If you follow engine tournaments, you'll find that although computers are very strong, there are still times when they do the most stupid of things. Here's an example from the most recent TCEC superfinal between Stockfish and AllieStein.
[FEN "2b1r3/6r1/p2p1k1p/Pp1PpP1P/1Pp2pP1/2P2P2/1K2R1B1/3R4 w - - 23 149"]
Try analyzing the position yourself before looking at what the engines say. What would you give this position?
If you said 0.00, I completely agree with you. After all, White can't make progress - all the pawns are firmly wedged. The only thing White can do is play g5, but that's just going to lose several pawns. Meanwhile Black is just as stuck. White can easily defend the pawn on d5. The only pawn break Black has is ...e4, but that will actually lose the game by not just throwing a pawn, but also opening lines for White to make inroads.
However, if you look at the engine evals, Stockfish gave +0.90, and AllieStein gave -0.50. If you look at the charts of how the evals vary, you'll see they're straight horizontal lines (at least Stockfish's is - AllieStein, being a neural network engine, has a less consistent but still mostly horizontal eval). This is what computer chess viewers derogatorily call "horizone". The engines aren't smart enough to realize they're not making progress. They will happily shuffle until the 50-move rule helps them realize that it's a draw. This kind of fortress situation is quite common, but can be avoided. If you're a human piloting Stockfish, you can see this fortress come up several moves before, realize that Stockfish is heading into a horizone, and choose a different move that keeps some winning chances in the position.
But that's not all humans can contribute! As it turns out, AllieStein spotted an ingenious way to make progress. Check out what happened 15 moves later. By this time, Stockfish had realized it wasn't making progress, and its eval had dropped to +0.38. Meanwhile AllieStein's eval had increased to -0.79. And then there happened ...
[FEN "2b3r1/8/p2p3p/Pp1PpPrP/1Pp1RpP1/2P2Pk1/2K5/3R1B2 w - - 53 164"]
1. Ree1 Kxf3
What kind of fanatic plays 164...Kxf3? Is AllieStein insane? Who the hell ventures into enemy territory with a lone king to grab pawns anyway? Black can't even back off through g5 because the square is currently blocked.
If you're a human piloting AllieStein, you'd immediately realize this is a pivotal moment. Either AllieStein is seeing something profound, or it's making a huge blunder. Before actually making the move, you'd try out variations to see what might happen. You might not be strong enough to tell yourself, but AllieStein is sufficiently powerful to tell that within a few moves Black's so-called advantage has cratered. After 164...Kxf3 165. Be2+, Black is dangerously close to being checkmated. 165...Ke3 166. Rf1 threatens Rf3+, Re1 and Bd1 checkmate. Black is forced to take desperate measures like ...Bxf5, after which his position collapses. AllieStein on its own didn't see this, and lost this game. But human + AllieStein is likely to have avoided this.
This is just an example of the more obvious ways a human can contribute. Correspondence players will be able to say more, since they get familiar with their engine's strengths & weaknesses, and feed their engine ideas on how to get to their strongest positions while avoiding the opponent's. From what I've gathered talking to correspondence players, the odds are that they will draw against someone playing only engine moves (since chess is a draw after all), but they won't lose either, and they will score wins. That's where humans can contribute.