A company with 500 employees buys ChatGPT Plus for everyone — $20 per person per month. That's $120K a year. Doesn't sound like much, right?
The problem was never the $120K.
The Subscription Illusion
Every AI company uses the same pitch: just the price of a coffee each month. For individuals, sure. But for enterprise decision-makers, this pricing model hides three risks that never make it into the contract.
First is the scale paradox. When 5 people use it, you're buying a tool. When 500 people use it, you're buying a black box system you have no capacity to audit. Nobody knows what data each employee is feeding into the API, nobody knows if prompts contain customer information, and nobody knows what the model providers are doing with those conversation logs. OpenAI's enterprise version promises not to train on data, but how do you verify that? Does the contract say "won't train" or "won't use for training"? Legally, those are universes apart.
Second is hidden costs. You think $20/month covers everything, but in practice you discover you need: extra security audit tools, AI usage policies, employee training on proper usage, and dedicated staff to manage API key permissions. Gartner estimates the hidden costs of enterprise AI deployment are typically 2-3x the license fee. That means the real cost per person isn't $20 — it's $40 to $60 per month.
Third is the most致命 — exit costs. Two years of Copilot, two years of Claude, two years of various AI plugins, and your workflow is embedded in these tools. The vendor raises prices 30%. What now? Switch? Your team has adapted to that interaction pattern, your automation scripts call that API, your document formats match that output. This isn't buying software — it's building yourself a maze with no exit.
Enterprise Procurement Logic Has Changed
Previously when you bought SaaS, you bought certainty: you knew what features you'd get, what the SLA was, where data lived. Now when you buy AI subscriptions, you're buying a "might be useful" promise. Models update, capabilities shift, outputs drift — the same prompt that works today might not work tomorrow.
Some companies are trying a different approach: instead of buying a pile of $20 personal subscriptions, they concentrate their budget on deploying controlled AI infrastructure. Self-hosted or rented private models, data stays in-house, costs are predictable. Higher upfront investment, but over three years the total cost isn't necessarily higher, and you keep control.
This doesn't mean every enterprise should self-host models. Subscriptions make perfect sense for small teams. But when your AI spending exceeds 5% of IT budget, it's time to sit down and honestly calculate this bill — not how much you pay monthly, but who you'll be locked in with three years from now.
As a post on Hacker News put it: Every AI Subscription Is a Ticking Time Bomb for Enterprise. The bomb might not go off, but putting a bomb in your basement and pretending not to see it isn't smart management.
Primary sources:
- Hacker News Discussion
- Gartner Enterprise AI Cost Analysis Framework
- OpenAI / Anthropic Enterprise Terms of Service