Most arguments about AI vendors are about talent—who has the better engineers, the slicker demo, the more impressive logos. That’s the wrong fight. The thing that quietly decides whether your AI work succeeds isn’t who you hire. It’s the shape of the contract you hire them under.
The contract sets the incentives. The incentives set the behavior. And over enough months, the behavior wins—no matter how good the people are or how good their intentions were on the kickoff call.
So before you compare engineers, compare billing models. Here’s how the traditional ones actually work underneath the proposal language, how credit-based delivery works instead, and why we moved our whole model to the second one.
How traditional consulting bills—and what each model optimizes for
There are really only three ways consultancies sell technical work. Each one sounds reasonable. Each one optimizes for something other than your result.
Hourly / time-and-materials
You pay for hours logged. This is the most honest model to invoice and the most dishonest one to live under, because an hour measures presence, not progress.
A senior engineer who has solved your exact problem before might finish in an afternoon. A less experienced one—or a more financially motivated one—spends two weeks “exploring approaches.” Under hourly billing, the second one earns more. You are paying a premium for inexperience and getting a discount on nothing. And the better the engineering gets—the cleaner the data model, the integration that removes manual steps—the smaller the future billable surface becomes. Durable work is, under hourly terms, the engineer sabotaging their own revenue. We wrote about this failure mode in detail in Outcomes, Not Hours.
Fixed-bid SOW
You pay a fixed price for a defined scope. Better in theory—the risk of “this took longer than expected” moves to the vendor. In practice, two things happen.
First, the vendor pads. They don’t know exactly how hard your environment is, so they price in a cushion, and you pay for their uncertainty up front. Second, the moment reality diverges from the document—and with AI on real, messy enterprise data it always diverges—you’re in a change order. The relationship turns adversarial at exactly the moment it needs to be collaborative, because every new discovery is now a negotiation. Most “the project went sideways” stories are really “we kept renegotiating the SOW” stories wearing a costume.
Open-ended retainer
You pay a recurring fee for availability. This is closest to ongoing delivery, but with nothing underneath it. A retainer with no defined unit of work pays for the vendor to be around, not to ship. There’s no shared definition of what a month should produce, so good months and coasting months cost the same, and neither of you can point at the work and agree it’s done.
Notice the pattern. In all three, the thing you pay for and the thing you actually want are different things. You want a system that exists and works. You’re paying for hours, or for a guess, or for availability.
What credit-based delivery actually is
Credit-based delivery starts from a different premise: price the work, not the time.
A credit is a standardized unit used to size the scope, complexity, integration risk, and acceptance requirements of an agreed deliverable. It is deliberately decoupled from hours. A credit is not a billed hour, and it is not an “AI hour.” It’s a measure of accepted work—how much delivered, working output a given outcome represents.
Here’s the part that matters in practice. Before any engineering begins, you see four things in writing for each request: the scope, the dependencies, a credit estimate, and the acceptance criteria. Then you decide whether and when that work enters the active delivery lane. Nothing starts on a vague promise, and nothing is “done” until it passes a test you could run yourself.
That changes the conversation from how long will this take and what’s your rate to what exactly are we building, how will we both know it works, and how big a bite of this month’s capacity is it. Those are the questions that actually predict whether you’ll be happy in ninety days.
This is the model behind Ryshe Forge: a named senior delivery team working a predictable monthly program with reserved capacity, sizing each deliverable in credits and shipping against acceptance criteria. We treat delivery capacity less like billable bodies and more like a reserved utility.
Side by side
| Traditional consulting | Credit-based delivery (Ryshe Forge) | |
|---|---|---|
| What you pay for | Hours, a fixed guess, or availability | Sized, accepted deliverables against reserved monthly capacity |
| Vendor’s incentive | More hours / protect the SOW | Ship accepted outcomes efficiently |
| Speed | Faster work can mean less revenue | Faster is good for everyone—the result lands sooner |
| Scope changes | Change-order negotiation | Re-estimated and re-prioritized, not re-negotiated from scratch |
| Visibility | A timesheet or a status deck | Written scope, dependencies, credit estimate, acceptance criteria—before work starts |
| Who does the work | Often junior delivery behind a senior pitch | The same named senior people, month to month |
| When it ends | At the SOW’s edge—right where the roadmap begins | It doesn’t; the roadmap keeps moving, and you control the lane |
| Work in progress | As many open threads as they can bill | Controlled active workstreams and clear WIP limits |
Why the credit model is better for the buyer—not just cleaner for us
It would be easy to dismiss this as vendor-speak for “pay us a recurring fee.” Done honestly, it’s the opposite—it puts more risk and accountability on us, not less.
You stop paying for the learning curve. If we underestimate how hard a deliverable is, that’s absorbed in the sizing we agreed to—it doesn’t reappear as a surprise on next month’s invoice. The party best equipped to estimate the difficulty carries the difficulty. That’s where it belongs.
Speed finally works in your favor. When the unit is an accepted outcome rather than an hour, there is no reason for anyone to pad. If our team finds a faster way to hit the acceptance criteria, you get the result sooner and we’re both fine with it. The cheaper-for-you path and the better-for-us path become the same path.
Focus replaces sprawl. We limit active workstreams on purpose. A controlled lane of work that ships beats a wide backlog of half-finished AI experiments that all look 80% done forever. You’d rather have three things in production than ten things in “almost.”
The renegotiation tax disappears. New discoveries don’t trigger a contract fight. They get re-estimated and you decide where they sit in the roadmap. Discovery becomes a normal part of delivery instead of a commercial event.
What it does not mean
Honesty cuts both ways, and a model that promised everything would be its own red flag.
Credit-based delivery is not endless, open-ended work, and it is not “everything is covered no matter what.” Reserved monthly capacity is a real, finite amount of delivery work—that’s what makes it predictable. Credits don’t map to a fixed number of human or AI hours. Defect correction within already-approved acceptance criteria doesn’t consume new credits, but genuinely new requirements or expanded scope are re-estimated before they enter the lane. Additional capacity starts only with written approval. The discipline is the point: clear terms and controlled scope are what let us actually commit to outcomes instead of hedging on hours.
And it isn’t always the right tool. If you need a single, truly fixed deliverable with a hard finish line and no ongoing roadmap, a scoped project is the honest answer—that’s what Ryshe Labs is for. If what you really want is bodies you manage yourself, that’s staff augmentation, and you should buy it as such. We help sort out which model a given problem wants before anyone signs anything.
The model is the message
We don’t publish rates, for reasons worth their own post—but the commercial structure isn’t a secret, and it shouldn’t be. The way you buy AI work shapes what you get. Buy hours and you’ll get hours. Buy a guess and you’ll get a padded guess and a change order. Buy accepted, sized, production outcomes against reserved capacity, and—if the people are good—that’s what tends to show up.
If the alternative on your table is building it in-house, it’s worth seeing the real number first — our build vs. buy calculator totals the hiring, ramp, tooling, and maintenance most teams forget to count, with no pricing wall.
If you want to pressure-test this against how you’re buying AI work today, that’s a good thirty-minute conversation. Bring the real situation, not the cleaned-up one. You can book a 30-minute fit call, and we’ll tell you honestly whether credit-based delivery fits—or whether a project, or no engagement at all, is the better call.
