>from the second game
Yes, it's worth mentioning that I prefer the second game to the third.
>10 turns to go around, just to somehow reach the archers
The archers need to be "assessed" first; rushing towards 10 gargoyles isn't very wise if the player has a significantly stronger stack(s), against which the damage from the gargoyles will be negligible. Furthermore, the AI evaluates its chances of reaching the archers; if there are none, the AI should inflict maximum damage on the player, rather than trying to attack what it's given.
In my opinion, the main problem is the assessment of the situation. Before the battle begins, the player usually has information about the possible outcome and builds their tactics based on this. At best, the AI can only review the preliminary results of the auto-battle.

Let's take a situation where the player has a stack of units, let's say level 1, which the AI sees as units from 1 to 5, and expert slow: obviously, it's not beneficial for the AI to divide its units into several stacks, but it does so because it doesn't know that there's 1 unit, not 5. As a result, all units are slowed, and the superior forces are killed by a single stack of divided units that is only slightly stronger.
This raises several questions:
1. Should the AI have access to information about the presence of "key" spells for the attacking player? Resurrect, slow, haste, etc.
2. Should a "decoy" of level 1 units be left? In this situation, the AI could distribute its 60 conditional centaurs not into 20/20/20, but into 1/54/1/1/1/1/1.
3. The AI could assess the player's SP, remaining mana, and perform a kind of predictive analysis: "the player can cast slow one more time for 3 turns," go through the player's spellbook options, and defend until the slow wears off, if the damage from the ranged stacks is insignificant and doesn't reduce the number of centaurs to the point where the player's stack can kill them without significant losses, or the remaining centaurs can still reach and inflict some damage on the player's key stack.

>like chess
I see it exactly that way; the AI should model a table of its possible actions and the player's actions for n+ turns until damage is inflicted, and choose the combination that inflicts the greatest strategic damage, adjusting it each turn, depending on changes in the player's actions. The difficulty level of the game will be determined by the depth of that n.