Sell AI agents without losing a billion: a unit-economics checklist

Autonomy is the easiest claim in enterprise software. It is also the fastest way to sell a project that never becomes a recurring outcome.

A vendor can ship many AI employees and still lose money if the operating model is built on custom work, weak retention, and slow cash. The buyer expects a worker. The vendor delivers a project. The project ships, the agent idles, the renewal dies, and the invoice arrives late.

Jinzhiwei is seeking financing on the Hong Kong Stock Exchange and reported cumulative ten-year losses of 1.259 billion yuan as of end-June 2026.

By June 2026, the company had deployed more than 2 million AI digital employees and served more than 1,600 customers.

Scale did not save the model. Revenue grew from 217 million yuan to 256 million yuan over three years, with an 8.7% CAGR versus a stated industry growth rate of 37.1%.

The problem was not demand. It was the shape of the revenue. More than 70% of revenue comes from project-based business, a model with a 40.3% gross margin, while the subscription model has a 96.3% gross margin and subscription revenue is only 16.8% and declining.

The customer base also weakened. Customer count fell from 781 at end-2025 to 439 in mid-2026, retention fell from 74% to 60%, and payment collection extended from 137 to 446 days.

Why autonomy claims break the P&L

When autonomy is a sales claim, the product becomes a promise. The enterprise signs for a digital worker, but the vendor still has to build, integrate, monitor, and defend the outcome. Every edge case becomes a service ticket. Every failed task becomes a support call. Every slow result becomes a renewal risk.

Project revenue feels safe because it is booked upfront. It is not. It consumes engineering, sales, and delivery capacity. It creates custom dependencies. It makes the next sale harder, not easier. The customer does not renew a project. The customer renews an outcome.

Agent vendor economics fail when the company sells autonomy but operates like a services shop. The market rewards measurable work, not impressive deployments. If the agent does not produce a result the buyer can track, the buyer will not pay for the next period.

The agent-outcome unit-economics checklist

Founders, product leads, and sales operators should use this checklist before selling another autonomous worker. It is not a marketing test. It is a cash test.

1. Define one measurable outcome. Pick the task the agent completes, the metric it improves, and the owner who signs off. Put the outcome in the contract. If the outcome is “more efficient,” the deal is a project. If the outcome is improved efficiency or lower cost, the deal can be recurring.
2. Price recurring outcomes, not projects. Charge for the work the agent does, not the time it takes to deploy it. Put the price on the result. Put the renewal on the result. If the customer stops using the agent, the revenue should stop too, but only after usage proves the outcome is not being delivered.
3. Cap support cost per active agent. Set the maximum support cost per active agent and review it when usage changes. Track the support hours, escalations, and manual overrides needed to keep one agent running. If support cost rises with every new edge case, the agent is not autonomous. It is a labor line with a robot label.
4. Require usage and retention thresholds. Set the minimum usage level and the retention trigger before the renewal date. Do not treat deployment as success. If usage drops, trigger a review, not a renewal.
5. Track cash conversion days. Set a target for cash conversion days and review it with product, not just finance. Measure the time from delivered outcome to paid invoice. Slow collection is a product problem, not just a finance problem. If the buyer pays late, the outcome is not valuable enough to protect.
6. Keep project revenue from becoming the majority and keep subscription gross margin in the top band. Set the project-revenue ceiling and the subscription-margin floor before the next sale. Project work can fund early learning. It cannot become the business. Subscription revenue should carry the margin. If the high-margin line is small and shrinking, the company is selling autonomy and buying services losses.

What to do next

Stop selling “AI employees” as a category. Sell the outcome the employee produces. Make the outcome visible in the dashboard. Make the price tied to the outcome. Make the support cost visible per active agent. Make the cash cycle visible to the founder, not just the CFO.

The market will keep rewarding vendors that can prove the agent works. It will punish vendors that can only prove the agent was deployed. The difference is not a better demo. It is a unit-economics model that survives contact with real usage, real churn, and real invoices.