Autopilot: how an AI ad manager should be constrained and held to account
An AI manager that can move money needs limits it cannot talk its way around. This guide explains the constraints on Autopilot: what it optimises, what it may never touch, how much autonomy it gets per action, how every change is recorded, and how you rehearse and stop it.
- Rewritten against the product as built: the policy levels, which kinds of change Nomy may apply on its own, and which always need a person.
- Removed features that are not built, such as a single kill switch and weekly experiments.
- Set the guardrails before you switch Nomy on: a daily spend cap, a bid range, a budget range and protected campaigns.
- Start on suggest only with the ad account in shadow mode; shadow mode is recommended for the first 48 hours.
- Move to act on high confidence when the proposals you approve match what you would have done; the default threshold is 80 percent.
- Read a daily budget as an average: Ads Manager allows up to 2x the daily budget in one day and 7x in seven days, so set the daily spend cap with that in mind.
- Judge Nomy on outcomes: every applied proposal is measured on the seven days after against the seven days before.
- Scheduled tasks: a daily brief, an optimisation run hourly, daily or weekly, and a weekly scout. Nothing runs until Nomy is switched on.
- Policy: suggest only (the default), act on high confidence (threshold 50 to 100, default 80) or full auto within guardrails, with overrides per kind for pause, activate, budget, bid, hints and daily cap.
- Only pause, activate, budget, bid, hint and daily cap changes can be applied automatically. New campaigns, new ad groups, creative, conversion event settings, objective or bidding changes and archiving always need a person.
- Guardrails: the daily spend cap, bid and budget ranges, change limits and protected entities apply to every change; step limits and cooldowns also apply to rules and Nomy.
- Every change is journaled with its source (user, rule, Nomy or sync) and status; an executed proposal can be reverted.
- Shadow mode records every intended change without sending it to the platform.
What is an AI manager for ads?
Software that does the recurring work of a campaign manager: it reads performance, conversions and stock, adjusts bids, budgets, context hints and status, looks for what is not advertised yet, and explains what it did. Nomy is Adsonomy's AI manager for ads inside ChatGPT. Every change it makes passes the same guardrails and lands in the same journal as a change made by a person.
The OpenAI Ads platform supplies the levers: fixed bids, daily or lifetime budgets, context hints, pause and activate, product sets. It supplies no judgement and no safety rails; there is no target CPA, no pacing, no test mode and no notification when something changes, as the Ads Manager versus API guide sets out. A manager that works those levers has to bring its own constraints.
What does Nomy do, and when?
Three scheduled tasks, each switched on separately: a daily brief, an optimisation run at the cadence you choose (hourly, daily or weekly) and a weekly scout. Nothing runs on a schedule until Nomy is switched on, and runs start at the hour you set in the workspace timezone.
The optimisation run builds a snapshot of the account and checks four things first: serving issues, stock-outs, bids against your target and spend pace. It then proposes changes, each with a rationale, a confidence and a risk. The scout looks for products and pages without ads, landing pages from Google Analytics 4 and moments in the calendar, and proposes new ad groups, context hints or research; it applies nothing itself. On request, Nomy also drafts narratives, ad group angles with hints and copy, and writes rules from a sentence for you to review.
What goal does the manager optimise?
The one you set in Nomy's strategy: an objective (sales, leads, traffic or awareness), a target (CPA, CPO, ROAS, CPM, CPC or spend pacing only) and a target value. A temperament sets the size of the steps: conservative takes half steps, balanced full steps and aggressive one and a half, never above the guardrail. You also give a growth phase, your gross margin and notes for the manager.
A single clear target matters because the manager trades everything else for it. A bid controller compares the measured value with the target and steps the bid toward it, within the maximum bid step and the bid range you allow. Every daily brief leads with numbers, so you can see how far the account is from the target.
Which guardrails are hard?
Guardrails are ceilings and floors in the account currency that every change is clamped to or blocked by. The daily spend cap, the minimum and maximum bid, the minimum and maximum daily budget, the maximum number of changes per entity and per day, and protected entities apply to every source: a person, a rule or Nomy. Rules and Nomy are also held to a maximum bid step, maximum budget increase and decrease per change, and cooldowns per kind of change.
Some limits sit on the platform itself, outside Adsonomy's reach. A daily budget is an average rather than a hard daily cap: Ads Manager states that maximum daily spend is 2x and maximum seven-day spend is 7x the daily budget, so a 65 euro daily budget can spend up to 130 euro in one day and 455 euro in seven. The API also sets a floor: a daily campaign budget on a euro account must be at least 15.00 euro. Account-level spend limit windows, up to 60 non-overlapping date ranges each with its own ceiling, are a capability OpenAI enables on the account, and Nomy has no action that changes them.
In Adsonomy, a bid outside the range is clamped and the journal marks the change as clamped. A change to a protected entity is refused. A budget that would leave no room under the daily spend cap is blocked with the reason, for example: "No room under the daily spend cap of 320.00: the platform may spend twice a daily budget in one day, and the other active campaigns already take 160.50 of the 160.00 that leaves." Everything created on the platform starts paused.
How does the policy work?
One policy for the manager, with overrides per kind of change. Suggest only, the default: "Every change waits for your approval." Act on high confidence: "Changes at or above the confidence threshold are applied; the rest wait for you." Full auto within guardrails: "Every allowed change is applied within the guardrails; a few kinds always wait for you."
The threshold runs from 50 to 100 percent and defaults to 80. For pause, activate, budget, bid, hints and daily cap you can override the policy per kind: follow the policy, always auto, always suggest, or never. Only those six kinds can ever be applied automatically, and only by an optimisation run.
Some kinds always need a person, whatever the policy: new campaigns, new ad groups, creative, conversion event settings, changes to objective or bidding type, and archiving, which the platform cannot undo. An approved new ad group is created paused. Creative, new campaign and fix-issue proposals are advice: approving one records your decision, and you carry out the work.
What does the journal record?
Every change, with when it happened, its source (user, rule, Nomy or sync), its kind, the entity, the change itself and its status: pending, applied, shadow, blocked, failed or reverted. The reason, whether a guardrail clamped it, the ad account and the actor are kept with it.
Because the platform has no webhooks, Adsonomy reads the account on a schedule: campaigns and daily results every 15 minutes, hourly impressions, clicks and spend every 30 minutes. A proposal that was executed can be reverted from the proposals list; creating and archiving cannot be undone. A proposal you approve is journaled with you as its source.
How do you rehearse and stop Nomy?
Rehearse with shadow mode, stop with the switches you already have. Shadow mode is a setting on the ad account: "Record every intended change without sending it to the platform. Recommended for the first 48 hours." Changes are journaled with the status shadow, and going live means typing LIVE to confirm.
There is no single kill switch. To stop automated changes, switch Nomy off, which stops its scheduled runs, or set the policy to suggest only; set a kind of change to never to exclude it; and add campaigns you want left alone to the protected entities. Changes Nomy already applied stay in place until you revert them or change them yourself.
What is in the daily brief, and how is Nomy checked?
Four short sections with fixed headings: what changed, what is at risk, what Nomy did (or would have done under the current policy and shadow state), and three numbers that matter today. At most 150 words, each section led by a number, using your own entity names and figures. When conversions are zero, the brief says why from the tracking set-up.
Nomy's own work is measured too. Seven days after a proposal is applied, Nomy compares the seven days after with the seven days before, and the dashboard counts what is waiting, what was applied automatically and the outcomes. The brief respects the platform's data timing: conversions need at least a day to settle, hourly data covers impressions, clicks and spend only, daily insights are kept for 365 days and hourly data for about 30 days. Read the platform guide for reporting details, or the shop guide and service guide for the signals the manager reads.
Frequently asked questions
Can the AI raise my spend limit?
Nomy has no action that changes account spend limit windows. It can change campaign budgets within your budget range and the daily spend cap guardrail, and it can propose a change to the account daily cap; set that kind to always suggest or never if you want it to wait for you.
What happens if I disagree with a change?
Revert it. An executed proposal can be reverted from the proposals list, and every change is in the journal with its source and before and after values. When you dismiss a proposal you can add a note on why, which is kept to teach the manager. Creating and archiving cannot be undone, and neither is ever done by Nomy on its own.
Is Autopilot a different product from Control?
No. Autopilot is a switch on top of Control. The rules, the journal, the guardrails and shadow mode are the same; the difference is that Nomy proposes and, where your policy allows, applies changes itself.
Adsonomy runs ads inside ChatGPT for shops and service businesses. Control edition: you run it with rules. Autopilot: Nomy runs it inside your guardrails. Free until 15 October 2026, then plans from EUR 49 a month.
Start free