← Blog
Company News
July 10, 2026 · 2 min read
Teaching Our AI Assistant to Actually Take Action
The first version of the ACBIZ AI Assistant could answer questions. That's a real capability, but it's also the easy half of the problem — an assistant that can only talk about your account, and can never actually do anything on it, still leaves you to go do the work yourself once the conversation ends.
The harder problem was letting it take real action safely. Opening a support ticket, saving something to a wishlist, requesting a savings withdrawal, cancelling an order — these aren't equally risky. Some are reversible and low-stakes. Others involve real money or a decision that's hard to undo.
The rule we built around: anything that touches money or cancels something you've already paid for always requires you to explicitly confirm with a real button tap before it happens. The assistant can prepare the action and explain exactly what it's about to do, but it never completes anything sensitive on its own judgment alone. Lower-stakes actions, like opening a support ticket or adding something to a wishlist, can complete directly, because getting one of those slightly wrong costs you nothing you can't immediately undo.
That distinction — not "can it act" but "which actions are safe to complete automatically, and which ones need a human to actually confirm" — is what took the most care to get right, and it's the reason the assistant can do considerably more than answer questions without asking you to blindly trust it with things that matter.
The harder problem was letting it take real action safely. Opening a support ticket, saving something to a wishlist, requesting a savings withdrawal, cancelling an order — these aren't equally risky. Some are reversible and low-stakes. Others involve real money or a decision that's hard to undo.
The rule we built around: anything that touches money or cancels something you've already paid for always requires you to explicitly confirm with a real button tap before it happens. The assistant can prepare the action and explain exactly what it's about to do, but it never completes anything sensitive on its own judgment alone. Lower-stakes actions, like opening a support ticket or adding something to a wishlist, can complete directly, because getting one of those slightly wrong costs you nothing you can't immediately undo.
That distinction — not "can it act" but "which actions are safe to complete automatically, and which ones need a human to actually confirm" — is what took the most care to get right, and it's the reason the assistant can do considerably more than answer questions without asking you to blindly trust it with things that matter.
#ai-assistant
#product