The $20M Prompt: Why AI Is Now the SaaS Killer

Building internal tools with AI beats buying SaaS whenever the job is mostly information retrieval, workflow glue, or a thin layer on top of data you already own. Buying still wins when you need a mature product, a support team, and someone else's legal liability. The gap between those two categories is shrinking fast, and a billionaire's yacht lawsuit just made it visible.

The $20M Prompt: What Actually Happened

Revolut co-founder Nikolay Storonsky is being sued by yacht broker Cecil Wright & Partners for roughly €17.5 million, about $20 million, in commission the firm says it's owed after Storonsky bought the 102.4-meter superyacht Nixie for around €350 million ($404 million) without going through them.¹

His defense is straightforward: he didn't need the broker's Rolodex. A ChatGPT search of publicly available information surfaced that Canadian businessman Patrick Dovigi had previously owned the yacht, giving Storonsky a thread to pull on his own instead of paying a broker to make the introduction.¹

Whether the courts buy that defense is a separate question. What matters here is the shape of the argument. A broker's entire value proposition was privileged access to information. A prompt got most of the way there for free.

Why This Is Bigger Than One Billionaire's Yacht

Strip away the yacht and the number, and the mechanism is the same one playing out in office software every day. Brokers, consultants, and plenty of SaaS vendors charge for one of two things: access to information you can't easily get yourself, or a workflow you'd rather rent than build. AI is quietly eating both.

A yacht broker sells access to ownership records and relationships. A niche analytics SaaS sells access to a dashboard you could otherwise wire up yourself. A $20,000-a-year vendor sells a workflow that a competent builder can replicate in Retool in a few weeks once the vendor's support queue gets slow enough.² The premium middlemen charge for "knowing things" or "having already built it" depends on the buyer not being able to do it themselves cheaply. AI just made that assumption a lot less safe.

How Much Are Companies Actually Spending on SaaS?

Here's how much money sits on the other side of this ledger. Average SaaS spend per employee hit $4,830 in 2025, up 21.9% year over year.³ The average organization now spends $55.7 million a year on SaaS and manages 305 separate applications.⁴ Renewal cycles have quietly become one of the largest recurring line items most companies never re-litigate.

That sprawl happened because buying was almost always faster than building. Nobody wanted to staff an internal engineering team to replicate Salesforce or Zendesk. For years, that math was correct. It's the "almost always faster" part that's now up for debate.

Build vs. Buy Has Already Flipped for Many Teams

This isn't a thought experiment. Retool's 2026 Build vs. Buy survey of 817 builders found that 35% of organizations have already replaced at least one SaaS tool with something they built themselves, and 78% plan to build more custom tools this year.²⁵ The categories most exposed are the ones you'd expect: workflow automation, internal admin tools, and BI dashboards, at 35%, 33%, and 29% respectively.²

The case studies back it up. ClickUp built six internal AI-driven automation tools connected to Salesforce, Zendesk, and Snowflake instead of buying more automation software, saving hundreds of hours of weekly work and cutting about $200,000 a year.²⁶ Startup discovery platform Harmonic rebuilt a $20,000-a-year third-party tool inside Retool after getting tired of waiting on vendor support, and now runs 33 internal apps wired into Salesforce, Gong, and Slack.² Notice the trigger for Harmonic wasn't just price. It was a vendor too slow to fix its own product, a governance failure on the vendor's side as much as a cost decision on the buyer's. This pattern shows up across the niche SaaS category that's bleeding customers to AI-built alternatives, not just in isolated case studies.

Wall Street Is Already Pricing This In

Investors noticed before most CFOs did. When Anthropic shipped specialized Claude Cowork plugins for legal, finance, and customer-support workflows in February 2026, software stocks sold off hard. Oracle fell about 4.2%, Salesforce about 3.3%, and Thomson Reuters and LegalZoom both dropped more than 15% in a single session.⁷⁸ The S&P 500 Software & Services Index was down roughly 20% year to date, and hedge funds had already shorted about $24 billion in software stocks.⁸

That's not a niche take. That's a market deciding, with real money, that AI-built internal tooling is a credible threat to recurring SaaS revenue, not a novelty. OpenAI's Sam Altman put it more bluntly, predicting SaaS is entering what he called the "fast fashion era," cheap, disposable, purpose-built tools replacing narrow use cases that used to require a subscription.⁹

The Shadow AI Problem: Ownership Without Governance

Here's the part that should worry engineering leaders more than the stock charts. Employees are already doing this, with or without a policy. Sixty percent of builders across seniority levels say they've built software, workflows, or automations outside official IT oversight in the past year.²⁵ Gartner projects that by 2027, 75% of employees will acquire, modify, or create technology outside IT's visibility, up from 41% in 2022.¹⁰

That's not a future risk. That's a description of what's already running in production at most companies, unmonitored. Every internal Retool app, every ChatGPT-assisted script, every ad hoc dashboard someone's ops lead built over a weekend is now an asset with no owner, no backup plan, and no security review. Shadow IT used to mean someone bought a rogue SaaS tool with a credit card. Shadow AI means someone built and shipped their own software, and it works well enough that nobody's going to ask permission to keep using it.

The Cautionary Tale: Klarna and the Limits of Rip-and-Replace

Before this turns into cheerleading, remember Klarna. In 2024, its OpenAI-powered customer service assistant handled 2.3 million conversations in its first month, roughly the equivalent of 700 full-time agents, and was projected to add $40 million in profit that year.¹¹ It was the case study everyone cited.

By May 2025, CEO Sebastian Siemiatkowski was telling Bloomberg the company had cut human staff too aggressively and was rehiring for premium support roles.¹¹ The AI could handle volume. It couldn't fully replace the judgment and trust that come with certain kinds of human contact. Replacing a tool isn't the same as replacing a team, and replacing a workflow isn't the same as replacing the accountability structure that used to sit behind it.

HubSpot co-founder Dharmesh Shah makes a related point from the builder's side. Despite running a company full of engineers who use AI coding tools daily, he says the number of core internal SaaS systems HubSpot has replaced with a vibe-coded variant is exactly zero.¹² His explanation isn't a lack of capability. It's a lack of focus: an engineering org has better things to do than re-litigate its CRM. Factory's CTO makes the same point from the vendor side, warning that agents can replicate a mature SaaS product feature by feature, but doing so is slow and expensive, and you still end up without the team of maintainers a real vendor has behind the product.¹³

What Does Building Internal Tools With AI Actually Look Like?

Strip away the hype and most AI-assisted building today falls into a few buckets:

• Code snippets and glue scripts. The most common use case by far: a script that pulls a report, reformats a CSV, or stitches two APIs together. Low risk, high frequency, and where most of the real time savings show up.
• Full internal apps built in low-code tools. Teams pairing AI coding assistants with platforms like Retool build actual applications, dashboards, and admin panels, often in days instead of the weeks a vendor eval used to take. That's the pattern behind how AI coding changed the build vs. buy math for a lot of teams this year.
• Feature-by-feature SaaS replication. The riskiest bucket. It's technically possible to have an agent rebuild a mature product's feature set, but it's slow, expensive, and leaves you without the maintainer team a real vendor has.¹³

Seventy percent of builders in Retool's survey had used ChatGPT for this kind of work, 56% had used Gemini, and 53% had used Claude. Fifty-one percent had shipped production software they built themselves using AI, and about half of those saved six or more hours a week.²⁵ The failure modes are familiar to anyone who's shipped code fast: hallucinated logic that looks plausible until it breaks in production, and technical debt that accrues quietly because nobody assigned an owner to what started as someone's side project.

Software as an Asset You Build, Not a Subscription You Rent

The deeper shift here isn't about AI coding speed. It's about what counts as a company asset. A SaaS subscription is a rental. Stop paying and you lose access, and the knowledge embedded in the tool's configuration walks out the door with the vendor relationship. An internal tool your team built is different: even a rough one is something you own outright, something you can inspect, modify, and keep running on your own terms.

That's the argument underneath every build-vs-buy decision worth taking seriously this year, and it's why platforms built for turning internal software into something teams actually own rather than perpetually rent are picking up real attention right now. Ownership isn't automatically cheaper. But it's the difference between an asset on your balance sheet and a recurring bill you never get to stop questioning.

What Should You Build vs. Buy This Year?

Don't panic-cancel your core systems of record. CRM, ERP, and anything with real compliance weight still benefits from a vendor's maintainer team and legal accountability, exactly as Dharmesh Shah and Factory's CTO argue.¹²¹³ Treat the following as genuinely up for grabs:

1. Workflow automation and glue tools. The single most-replaced category in Retool's survey, and usually the lowest-risk place to start.²
2. Internal admin panels and BI dashboards. High cost, low differentiation, and often built by a vendor slower to fix bugs than your own team would be. It's the same shift behind AI-generated dashboards eating niche analytics SaaS.²
3. Anything a vendor is slow to support. Harmonic didn't leave its vendor over price. It left over speed. That's a signal worth tracking across your own renewal list.²

The honest move for most leaders is to formalize a build lane rather than pretend shadow building isn't already happening, because 60% of it already is.²⁵ Audit what's out there, give it an owner, and decide deliberately what gets built versus bought, instead of discovering the answer in a lawsuit.