Skyfall's AI boss doubled revenue. With marketing homework

A startup buys small software businesses and lets a network of AI agents run them. Sounds like the end of management. Look at what the agents actually changed, though, and it's something far less dramatic: marketing homework the previous owners never got around to.

What happened

• Who: Skyfall AI, a San Francisco startup co-founded by Sam Pasupalak and Kaheer Suleman, among others. The pair founded Maluuba, which Microsoft bought in 2017.
• What: In July, Skyfall said it would spend up to $1 million on a small software or e-commerce business and double its revenue in six months. Now, per TNW, it has unveiled its system, Steve-001, and first numbers from three acquired businesses.
• Business 1: a social media API tool, acquired in July. Monthly revenue, per Skyfall: $3,649 at takeover, $8,307 in October, up 127.7%.
• Business 2: a Shopify referral app, acquired in August after five years of decline. From $8,634 to $8,379 a month, about 3% lower. Skyfall calls it stabilization.
• Business 3: too new to report.

The AI boss does what every good CMO does first

What I find interesting is the list of interventions. No brilliant new strategy. Instead: answer signup questions instantly. Stop promoting one-off credit packs, which Skyfall says brings about four new subscriptions a week. Offer cancellation discounts only to newer users. Rewrite the App Store listing. Describe the app so AI assistants can recommend it more easily. And switch off ads that drove clicks but no purchases, saving over $1,000, according to Skyfall.

Sound familiar? These items top almost every audit of a small business. They just never get done because the founder is buried in operations. My rule of thumb: AI doesn't fix broken ways of working, it widens the cracks in the foundation. This is the flip side. Whoever finally fixes the foundation sees results fast, human or machine. Choosing placements deliberately instead of trusting the automation was also the biggest lesson from our own Meta tests.

The real product is the approval logic

More interesting than any revenue curve is the policy engine. Every action lands in one of four tiers: just do it, do it and alert the owner, wait for approval, or stop. Refunds up to $30 go through automatically; anything larger needs a person. On top of that, Steve tests each move in a model of the business first, with a forecast, a timeframe and a confidence level.

That's human in the loop as an operating manual, not a gut feeling. And it's exactly what most marketing teams lack as they hand agents access to campaigns, budgets and customer emails.

The catch

All numbers come from Skyfall, and nobody has audited them. The base is tiny: going from $3,600 to $8,300 is a nice curve, not proof that AI can run a corporation. The second business is still slipping slightly. And Skyfall hasn't said whether anyone worked at the businesses it bought. A public dashboard per business is promised. That's when this can be judged properly.

What I'd do now

• Use Steve's list as an audit. Onboarding questions, cannibalizing offers, discount logic, store copy, ads without purchases. Five items, one week.
• Define four approval tiers. For every agent or automation: what can it do alone, with a heads-up, only with sign-off, never? With money limits.
• Write for AI assistants. Make your product page and store listing clear enough that an assistant can say in two sentences who you're good for.

Bottom line: The AI boss doesn't win on genius but on discipline. You can have that without it.