The leash
How an AI teammate is kept on a leash.
Handing an inbox to AI is a real leap of faith. Here is exactly what stops Podley from doing anything you did not agree to - not a promise, but the way it is built. Every claim on this page is enforced in code, not policy.
The model proposes. Code disposes.
The AI can only ever suggest an action. A separate, plain rulebook decides whether it actually runs.
When a customer email comes in, the AI reads it and writes down what it wants to do: draft this reply, refund this order, change this address. That is a proposal, nothing more. It does not have its hands on the buttons.
Between the AI and anything real sits a small piece of code that is not the AI and cannot be argued with. It reads the proposal and decides, on fixed rules, what is allowed to happen. The AI cannot talk its way past it, because the code never asks the AI for permission. It checks the request against the rules and either lets it through, holds it for you, or refuses it.
This matters because a clever email, a prompt-injection attempt, or a model that simply has a bad moment can all produce a bad proposal. None of them can produce a bad action, because the rulebook, not the AI, has the final say.
A fixed list of things it is even allowed to attempt.
Only 33 specific support actions can run. Anything outside that list is refused, by name.
The rulebook starts from a closed list: there are exactly 33 support actions Podley can carry out, and they are all ordinary customer-service moves - issue a refund, cancel an order, swap a size, apply a discount, send a reply, tag a case.
If the AI ever proposes anything not on that list - because it invented an action, or reached for an internal tool it was never meant to use - the code rejects it outright as out-of-manifest. It does not improvise, and it does not fall back to "close enough."
That is why there is no action for editing your products, your themes, your storefront, your pricing, or your store settings: those actions are not on the list, so there is no path for the AI to reach them, even if it tried.
Money doesn't move without your say-so.
Podley starts in Training mode: it drafts everything and sends nothing on its own. Even after you grant autonomy, hard dollar caps stand.
Out of the box, Podley runs in Training mode. It reads, it researches the order, and it writes a full draft reply for you - but it sends nothing and moves no money on its own. You review and approve. If Podley ever cannot read your settings, it fails back to Training on purpose, so the safe state is the default state.
When you decide you are ready and switch on autopilot, the caps do not go away. A refund above your auto-approve limit is held for you, not sent. A refund on an order whose total Podley could not confirm is held. Several small refunds that would add up past your limit are held as a group. A second refund on an order that was already partly refunded is held. And a rolling 24-hour ceiling stops a drip of small refunds from adding up to a runaway day.
Free product counts as money leaving the business, so a replacement order is measured against the same limits as cash - and a padded, multi-item reship is valued by everything in it, not by one unit.
One more line: anything that looks like a legal threat or an escalation never pays out automatically. It goes to a human, every time.
It can't promise something that didn't happen.
A deterministic check reads every draft and flags a reply that claims an action is done when the system only held it for your approval.
A support reply can lie without anyone intending it to. "I have refunded you" reads fine - until you notice the refund was actually held for your approval and never ran. That gap is where trust breaks.
So a separate check reads the finished draft against what actually happened this turn. If the draft says a refund, an address change, a cancellation, or an exchange is already done, but the action was held for your approval or handed to a human, the check catches it. It only fires when the system can prove the claim is false, so it never second-guesses a reply about something that genuinely was completed.
The same check catches promises of follow-ups nothing will deliver - "we'll email you when it ships," "I'll keep an eye on it" - unless there is a real timer or a real notification standing behind the words.
Hitting your plan limit means drafts, not a surprise bill.
When you reach your monthly resolution cap, auto-send switches off and Podley keeps drafting. It never bills you past the plan you chose.
Your plan includes a set number of resolutions each month. You can also set your own ceiling below that. When you reach whichever limit comes first, Podley does not keep spending on your behalf and it does not silently ring up overages.
Instead, auto-send switches off and Podley goes back to drafting for you to review and send by hand. You get an email at 50, 80, 95, and 100 percent so nothing is a surprise. The work does not stop; it just waits for your click.
The hard line
What Podley can never do.
Touch your products, themes, storefront, pricing, or store settings.
Run any action outside the fixed list of 33 support actions.
Move a refund or free reship past the dollar limits you set.
Refund an order whose total it could not confirm - that is held, not guessed.
Pay out on a legal threat or escalation - those always go to a human.
Bill you past your plan - reaching the cap drops it to draft-only, never an overage.
None of this depends on the AI behaving. The allowlist, the caps, the promise check, and the draft-only fallback all sit outside the model and run on fixed rules - so the worst a bad proposal can do is get held for you.