The Drug Dealer Model Behind AI Subscriptions, Agents and Token Costs

I started with one OpenAI subscription. At the time it felt like a useful extra tool, something I could experiment with, speed up a few bits of writing, maybe use for research or the occasional technical question. It did not feel like infrastructure. It did not feel like something my working day would start to organise itself around. It was just another subscription.

That has changed. I now have two OpenAI accounts, plus Anthropic and Gemini, and I regularly hit the five-hour and weekly usage windows when using Codex. That did not happen because I was casually playing with AI tools. It happened because they became genuinely useful. They got good enough to change the speed and quality of what I could deliver. They helped me move faster, test more ideas, write more, build more, review more, and take on work that would previously have been slower or more expensive to get through.

That is where the uncomfortable part starts. The better these tools get, the more dependent you become on them. At first they feel optional. Then they become part of your workflow. Then they become part of your delivery capacity. Once that happens, the commercial power shifts. If the price goes up, I do not have a clean choice between paying and not paying. I either absorb the higher cost or accept a productivity hit.

That is what I mean by the drug dealer model. The product does not have to be bad to follow the same commercial pattern. Make it easy to start. Make it cheap enough that nobody thinks too hard about it. Make it useful enough that people come back every day. Then, once it is embedded into how the work gets done, the customer has much less room to move.

We have seen versions of this before with cloud platforms, CRMs, productivity suites and SaaS tools. The difference with AI is that it sits much closer to the work itself. It is not just storing your documents or managing your pipeline. It is helping you think, write, code, research, analyse, summarise, plan and make decisions. That makes the dependency more subtle and more powerful.

For small businesses, this matters. AI adoption is often sold as a simple productivity win: plug this in, connect your data, automate the boring work and save time. Some of that is true. The productivity gains can be very real. I would not be using these tools heavily if they were not helping. But the commercial model underneath the productivity story deserves more scrutiny.

The first stage is access. Free trials, low monthly subscriptions, bundled features, generous usage limits and easy integrations make the decision feel low-risk. The second stage is habit. Teams start using the tool for meeting notes, customer emails, reporting, internal research, drafting, support replies and process shortcuts. The third stage is dependency. The business no longer just uses the software. Parts of the business now expect it to be there.

The next step is even more interesting, and probably more dangerous commercially: standalone agents. It is one thing to use AI as a tool you open when you need help. It is another thing to have agents running parts of the business in the background. Agents that monitor inboxes, update CRMs, draft responses, check data, write code, process documents, chase tasks, generate reports or handle customer workflows. At that point, AI is no longer just helping a person do the work. It is becoming a worker inside the operating model.

That creates a different kind of dependency because agents do not just cost a monthly subscription. They consume tokens every time they think, read, write, search, retry, call a tool or hand work back for review. The better they get, the more useful they become, and the more tempting it is to delegate whole chunks of operational work to them. But every delegated process creates a recurring usage cost. If an agent is checking messages every few minutes, reviewing data, summarising context and making decisions, the cost is no longer occasional. It becomes part of the cost base of running the business.

This is where the pricing risk becomes sharper. If you have a few AI subscriptions for personal productivity, a price rise is annoying. If you have agents handling lead qualification, customer support, reporting, admin, research or delivery workflows, a price rise hits the operating model directly. You either pay more to keep the same level of automation, reduce usage and take work back manually, or rebuild the workflow somewhere else. None of those options are painless once the agent has become part of how the business runs.

That is why token cost matters. It is easy to talk about AI agents as if they are digital staff, but digital staff still have a cost per action. Every background task, every retry, every context window, every tool call and every check adds up. Businesses need to understand that before they hand over important processes. Otherwise they may wake up with a system that works brilliantly, but only because it is constantly consuming paid inference in the background.

Once that dependency forms, future price rises, usage limits, model changes or product restrictions are no longer minor annoyances. They become operational risks. If your team has built its internal workflows around one AI platform, switching away is not just a billing decision. It means retraining people, rebuilding prompts, reworking automations, changing processes, losing context and accepting a period of slower delivery.

This does not mean businesses should avoid AI. That would be the wrong lesson. The gains are too useful to ignore, and the companies that refuse to engage with the technology will probably fall behind. The point is that businesses should be more deliberate about where dependency is forming. They should know which processes rely on which tools, what data is being sent where, which outputs can be moved, which workflows need approval, and what would happen if a provider changed the rules.

The practical answer is to treat AI like infrastructure, not a novelty subscription. Useful infrastructure always creates some dependency, but good businesses understand their dependencies. They document them. They design around them. They avoid putting all of their operational knowledge, workflow logic and delivery capacity into one vendor's black box without thinking through the consequences.

The companies that get this right will not be the ones trying every new AI feature as soon as it appears. They will be the ones building useful, grounded systems around their own work. They will use AI to reduce admin, connect tools, improve reporting and speed up delivery, but they will keep ownership of the process. They will treat AI as a powerful layer in the business, not the business itself.

That is the balance I think more small businesses need to find. Use the tools, because the value is real. Use agents too, where they genuinely remove work or improve delivery. But understand the dependency and the cost model before they become invisible. The first hit might be cheap. The habit is where the cost starts. The agent running your business in the background is where the bill can become structural.