Running What You Can't Fully Control: The AI Cost Visibility Problem
You built something people want. Users are signing up. Revenue is growing. Then your OpenAI invoice arrives, and suddenly that $50/month customer cost you $90 to serve.
Welcome to the new reality of AI-powered SaaS.
The Control Paradox
Traditional software had a beautiful simplicity: build once, distribute infinitely. Marginal costs approached zero, and gross margins of 80-90% were standard. Every new customer was essentially pure profit.
AI broke that model completely.
Every API call to OpenAI, Anthropic, or Google has a price tag. Every customer query consumes tokens. Every feature interaction burns compute. According to Bessemer Venture Partners, AI companies now average 50-60% gross margins compared to traditional SaaS at 80-90%. That's not a minor adjustment—it's a fundamentally different business.
The uncomfortable truth? You don't control your most significant variable cost. Third-party LLM providers set prices, change models, and adjust terms. Your margins live at their mercy.
The Visibility Gap
Most founders discover this problem the hard way. One CFO from a data infrastructure company told Metronome: "We're not monetizing AI to juice revenue. We're monetizing to avoid eating $10k of costs on a $500 plan."
The pattern repeats across the industry. Companies charge flat subscription fees while incurring wildly variable AI costs per customer. Some customers barely touch AI features. Others hammer them relentlessly. Without granular visibility, you're flying blind.
Zylo's 2025 SaaS Management Index found organisations spent an average of $400k on AI-native apps—a 75% year-over-year increase. More concerning: finance teams frequently aren't notified until after charges are incurred, making forecasting nearly impossible.
The Power User Problem
Here's what kills margins quietly: your top 20% of users typically consume 70-80% of your API spend but represent only 20-30% of revenue. This inverted profit structure means growth doesn't automatically improve unit economics the way it does for normal software.
Think about that. Acquiring more customers might actually accelerate losses if you can't identify which customer segments are profitable.
A startup running 500 GPT-4 queries per user session that cost $5 in 2023 now pays less than 50 cents—but they're still losing money because they never built the attribution layer to understand costs at the customer level.
What Actually Works
The companies solving this share common approaches:
Track everything from day one. You cannot optimise what you don't measure. Per-customer, per-feature, per-model cost attribution isn't optional anymore—it's survival infrastructure.
Separate value metrics from cost metrics. Customers understand outcomes (tickets resolved, documents processed, tasks completed). Let those drive your pricing. Use token counts and API calls to drive internal limits and profitability analysis.
Build guardrails before you need them. Rate limiting, tiered access, and usage caps feel restrictive until a single power user craters your monthly margins. Design constraints into the product architecture, not as afterthoughts.
Accept that margins are now moving targets. Traditional SaaS CFOs managed relatively stable cost structures. AI requires continuous recalibration as models change, pricing shifts, and usage patterns evolve.
The Path Forward
Running an AI business means accepting a fundamental tension: the technology that creates your value also creates your most unpredictable costs. You can't fully control LLM pricing, model availability, or customer usage patterns.
What you can control is visibility.
Knowing exactly which customers are profitable, which features drive costs, and where your margins actually live turns an unpredictable cost structure into a manageable one. The companies that build this instrumentation early survive. Those that don't become cautionary tales about subsidising customers with investor money until both run out.
The AI margin squeeze isn't going away. The question is whether you'll see it coming.
Building AI-powered software and struggling with cost visibility? tknOps provides precise token tracking for multi-tenant AI applications, helping you understand your true per-customer profitability.