The AI Iceberg: Why Enterprise AI Costs 50x More Than You Think

Since establishing The AR Dept. at the start of the year, I've seen firsthand how the current AI arms race between OpenAI, Anthropic, SpaceXAI, etc. is creating a dangerous misconception when it comes to deploying AI within an organisation. Consumer-facing apps have conditioned users to expect an infinite well of capability for the price of a streaming subscription. But when organisations attempt to transplant this consumer convenience into the engine room of a business, the pricing shifts dramatically. To illustrate this gap, we ran a test over the summer, and the findings suggest that the enterprise dream of cheap, limitless automation is, for now, an expensive fiction.

The Test

We used a Claude Max 20x plan (cost: £150 a month) to determine the gap between subsidised consumer access and direct API costs. The experiment involved a dummy repository for a personal health application, Project 90, where we fed development tickets to a worker agent via a GitHub project board for thirty days.

The agent churned through code, testing and reviewing around the clock, with every request, prompt, and output token logged and priced against Anthropic’s API rates, the exact unit prices a business pays when integrating these models directly. The final bill showed a significant gap between our subscription cost and the market value of the compute, which looks good on paper but is unlikely to be sustainable in practice.

We used a GitHub project board as our command centre. We populated this board with a constant stream of development tickets, which were then fed to our worker agent. This configuration ensured the agent remained in a state of perpetual engagement, processing tasks without the pauses inherent in human-in-the-loop workflows.

By automating the intake of these tickets, we forced the agent to operate at a tempo that mirrors the production environment of a modern software firm, effectively removing the idle downtime typical of casual user behaviour. This approach allowed us to observe the agent's performance and consumption patterns under a consistent, high-intensity load, providing the raw data needed to calculate the real-world operational cost.

The Results

Over the thirty-day cycle, the agent processed 12.3 billion tokens and fielded 105,638 individual requests. When calculated at standard API list pricing, the consumption reached £8,443. To maintain perspective, 23 million output tokens equate to roughly 17 million words, or reading the entire Harry Potter series sixteen times. One agent subscription and worker consumed just under eight and a half thousand pounds' worth of server power for the price of a single £150 subscription plan. That is a 50x discrepancy.

The Conclusion

The economics are straightforward. Anthropic is a commercial enterprise with investors to satisfy, and two prices for identical work cannot coexist indefinitely. Nobody outside the company knows what a token costs to produce, but the gap is most likely held open by some mix of three things: venture capital buying market share by the billion, API margins that carry the subsidy, and the fine print, which says the advertised limits assume ordinary, individual use. We saw the same trajectory a decade ago with ride-hailing and food delivery. Subsidised convenience is followed by a price correction, and the true cost of production becomes the only reality.

If your business buys a handful of seats for drafting and document summaries, the current model is good value and will stay that way for a while. The danger starts when you build an operational process on top of it. Automating client onboarding or invoice reconciliation needs more than a personal login. It needs an audit trail, governance, guaranteed uptime and a bill that scales with the work, none of which a consumer plan is designed to give you. That is why none of the agents The AR Dept. run for clients go anywhere near a personal subscription. They run on the API, metered by the token, with every action logged, and we budget for them at that price.

So should you. When you plan next year's budget, do not anchor it to the artificial reality of a cheap monthly plan, and if a supplier promises to automate your back office on flat-rate logins, they are handing you a system built on borrowed time. If you would like to discuss the real cost of deploying AI, please get in touch.