AI Policy

Eleven things, and there is no twelfth: your lead form definitions without submitted answers, your ad accounts and their campaigns, a single campaign in detail, your account's past as far back as each network keeps it, your site's analytics and product feed and search terms, Google Analytics on its own, the objects the network itself holds (audiences, pixels, creatives), the media you uploaded here, each network's current publishing rules, what changed on the four networks this week, and your connected ad accounts as the networks hold them right now, read live when the stored copy cannot answer or you ask how things stand now.

It also searches the web, pinned to your own country rather than the whole of it.

Every one of those is your own account answering. Nothing is drawn from another customer's data, and there is no benchmark table: a figure is compared to your own history or it is not compared at all.

It cannot change anything in your ad accounts: every change it proposes is a card you press. The database connection it reads King's stored copy through has had writing removed from it, and throws an error the moment anything attempts to write. Not "is not allowed to": cannot.

A live read is kept to reading by King's own code rather than by a missing ability, because the access you granted a network could also write. King builds each live request itself, for your connected account, and sends only requests that read. Two other kinds of request can come with a live read, and neither changes anything in any account: a request for a new key when the one King holds has expired or is about to, described below; and, on Google, when your account is reached through a manager account, a search for that manager — King lists the accounts your Google sign-in can open directly, reads each one's name, currency, time zone and whether it is a manager account, and reads each manager account's list of client accounts.

There is no delete verb anywhere in the assistant. It cannot remove a campaign, an ad, an audience or an asset, and there is no instruction that unlocks one.

A duplicate of a campaign is always created paused, so nothing starts spending because you were exploring an idea.

Everything that changes anything, with one exception you switch on yourself, described in the next section. The assistant proposes by drawing a card showing exactly the request it would send, and the card is built by the same code that would send it — so what you approve is the request, not a description of it.

On a campaign that already exists it can propose pausing it, resuming it and moving its daily budget; on Google, changing its countries, its ad text or its target ROAS or target CPA; and on Snapchat, pausing, resuming or renaming a campaign, an ad squad or an ad, setting when it starts and ends, changing an ad squad's daily budget and bid, and changing a campaign's countries, its age range, its languages or its ad text — each one a card you press. It can also propose a new campaign, which it builds in order once you press: the media, then the creative, then the campaign, the ad set and the ad. If any step is refused, the steps already made are unwound rather than left behind. On Snapchat it can also propose uploading one file from your media library to your ad account's Snapchat media library, without making an ad: the file is sent to Snapchat only when you press that card, and nothing is published.

Outside that one exception, every button that spends money is pressed by you.

Only inside an automation you switched to act mode yourself. Every automation is set to propose until you change it, and in propose mode it tells you what it would have done and sends nothing to any network.

Three automations can be switched to act: Raise what is working, Taper what is not working, and Pause what is not working. In act mode they can raise a campaign's daily budget up to a ceiling, lower it, or pause the campaign, and nothing else. On Google, automations only propose, in act mode too: a change to a Google campaign is shown to you as a proposal, and it is sent to Google only if you make it yourself, from the campaigns screen. A budget moves one step at a time, within a weekly limit. Each of them decides by fixed arithmetic on your own figures, not by asking the AI model, and every change it makes is recorded with what was there before.

Everything else waits for your approval. Nothing here queues a multi-step plan and runs it while you are asleep, and the assistant never acts on its own: it answers when you ask it and proposes when you ask it, and that is the whole of its initiative.

The rest of the scheduled work reads: it reads your connected accounts on a fixed cadence so the figures are current, and every other automation reads and reports without changing anything.

Nothing into your ad accounts. What it writes goes into King's own records: one sentence into its own memory, when you tell it something about your business worth keeping; and a new key for a network, when a live read finds that the one King holds has expired or is about to, which King stores so the connection keeps working. If the network has already replaced the old key and the new one cannot be stored, the connection cannot work any more, so King marks it as needing to be reconnected and tells you.

King also keeps its own record of the work: the conversation, what each answer cost, and each card the assistant asked to show you and was refused, with the reason. And it counts each question and each live read, against the limits on how often they may run, and the requests made to each network each day. A count is a number and carries nothing from your ad account.

Your monthly report is arithmetic: nothing in it is composed, so nothing in it can be invented. The single paragraph the assistant writes on top of that report is checked afterwards, number by number, against the report's own figures — one figure that is not on the report and the paragraph is dropped and the report goes out plainer.

The report also names, at its foot, any network it could not read. A total that quietly excludes a broken connection is a wrong number with a clean bill of health attached.

The assistant runs on Anthropic's models, under Anthropic's commercial terms. Your conversations and the account data in them are sent there to produce an answer.

We do not sell your data, and we do not use one customer's advertising data to answer another customer's question. The privacy policy is the fuller statement and this does not replace it.