Ask Me First, or Don't Ask

Ask an AI agent to set up an ad campaign on your best sellers and "get it running", and watch what a good one does: everything except the part that costs money. Ours read a month of real orders, picked five listings that actually sell, built the campaign — and delivered it paused, with a note that nothing is spending until an approval waiting in the dashboard is answered.

People are rarely nervous about an AI writing a product description. They're nervous about what happens when it can spend money and send email in their name. That's the right thing to be nervous about, and it deserves a straight answer rather than reassurance.

Here is ours. SellerClaw — a team of AI agents connected to the stores you already sell on, plus your suppliers and ad accounts, run from a chat — puts everything that spends or speaks for you behind one setting, with the consequences written out.

Two modes, and under each one the exact list of what changes.

Two modes, no small print

Ask me first is the default. Nothing spends or sends until you say yes: ad spend and campaigns, email and customer replies all wait for your answer.

Do not ask is the other end, described honestly rather than sold: campaigns go live and budgets rise without confirmation, email and replies are sent as drafted.

The setting is deliberately boring and global: one switch for the whole team, taking effect on the next action. No per-feature exceptions to memorize, no mode where a specialist has powers the supervisor doesn't.

The ask names what will actually run

A gated request isn't a vague "proceed?". Here is the campaign from the opening, as it reached the owner — built, priced, and explicitly not spending:

Second line of the report: nothing is spending yet. The campaign card at the bottom of the message says Paused.

The request itself names the platform, the campaign and the exact change, and carries a deadline instead of hanging open indefinitely. Email works the same way, and there the review is even more concrete — the exact words about to leave in your name, from the address they'll leave from:

The Message block is the review: these words, from this address, to this recipient — or nothing goes out.

A sentence counts as much as the button

You can press Approve or Reject — or just answer in the chat. "Hold off — don't launch it this week" did what the Reject button would have done: the agent withdrew the request and stood the campaign down, and the reason travelled with it.

One sentence instead of a button — and the confirmation that there's nothing left to approve.

A yes works the same way, with one guard: a reply in chat is checked against what was actually staged before it counts. "Yes, go ahead" approves the thing that was proposed, not whatever the conversation has drifted to since.

And the quiet mode leaves the same trail. In "Do not ask" the request is still created — your setting answers it instead of you, and the history says so in plain words:

The second entry closed without a question — and the record states exactly that: you were not asked before it ran.

Whichever mode you pick, the things only you can physically do — paying an invoice from your card, connecting an account — always come to you. "Do not ask" removes the confirmation step for work the agent already knows how to do; it does not let the agent authorize things it has no way to authorize.

What the gate doesn't do is rescue a bad instruction. Approve a campaign with the wrong budget and it launches with the wrong budget — the gate protects against surprise, not against your own typo.

What changes for you

You stop having to accept "useful" or "safe" as a personality trait of the tool, and get to set it — one you can change on a Tuesday and change back on a Friday.

Start in Ask me first — it's free to try, and the credits are enough to watch it work for a while. When you notice you've approved twenty routine things in a row without reading them, that's the signal to loosen it. Not before.