The AI Versus SaaS Story Is Missing a Ledger

The claim arrived in familiar packaging. A crypto media outlet published a story this week asserting that small businesses are replacing Salesforce and HubSpot with custom AI tools for “pennies on the dollar.” No model names. No customer deployments. No cost model. No timeline. For anyone who spent 2017 auditing ICO smart contracts, the odor is unmistakable: a narrative running ahead of evidence. On paper, the case is seductive. LLM inference costs have fallen by an order of magnitude since 2022. A small sales team can now generate follow-up emails, summarize discovery calls, and flag hot leads for a few dollars a month. The marginal unit of software — the text token — is nearly free. So why pay $75 per seat for a CRM? This is the question the article raises. It does not answer it. It offers a headline, a belief, and no data. The missing evidence matters because the cost comparison is wrong. The marginal cost of inference is not the total cost of ownership. A CRM is not a language model. It is a data layer with permissions, audit trails, integration contracts, and regulatory commitments. When a small business “builds a custom AI tool” on an LLM API, it gets the generation layer. It still needs the data foundation, the workflow engine, and the failure handling. Those components are not pennies. Start with the technical stack. The custom AI tool described in the article, if it exists, is almost certainly not a custom model. It is a workflow: LLM API plus retrieval-augmented generation, function calling, and a low-code orchestration layer. That is combinatorial innovation, not architecture innovation. It can be built in weeks, but it can also be copied in weeks. The moat is zero. The real platform dependency has only shifted. Instead of paying Salesforce, the small business now pays OpenAI or Anthropic and sends its customer data through their APIs. That is not independence. It is a change of landlord. I learned this lesson in 2017, when I spent six weeks writing Python scripts to verify ICO token distributions. The scripts caught three calculation errors and stopped a $200,000 investment in a fraudulent project. The key insight was not the code. It was the difference between a demo and a system. A smart contract can look immutable until you test edge cases. A custom AI CRM can look cheap until you test data migration, permission inheritance, and the moment an employee leaves. The cost curve bends upward exactly when the business starts to depend on it. The “pennies on the dollar” framing hides the second ledger. Data cleansing, legacy migration, integration engineering, error handling, and ongoing maintenance all sit outside the API bill. Add the cost of a data breach. Add the cost of a hallucinated promise in a sales email. Add the cost of a prompt injection attack that leaks a customer list. CRM data includes contact details, transaction records, and contract terms. Sending that data to a third-party model creates GDPR and CCPA exposure. Traditional SaaS includes enterprise security commitments, SLA guarantees, and audit certifications. The custom AI tool includes none of that unless it is engineered in — and engineering is exactly what the headline leaves out. Exit strategies are written in ice, not in hope. Capital preservation starts by counting the liabilities. Competitive response makes the story worse for the “cheap replacement” thesis. Salesforce has Einstein. HubSpot has AI embedded across its tiers. Incumbents are not stationary. They will bundle AI into existing subscriptions, narrow the price gap, and convert the cost advantage into a feature comparison. The more interesting dynamic is not replacement; it is downgrading. Small businesses may keep the CRM as a contact directory but move their real sales logic into AI-native tools. That is a serious revenue threat to incumbents, but it is not the clean “pennies on the dollar” revolution the headline implies. Here is the blind spot. The article frames the winner as small business. It is not. The winner is the model layer. Every custom tool built on an LLM API accelerates the dominance of the platform underneath. The small business gets a convenient workflow. The platform gets proprietary business data, usage patterns, and pricing power. If the API price doubles after the tool is embedded, the pennies narrative dies. This is exactly what we see in crypto when applications build on layer-2 networks without controlling the base layer: the application looks independent until the base layer changes a parameter. The same logic applies to the AI-crypto narrative. “Decentralized AI” often means a centralized API wrapped in a token. A ledger without a story is a spreadsheet. A story without a ledger is a candle in a data center. The contrarian position is not that AI tools fail. It is that the disruption story is being told too late and in the wrong direction. The true signal is not “small business builds custom AI.” It is vertical AI-native tools undercutting SaaS entry points at the low end. Those tools will either be acquired by incumbents or forced to move up-market, where enterprise compliance demands destroy their cost advantage. Small business will not escape the platform economy. It will simply switch from buying seats to paying per outcome, or per token. So track three numbers. First, the number of small businesses running customer data through production LLM workflows for more than six months. Second, the retention and churn rates of those deployments. Third, the AI attach rate inside Salesforce and HubSpot. When those numbers exist, the “pennies on the dollar” story becomes a testable thesis instead of a mood ring. Until then, treat it as narrative. In a bull market, narratives outrun audits. In a bear market, audits survive. Exit strategies are written in ice, not in hope.