Deploying AI agents in enterprise: ship fast and iterate vs build the 'right' architecture first?

There's a debate I keep having with enterprise AI teams.

Side A says: 'Ship something working in 4 weeks, prove value, then refactor. Enterprises want results, not perfect code.'

Side B says: 'One bad architecture decision in enterprise AI costs you 6 months of rework. Get it right before you scale. The cost of fixing later is 10x.'

Both have merit. Both have horror stories.

→ Ship fast and iterate — perfect is the enemy of deployed

→ Architecture first — enterprise AI has too much tech debt risk

→ It depends on the use case (explain in comments)

→ The question is wrong — you can do both with the right team

I've seen both approaches fail and both succeed. What I've noticed: the right choice usually depends more on the organization's risk tolerance than the technology stack.

What's your take? And if you've lived through the consequences of either choice — I'd love to hear the war story.