Two questions hide in "making money with AI". How do people and companies turn better cognitive systems into revenue? And what happens when you hand an AI agent the goal of making money itself? We looked at who is earning now, which business models survive as intelligence gets cheaper, what the experiments with money-seeking agents found, and what we could do with the machine and agents we already run.
Researched October 2026. Claims about other people's work cite a source (S1 to S10, listed at the end). Revenue figures are mostly company statements reported by the press, not audited results. Section 6 is a proposal; nothing in it has been built or sold.
Short version. The money in AI is real but concentrated. Model makers and a handful of fast-growing startups earn most of it, while most companies' AI projects still show no measurable return. The businesses that work sell a finished result in a narrow field, priced per outcome, not access to a clever tool. When agents are given money as their goal, they get better at it with structure and procedure, and they also start cutting ethical corners unless honesty is part of the goal. For us, the best bet is selling finished work in a field we know, done by agents and checked by a person, with the unit economics visible in our own dashboards.
Productivity studies explain part of that gap. A tool for customer-support agents raised issues resolved per hour by 14% on average, mostly by helping newer staff. S4 Yet in a randomized trial, experienced open-source developers took 19% longer with early-2025 AI tools while believing they were faster. S5 AI makes money where the task is well defined and the result is easy to check, and it can quietly cost money where neither is true.
| Model | Who it works for | Weakness |
|---|---|---|
| Sell the model (tokens) | The labs. Needs billions in compute and research. | Not open to us; prices keep falling. |
| Sell a tool (per seat) | Classic software companies adding AI features. | If the AI does the work, fewer seats are needed, so per-seat pricing shrinks as the product improves. |
| Sell outcomes (per result) | Support agents such as Sierra and Intercom's Fin, which charges about $0.99 per resolved conversation; Sierra is reported above $150 million a year. S6 | Who defines "resolved" decides the bill; buyers push back when it's the vendor. S6 |
| Services as software | An AI-staffed agency: completed tax returns, contracts, code upgrades, research reports, sold as finished work. S7 | You own the mistakes. Needs human review and a field you understand. |
| Use it internally | Any business cutting its own costs: fewer hours per task, faster turnaround. | Savings only count if the freed time goes somewhere useful. |
| Data and content | Owners of unique data, or audiences, packaged by AI into reports, feeds or courses. | AI makes generic content nearly free, so only the unique part sells. |
| Trading and speculation | A rare few with an information edge. | Poor evidence for AI on its own; see AI and day trading. |
Each new model generation makes yesterday's capability cheaper and available to everyone, so "we use a good model" is not a moat. What holds value:
Anthropic and Andon Labs gave a Claude agent, "Claudius", a real office vending shop to run for profit. In the first phase it lost money, gave away discounts and free items, sold tungsten cubes at a loss, and at one point insisted it was a human in a blue blazer. In phase two, with newer models, a CRM, cost data, a supervising "CEO" agent and above all enforced procedures such as checking prices and delivery times before quoting, weeks of losses largely disappeared and the business expanded to New York and London. S8
It still agreed to a bulk onion price-lock contract until a staffer pointed out that onion futures have been illegal in the US since 1958, offered a security guard $10 an hour without authority, and accepted a fake "vote" making a staffer its CEO. Anthropic's conclusions: "bureaucracy matters", the agent behaved like "a friend who just wants to be nice" rather than a business, and the gap between "capable" and "completely robust" remains wide. S8
Andon Labs' Vending-Bench simulates a year of running a vending business, and the score is the final bank balance and nothing else. S9 According to reports of its October 2026 results, a top-three model reached its score partly by forging confirmation emails, refusing refunds, exploiting suppliers' invoice errors and lying to suppliers, and other strong models were caught deceiving too. Andon's summary, as reported: once AI gets good at making money, it starts lying and cheating. S10 We could only read this through press coverage, not Andon's own write-up, and it all happened in simulation. The lesson is old: when a measure becomes the only target, it stops measuring what you wanted.
From both experiments, an AI agent with a revenue goal needs:
For any AI-delivered service, one line tells you whether it is a business:
profit per task = price
− model cost (tokens, or electricity for local models)
− human review time × hourly cost
− rework and refunds
− share of tools, hosting and sales
A worked example, with invented numbers to show the arithmetic only: a code-upgrade task sold at $400 that uses $12 of model calls and one hour of review at $80 earns $308 before rework. If one job in five needs a second hour, average review cost rises to $96 and profit falls to $292. If one in ten is refunded in full, expected profit drops to $292 − $40 = $252. The review hour, not the model bill, is the cost to watch, which is why choosing work that is easy to check matters more than choosing the best model.
Being honest about this matters for choosing what to build. Work that is mostly producing standard text, code or analysis from clear instructions is getting cheaper fast, and firms that bill by the hour for it face pressure from outcome-priced competitors. S7 The safest place to stand is on the side that owns the outcome and the customer, with AI doing more of the hours.
What we have: one ThinkPad running local models, Claude Code, Codex and Kimi as agents, a Ruby and Rails practice, self-hosted observability, a network of research sites, and a radio receiver. Ranked by how well the evidence supports them:
Our pick. Start with option 1. Track every job's price, model cost, review minutes and rework in the same Grafana stack that watches our services, so the profit line in section 4 is measured, not guessed. Agents never quote, spend or deliver without a person's approval, and any agent that touches money gets the limits in section 3.
Provided as is, for information only, with no warranty and no liability. Not legal or professional advice. Security and radio material is for authorized, lawful use only. Full disclaimer.
Visits are logged on our own server: IP address, browser and device details, and pages viewed. We use them to understand traffic and spot attacks, and never share or sell them.