AlphaZero) came along which was not designed with If-Then rules so was not limited by the imagination of the programmers either tactically or strategically and blew all the previous ES systems out of the water. But real world problems are seldom well defined and the set of potential states of the system can be large and varying in size.

Note that there are many well defined real world problems that ES can tackle and A.I wouldn't provide any additional advantage but the number of problems that A.I potentially could handle where ES would be sub-optimal is much larger.

If-Then rules cannot exhaustively cover all possibilities that can be encountered in real world situations. One of the hallmarks of A.I is the ability to generalize across problem domains, not just individual problems like Vision or Natural language or Automated driving. Specifically, a true A.I system would not need a programmer to specify each and every problem that it might have to solve or even the specific conditions within each problem that it might encounter.

Consider this, the number and type of problems that potentially exist is virtually infinite. So, to be able to exhaustively catalog all the potential problems is virtually impossible. Furthermore, to be able to then write If-Then rules for all the problems and their individual conditions is a step beyond the impossible. Automated driving is a good example as a problem where there can be an infinite number of If-Then rules to write down if one were to go that route but you can always create a new situation that the programmer didn't think of like what should the system do if a meteor crashes in the middle of the road.

What A.I promises is automation of reasoning ability in domains and problem sets that are broad. In some very narrow fields where the possibilities are countable and finite and it can be taken for granted that the system exists in a pre-defined set of states, one could use If-Then rules to surpass human ability. For example, in Chess simple rule based systems could surpass human abilities a long time ago. But even in this well defined game, a rule based expert system worked well until a new system (AlphaZero) came along which was not designed with If-Then rules so was not limited by the imagination of the programmers either tactically or strategically and blew all the previous ES systems out of the water. But real world problems are seldom well defined and the set of potential states of the system can be large and varying in size.

Note that there are many well defined real world problems that ES can tackle and A.I wouldn't provide any additional advantage but the number of problems that A.I potentially could handle where ES would be sub-optimal is much larger. Also check out LCS that can synthesize rules through learning and exploration.