- 01Recovered cost is the yardstick: the hours per week a piece of work absorbs, multiplied by the loaded salary of the people absorbing it, gives the yearly cost of doing it by hand. An internal agent is priced against that figure.
- 02Per-seat SaaS scales with headcount rather than with work. Time-and-materials scales with uncertainty. Build-it-yourself platforms assume an AI operator on your payroll. None of the three is priced against the work.
- 03The real comparator is not another vendor. It is doing nothing, or hiring one more operations person to absorb the same hours.
The first question an owner asks about an internal agent is what it costs. It is the wrong first question, because the number has nothing to compare against. The right first question is what the work costs today. Once that figure exists, every pricing model in the market can be held up against it, and most of them fail the comparison.
The recovered-cost method. Take one piece of internal work: the weekly leadership report, the invoice-to-PO check, the stock reconciliation across store and warehouse. Count the hours it absorbs each week, across everyone who touches it. Multiply by the loaded hourly cost of those people, meaning salary plus the employer's overheads, not the figure on the payslip. Multiply by the working weeks in a year. That is the yearly cost of the work as it is done now. It is a real number, it is already being paid, and it sits in the accounts under salaries where nobody sees it as a line.
Run it once and the shape is usually surprising. A report that takes an operations manager six hours every Monday is not a six-hour problem. It is three hundred hours a year of one of the more expensive people in the building, plus the hours of the three people who send them the inputs. The agent that assembles that report overnight is priced against three hundred hours, not against a software category.
This is how Hubzoid prices. The engagement is fixed-price, and the price is set against the recovered cost of the work in scope, so that payback lands inside the first year on the arithmetic above, before any second-order effect is counted. The site does not publish the figure, because it is different for every company, and it is different for a reason: the work is different. What is constant is the method, and the method is written down before the build starts. Now hold the other models against the same yardstick.
01. Per-seat SaaS. The price scales with the number of people who log in, not with the work that gets done. Internal agents are used most heavily by a small number of people. The owner reads the briefing. The accounts desk runs the invoice check. Charging by seat punishes exactly the rollout an owner wants, where the agent reaches the whole team, and rewards the rollout nobody wants, where it stays with the three people who set it up. The enterprise platforms in this category also carry seat minimums that a 200-person company cannot meet honestly.
02. Time-and-materials. The dev-shop model. The estimate is an estimate, the hours are real, and the risk of the difference sits with the buyer. The incentive runs the wrong way: the longer the build, the larger the bill. It also means the number to compare against recovered cost is unknown until the end, which makes the comparison impossible at the moment it matters, which is before signing.
03. Build-it-yourself platforms. Priced low, sold monthly, and honest about their assumption only in the fine print: someone on your side will design the agent, connect the systems, write the checks, watch the output, and fix it when the ERP export changes shape. That someone is an AI operator. Most owner-led companies of a hundred to a thousand people do not have one and are not going to hire one. The subscription is cheap. The unfilled role is not.
Build versus buy. There is a real case for building internally. It holds when the company already has an engineer who will own the agent for the next two years, has the access to production systems that ownership requires, and has leadership patience for the first three months of it not being right. When those three hold, build. When any of them do not, the data is clear: firms using an implementation partner scale AI at roughly twice the success rate of pure internal builds. The gap is not talent. It is that a partner has already made the mistakes on someone else's clock.
What a fixed-price engagement should include, if it is priced against the work, is the following. Anything less and the price is against a demo.
01. A discovery map. How the company actually decides: the tasks that repeat, the approvals that live in one person's head, the thresholds nobody wrote down. Confirmed by the people who hold them, and the document every later step reads.
02. A catalog. Candidate agents, more than will be built, each one tied to a decision, the systems it reads, and the hours it recovers. Selection means something only when real candidates lose.
03. The first agent, and the hub it runs on. One agent on real work, on the surface the team already uses, with access gated by group and decisions recorded. Not a pilot in a sandbox. The production shape from day one, at a scope small enough to prove.
04. Handover. A runbook, a repo of versioned decision logic, and a team that knows how to change a threshold without calling anyone. The engagement ends with a handover, not a subscription.
05. Ownership. Deployed in your own cloud, on an open-source core under MIT, with the work product yours under the contract. If the vendor disappears, the agent keeps running.
The comparator that matters is not another vendor. It is the two things an owner will actually do instead. Do nothing, and keep paying the recovered-cost figure every year, quietly, under salaries. Or hire one more operations person to absorb the same hours, at a loaded cost that is itself the yardstick, with a notice period, a ramp, and the same knowledge locked in one more head. An internal agent priced against the work has to beat both of those on the arithmetic. If it does not, do not buy it, from Hubzoid or from anyone.
Run the number before the demo. Pick the one piece of work that most annoys the most expensive person in the building, count its hours honestly, and multiply. Bring that figure to every conversation about agents, and ask each vendor to price against it. The ones who can will show their working. The ones who cannot will change the subject to features.