Why Your Company Doesn't Need the Most Expensive AI Model

Summary

The speaker argues that companies should buy AI capability based on demonstrated task performance—not assume every new, more expensive model is a necessary upgrade. As cheaper models become “good enough” for more workflows, businesses can expand AI usage without increasing their bills. This creates a more price-sensitive market for model providers and puts pressure on companies using AI to pass savings or added value on to their own customers.

Business Strategy and Operating Playbook

• Evaluate models against real work. Define acceptable performance for specific tasks, test models on representative examples—including difficult cases—and choose the least expensive option that passes.
• Compare total cost, not just model price. Include human inspection and correction, tooling changes, and the number of iterations needed. A cheaper model may still be less economical if it requires substantial rework.
• Route work by complexity. Use lower-cost models for routine requests and reserve premium models for cases that genuinely need their capabilities. For example, cheaper models may handle most standard customer-support questions, with a router sending harder cases to a more capable model.
• Use evaluations in procurement and vendor negotiations. Evidence that a cheaper model meets requirements gives buyers leverage to switch tiers, compare providers, or negotiate discounts while staying with the same supplier.
• Invest in durable capabilities. Prioritize organized customer information, integrated systems, reliable quality evaluation, and maintainable workflows. These can make future models easier to adopt and may outlast technical workarounds designed for a particular model’s weaknesses.
• Plan for model improvements. Before investing in a complex AI system, ask how long its advantage is likely to last and whether a stronger future model could make it unnecessary.
• Prove value beyond access to AI. As AI lowers service-delivery costs, customers may expect lower prices. Businesses need to identify and communicate the additional value they provide rather than relying on AI use alone to protect margins.

Metrics and Examples

• Ramp spending data: An October report described almost no token-spending growth in August and September despite rising usage. Ramp’s November 11 update showed AI spending up 1.65% in October, which the speaker characterizes as effectively flat for three consecutive months.
• Adoption: AI adoption in Ramp’s sample reached 56.8% of businesses.
• Customer support: The share of spending attributed to customer success and support rose from under 1% in June to 13% in October. The speaker interprets this as a sign that these use cases are becoming practical for more companies.
• Model mix: Advanced models accounted for nearly 70% of spending in Ramp’s November report, with their share increasing slightly in October. The speaker notes that high spending on premium models does not necessarily mean they perform most of the work.
• Token-cost illustration: At $5 per million tokens, one billion tokens cost $5,000. At $1 per million, the same budget buys five billion tokens. Usage can therefore rise fivefold without increasing spend if work shifts to a cheaper tier.
• Coding example: Evaluate whether a model makes correct changes in the codebase and how much human verification it needs. If a cheaper model passes, use it. If it fails, the evidence may justify a premium model—or show that several cheaper-model iterations still cost less overall.
• Service-margin example: A service priced at $10,000 with $6,000 in delivery costs yields $4,000 in gross profit. If AI reduces delivery costs to $4,000, the provider could initially retain the savings, but competitors and customers may create pressure to lower prices.
• Infosys example: The company reported rapidly growing AI revenue, while customers pressed for AI-related productivity gains to show up in pricing. Its operating margin remained at 21%, illustrating that AI adoption does not automatically translate into higher margins.

Market Implications

• Lower prices can make previously uneconomical tasks worthwhile, potentially expanding overall AI usage—a dynamic the speaker associates with Jevons’ paradox.
• More usage does not guarantee proportionally higher revenue for model providers. Buyers may consume more tokens while shifting to cheaper models, demanding discounts, or negotiating enterprise contracts.
• The speaker expects AI demand to keep growing in 2027, but argues that technical progress alone will not establish commercial returns. Providers must manage research and infrastructure costs while demonstrating that customers will pay for the value delivered.
• The likely result is a combination of rapid AI adoption and tougher purchasing scrutiny: more workloads move to AI, while buyers become less willing to pay for capability they cannot show they need.

Sources

• Presenter: Nate B. Jones
• Sources cited in the video: Ramp’s October report and November 11 update; Infosys’s October earnings announcement