"Code is law, but bugs are reality."
That sentence was running through my head as I parsed a Crypto Briefing article claiming small businesses are replacing Salesforce and HubSpot with custom AI tools for "pennies on the dollar." The article contained no model names. No cost breakdown. No customer data. No security analysis. It was a title assertion plus two paragraphs of generalized commentary. In protocol engineering, we call that a state transition without a proof. The underlying signal deserves analysis, but the narrative around it is dangerously compressed.
What is the real signal? Small businesses are indeed building lightweight AI workflows around LLM APIs. They are writing follow-up emails, summarizing sales calls, and scoring leads. These are high-frequency, text-heavy, narrow tasks. For those tasks, custom AI tools can be dramatically cheaper than a Salesforce seat. But "cheaper per task" is not "cheaper total system." Anyone who has audited smart contracts knows the difference between gas cost and verification cost. A transaction costs a few cents to execute. Proving it is correct costs weeks of manual invariant tracing. In 2019, I spent three months dissecting Uniswap v1's constant-product market maker, finally finding an overflow in eth_to_token_swap_input by tracing the arithmetic by hand. The execution cost was negligible. The audit cost was not.
The same asymmetry applies to AI-powered CRM replacement.
Technical context matters here. Small businesses are not training foundation models. They are composing existing models with retrieval-augmented generation, function calling, and low-code orchestration. This is compositional innovation, not architectural innovation. In blockchain terms, it is the difference between deploying an appchain and securing a base layer. Anyone can spin up a chain with OP Stack. The security model, however, still inherits from Ethereum. Similarly, a "custom AI tool" built on OpenAI or Anthropic APIs inherits its reliability, pricing, and data handling from those platforms. The value does not fully flow to the small business. Some of it flows to the model API provider. The article ignores this dependency.
The phrase "pennies on the dollar" is the core problem. It refers to marginal inference cost, not total ownership cost. To replace a CRM, a small business must build a data model for leads, accounts, opportunities, and activities. It must integrate email, calendars, and billing systems. It must manage permissions, handle API rate limits, and log every interaction for audit. It must maintain the system as models and APIs change. Those are engineering costs. They scale with process complexity, not with token count. A ten-person sales team that wants to replace Salesforce is not writing a prompt. It is building a mini-CRM. The initial build might cost a few hundred dollars in API calls. The integration and maintenance will cost an order of magnitude more, and they will recur every quarter.
Let me be specific about where replacement actually happens. Based on my experience auditing data-dependent systems, I would split the CRM surface into five layers:
- Sales email drafting and call summarization: high replacement potential, 40-70 percent in 6-18 months. This is where AI shines.
- Lead data entry and deduplication: moderate, 30-60 percent. Depends on API integrations and data quality.
- Full customer lifecycle management: low, 10-20 percent. Cross-department workflows require governance.
- Forecasting and revenue analysis: below 10 percent. Data quality and model drift make AI outputs unreliable.
- Compliance, audit, and permissions: below 5 percent. Regulatory risk is too high.
This is not a single "replacement" curve. It is a spectrum. The article compresses all of these into one dramatic claim. That compression hides the fact that even the most successful AI deployment still needs a reliable system of record. AI tools are execution layers. They need a data substrate underneath them. The substrate can be Excel, Airtable, a database, or still Salesforce. The article assumes that because AI writes the email, the CRM is dead. In reality, the CRM's value is not writing emails. It is holding the truth about customers.
That is why I find the competitive analysis in the article especially shallow. Salesforce and HubSpot are not stationary. They are adding AI features to their existing platforms. They are defending their pricing power with Einstein and embedded assistants. Their moat is not the UI. It is the data schema, the integration ecosystem, and the compliance certifications. Those assets take years to accumulate. A small business building a custom tool on an LLM API has none of those assets. It can start quickly, but it will hit a wall when the business needs a second person, a third department, or a SOC 2 report.
Here is the contrarian angle: the article's "small business builds custom AI" framing is a fantasy. Most small businesses do not have the talent or patience to build and maintain custom AI workflows. They will rent those workflows from a new generation of vertical AI startups. Those startups will charge less than Salesforce per seat, but they will still be platforms. The real shift is not from SaaS to self-sovereignty. It is from one platform to another, with the model layer extracting a new tax on top.
That has a crypto parallel. The crypto community loves to talk about replacing trusted intermediaries with code. Then it builds protocols on top of AWS and Infura. Code is law, but bugs are reality. The same happens with AI: the "custom tool" is just an API call to a centralized model provider. If that provider changes pricing, deprecates a model, or adjusts its safety filters, the tool's behavior changes overnight. No small business can fork the model. They are tenants, not owners.
The security blind spot is even more serious. CRM data is among the most sensitive data a small company holds: contact details, contract terms, financial history, support logs. Feeding that data into a third-party LLM API creates legal exposure under GDPR and CCPA. It creates prompt-injection attack surfaces. It creates data-residency questions. The article celebrates low build costs while ignoring these liabilities. In my audits, I have seen teams treat compliance as an afterthought. That is exactly how customer data ends up in a training set.
"Zero-knowledge isn't mathematics wearing a mask," as I often tell colleagues. It is a discipline that guarantees a fact without revealing the fact itself. Most AI tools have no such guarantee. They give you a probability, not a proof. For marketing emails, probability is fine. For contracts and compliance, it is not. The market doesn't reward truth; it rewards settlement. And settlement requires an auditable record of who made which decision, based on what data, with what guardrails.
So what should we actually watch? Not the article's headline. The measurable trend is the growth of vertical AI sales tools with real adoption. I want to see customer retention after 12 months. I want to see the actual cost of integrating and maintaining those tools. I want to see whether Salesforce and HubSpot lose small-business logos to AI-native startups. Until then, "pennies on the dollar" is just a narrative token. It has no backing collateral.
The takeaway is simple: the old SaaS pricing model is under pressure. The pressure is real. But "custom AI tools" are not the alternative. They are the opening move in a re-platforming game. The winners will be those who combine cheap inference with expensive trust — auditable data flow, compliance-grade security, and deterministic state. That is hard. That is expensive. And that is why Salesforce will not die this quarter. The article's mistake is setting the state transition at zero. Reality has gas costs.