Operating record 001
We gave an AI executive $300 and an outcome. It built the ad campaign itself.
On August 19, 2026, Granotic’s AI growth executive was not asked to write ad copy. It was given an outcome - qualified traffic - and a deliberately bounded $300 experimental authority envelope, with real money behind it. The experiment began without a pre-built paid-acquisition execution path: no Granotic Google Ads account, no campaign, no campaign-management integration. This is the primary-source record of how an AI executive established that path and launched an international, multi-market acquisition experiment - the boundaries where it stopped, the platform defaults it caught, and the things that went wrong. The experiment is running now. We are publishing the method first; the results will be published when the data means something.
The campaign has begun serving and the end-to-end acquisition path - ad, landing page, attribution, product activation - is operational, under a $30/day budget and a $300 total ceiling. We will not publish performance numbers until enough spend and traffic exist to support an informative statement. This page will then be updated with total spend, impressions, clicks, cost-per-click, performance by search intent, Granotic starts, companies founded, meaningful activation, paid conversion if any - and the final scale, iterate, or kill decision. Whatever the result is, it will be published.
Three words that change the job
Most founders use AI the same way: ask for an artifact, receive an artifact, then do everything around it yourself. Ask for ad copy and you get ad copy - but you still open the ad platform, create the account, find the budget setting, discover what the platform enabled without asking, and become, in practice, the media buyer. The AI did a task. You ran the operation.
This experiment started from a different sentence. The founder said, in effect: get us qualified traffic - and approved a deliberately bounded $300 experimental envelope for it. No keyword list, no channel decision, no campaign brief. And no inherited machine to run it with: at the start of the mandate there was no Granotic Google Ads account, no active paid campaign, no campaign-management integration, no prepared paid-acquisition operating system. The difference between those two openings - “write me ads” versus “get us traffic” - is the difference between an AI tool and an AI executive, and it is the subject of this record.
An AI executive is an AI system that owns a business outcome rather than a task: it investigates the current state, decides what work the outcome requires, executes that work through real business systems, and stops for human authority only where authority is genuinely human - identity, money, law, and irreversible acts.
The distinction has three levels, worth separating precisely, because the market currently sells all three under one word. An assistant answers when asked and stops when the conversation ends. An agent completes a defined task through tools and reports back. An executive is accountable for an outcome over time: it must notice what is missing, sequence its own work, operate the systems, watch the results, and defend its decisions with evidence. Nothing about that third level is guaranteed by model quality. It is produced by an operating structure - and the structure is what this experiment tested.
The authority envelope
Delegating an outcome that spends real money requires an explicit contract. Ours had a name inside the workstream: the authority envelope. This was not a sandbox and not a simulation - a real card sat behind the account, and the executive was allowed to operate autonomously against it, inside the walls.
An authority envelope is the explicit boundary inside which an AI executive may act without asking: what it may spend, what it may configure, what it must never do, and which events hand control back to a human. The envelope is approved before execution - so autonomy is a granted, bounded thing, not an assumed one.
The envelope for this experiment, verbatim from the working record: maximum total spend $300; maximum daily budget $30; no Display Network; no Search Partners; no political advertising; no income or get-rich claims; no pricing changes; no expansion to another paid platform. Inside those walls, the executive could create, configure, launch, pause, and rebalance freely. Outside them, everything waited for the founder. Boundaries as policy - not as hope - are what make outcome delegation rational. It is the same principle Granotic applies to every executive in an autonomous company: autonomy is data, inspectable and adjustable, never a vibe.
Establishing the execution path
It audited before it acted. The first move was not creative. The executive inspected the company’s actual acquisition state and found an asymmetry worth being precise about: measurement already existed and was good - analytics on both website and product, campaign parameters captured on every entry point, a cross-domain identity handoff connecting an ad click to eventual product activation. Paid-acquisition execution, by contrast, did not yet exist as a path: no ad accounts, no billing, no platform access, no campaign tooling. Measurement-ready; execution path not yet built. That audit reshaped everything that followed - the bottleneck was not ideas, it was infrastructure and authority, so establishing the execution path became the work.
It mapped which steps genuinely required a human. From the audit came a short list of acts the executive classified as founder-only, before touching anything: creating the Google Ads account (legal terms), attaching payment (financial authority), advertiser identity verification (a person’s identity), the EU political-advertising declaration (a legal statement), and the final “go” (spend authorization). Everything else - keyword research, campaign architecture, ad copy, targeting, bidding, guardrails, monitoring - it claimed for itself. The founder’s total contribution to standing up a live advertising operation was roughly an hour of genuine authority acts, guided step by step. The founder never once operated the campaign.
It designed an experiment, not just ads. The campaign is a message-discovery instrument: three ad groups testing three different mental models a founder might search with. BUILD (“ai business builder” - a tool that builds it for me), IDEA (“turn my idea into a business” - I have something real and no path), COFOUNDER (“ai cofounder” - I want an entity that carries this with me). Eighteen keywords, exact and phrase match only, a $3 cost-per-click ceiling, and - deliberately - no smart bidding: the campaign’s true conversion event lives server-side in the company’s own analytics, invisible to the ad platform, and an optimizer that cannot see the goal optimizes toward noise. Which framing wins is a hypothesis under test, not a claim.
It went international on day one. The experiment launched as an international, multi-market test across seven English-speaking and English-capable markets - the United States, the United Kingdom, Canada, Australia, New Zealand, Singapore, and Ireland - deliberately configured so the evidence can begin showing where founder demand actually lives, rather than assuming it. Not global, and not claimed to be: seven named markets, chosen for language and founder density.
It armed its own kill rules before launch. Any intent cluster that spends $80 without producing a single click on the landing page’s call to action - “Tell Casper my idea” - dies. If $180 of the envelope produces zero founding starts anywhere, the experiment itself dies, and the conclusion is recorded as “this message does not convert on this channel” - a result, not a failure to hide. At $300 the campaign pauses unconditionally. Deciding how you will stop before you start is most of what separates an experiment from a gamble.
The defaults tax
Here is the section that is useful even if you never delegate anything to an AI. Building one Search campaign, the executive encountered - and corrected - six defaults that would each have silently spent money against the experiment’s interest. Google’s campaign flow enables them unless you notice:
| Default (enabled or preselected) | What it would have done |
|---|---|
| Search Partners: ON | Spread budget across hundreds of non-Google search sites - junk traffic for a message-discovery test. |
| Display Network: ON | Leaked search budget into banner inventory with entirely different intent. |
| “Presence or interest” targeting | Served ads far outside the seven target markets to anyone “interested in” them - contaminating the geographic evidence. Set to presence-only. |
| Campaign language default | The preselected campaign language did not match the intended international English-speaking audience - an English-keyword campaign would have served non-English-language users. Caught and set to English before launch. |
| AI Max keyword expansion | Let the platform’s AI broaden matching beyond the 18 chosen keywords, dissolving the three-cluster design. |
| Enhanced conversions: pre-checked | A consent checkbox agreeing to send hashed customer data to Google - a data-sharing decision that belongs to the company, not to a default. Unchecked. |
None of these are scandals; they are growth defaults, tuned for the advertiser who configures nothing. But they compound into a quiet tax on everyone who “just sets up ads.” An executive that owns the outcome has a reason to fight every one of them. A tool that was asked for ad copy does not.
What went wrong - and why that is the point
The honest version of this story includes three incidents, and they are more instructive than the parts that went smoothly.
The interrupted save. Mid-build, Google demanded an identity re-verification - a founder-only act. After the founder completed it, the campaign editor looked correct but the review screen disagreed: it showed targeting as “all countries and territories.” The executive stopped the launch, forced a full reload to read true server state, and found the seven-market targeting intact - but one legally required field, the EU political-ads declaration, had silently reverted during the interruption. It was re-set and re-verified. Without that stop-and-verify reflex, the experiment would have launched worldwide - the single most expensive mistake available that day.
The measurement trap. Immediately after publishing, the platform offered a helpful-looking step: “finish conversion setup” by reconfiguring the company’s live website analytics tag - with a warning, in small print, that existing tag settings would be lost. The executive declined. The company’s production measurement was demonstrably working and someone else’s domain; no campaign convenience justified touching it.
The zero-spend launch. The platform’s publish flow offers no “publish paused” option - the moment you publish, you can spend. The executive refused to rely on ad review as a safety net and engineered its own: it set the campaign’s start date two days into the future before publishing (a scheduler-level guarantee of zero spend), paused the campaign as a second independent lock, finished building the remaining ad groups and a 39-term negative-keyword wall while the campaign was inert, re-verified everything against server state, and only then moved the start date to launch day and activated.
The lesson we would offer anyone building autonomous systems: autonomy is not the absence of problems. It is the ability to detect, reason about, and recover from them - without handing the whole operation back to a human.
Where the executive stopped
- Identity: account creation, verification challenges, and two-factor setup were the founder’s, every time. The executive never touched a credential.
- Money: the payment method was attached by the founder; the spending envelope was approved by the founder; the executive operated only inside it.
- Law: terms of service, the advertiser’s legal identity, and the EU political-advertising declaration were presented to the founder with exact wording, never answered on his behalf.
- Irreversibility: permanent account choices (currency, time zone) and the final launch confirmation stopped for an explicit yes.
This is the practical answer to “what should an AI be allowed to do in a real business?” Not a philosophy - a working division of labor. The human is not the operator; the human is the authority. Everything mechanical, analytical, and procedural belongs to the system. Everything that binds the company legally, financially, or irreversibly belongs to a person.
The honest limits
This experiment also documented, precisely, where autonomy ends in August 2026 - and publishing that seems more useful than hiding it. The executive cannot create accounts or handle credentials (by design - identity is human). It cannot complete CAPTCHAs or bot checks (also by design). Its platform access ran through a browser session a human had authenticated, because the official ads API requires an approval process that takes weeks. And its monitoring loop, while automated, currently depends on infrastructure that must be made fully persistent - a gap recorded as an engineering dependency the moment it was noticed. An autonomous company, today, is a system of explicit human gates connected by wide autonomous corridors. The corridors are widening monthly. The gates - identity, money, law - should not.
Why this is different from “AI made my ads”
Generating campaign assets is a solved problem and has been for years. What this record documents is different in kind: an AI that discovered the work, established the infrastructure, sequenced the execution, operated the platform, defended a budget, enforced its own stopping rules, and knew which decisions were not its to make. The artifact was never the point. Ownership of the outcome was.
This single controlled experiment does not prove the operating model. It tests one AI executive operating one acquisition channel under a defined authority envelope, published with its method exposed and its results pending. But if you remember three things from this record, make them these: the AI executive was given an outcome, not a campaign task; it established the paid-acquisition execution path itself, operating real money under bounded authority; and it independently built and launched an international, multi-market acquisition experiment whose market result will be published here when the evidence is real. That shape - outcome in, authority envelope around, evidence out - is how every Granotic executive works, whether the outcome is traffic, a brand, or a functioning company. This was simply the first time we let one buy its own ads and wrote everything down.
If the interesting question after reading this is not “how do I configure Google Ads” but “what outcome would I hand to an executive team that works like this” - that question has a place to go. From idea to functioning company →