The one question that separates them
Two firms bid your build. One quotes $185 an hour for three engineers. The other quotes $205 an hour for two. Both send capable résumés, both say the word partnership more than once, and both promise senior people. Procurement reads this as a 30 percent price difference and picks the cheaper column. What actually differs is who owns the design decisions. In one arrangement your staff decides what gets built and the vendor executes it. In the other the vendor decides what gets built and answers for whether it works. Everything else about the two engagements follows from that single fact, including the part where the cheaper one costs more.
The confusion is not anyone's fault. Both models sell hours of engineering and use overlapping vocabulary. Staffing firms learned that buyers want a partner, so their proposals say partner. The words converged. The businesses did not.
The test is who holds the pen on architecture. Ask both firms: if your engineer looks at our data model in week three and concludes it will not support the feature we asked for, what happens? A staffing vendor's honest answer is that the engineer raises it with your technical lead and does what your lead decides. That is correct behavior for that model. A partner's honest answer is that they come back with a recommendation, a cost of changing course, and a position on which option they will stand behind. One is renting you hands. The other is renting you a decision.
You are probably here because
- Two proposals came back with similar numbers and you cannot tell what the difference buys
- The last engagement delivered exactly what was asked for and it was not what you needed
- Your best architect is spending half their week writing tickets for contractors
- Something government-adjacent got added to the requirements and nobody on the bid seems to know what 800-171 costs
All four are symptoms of buying the wrong model rather than the wrong firm. The sections below give the questions that reveal which model a proposal actually is.
What a staffing vendor is genuinely good at
Staff augmentation is a real product with a real customer, and it is badly served by being treated as the discount version of something else. A staffing firm's core competence is a recruiting pipeline. They can find a Kubernetes engineer with a public-trust background in eleven days, which is a capability most companies cannot build and should not try to. Onshore rates typically land between $70 and $150 an hour depending on skill and clearance, meaningfully below what a firm carrying design responsibility can charge, because the vendor is not carrying that responsibility.
This model is correct when your bottleneck is capacity and your design is already settled. You know the architecture. You have a technical lead who can review pull requests, answer design questions within a day, and tell a contractor that their approach is wrong. You need four more sets of hands for eight months. Buy seats. Paying an engineering firm's margin for work your own architect has already designed is spending money on judgment you already own.
The model breaks in one specific way, worth naming because it does not look like breakage while it is happening. Staff augmentation transfers the design load onto your organization without anyone writing that down. Your senior architect becomes a full-time ticket writer and reviewer for six contractors. Velocity charts look excellent, and your most expensive internal person is doing the least valuable version of their job.
What an engineering partner actually sells
An engineering partner sells a judgment call and stands behind it. The deliverable is not hours. It is a working system with acceptance criteria attached, and the firm eats the cost of getting there when its own estimate was wrong. That is the entire economic difference, and it is why the rate is higher. Somebody is holding risk that a staffing arrangement leaves with you.
Concretely, this shows up in four places. The firm writes the acceptance criteria and signs up to them. It decides the architecture and defends it. It absorbs rework caused by its own design mistakes rather than billing change orders. And it is on the hook for the system working in your environment, not for a component working on a laptop.
The model is wrong for you in cases worth stating plainly. It is expensive for commodity work. Paying a senior firm's rate to build a standard admin interface is a waste of your money and their attention. It does not scale on demand, so asking a twelve-person firm for fifteen engineers by November produces either a no or a subcontracting arrangement you did not ask for. And if you already have a competent architecture and a technical lead with capacity, you are paying twice for the same function.
The five questions that reveal the model
Rate cards do not distinguish these firms. These questions do, and each one has a mechanism behind it rather than a preference.
Who writes the acceptance criteria? If the answer is that you write them and the vendor executes against them, you are buying staffing regardless of what the cover letter says. A partner writes criteria because a partner is committing to hit them. Watch for the criteria themselves: "improves analyst throughput" is a slogan, "reduces median case triage from 14 minutes to under 6 on a frozen 200-case regression set, verified by your technical lead" is an engineering commitment.
What happens when the estimate is wrong? Every estimate is wrong. The question is who pays. Under a time-and-materials structure, the answer is always you, which is honest and appropriate when discovery genuinely dominates. Under a firm-fixed-price structure, the answer is the vendor, which is why fixed-price bidders push hard on scope definition before signing. A firm that will not discuss this question is one that has not thought about it.
Name the people, and tell me what else they are doing. A staffing vendor names candidates and their availability, which is exactly right. A partner names two or three people, their other commitments, and who is accountable when one of them is out. A firm that answers with role titles rather than names is telling you the delivery team does not exist yet, and you are buying a promise to staff.
Who owns the design decision your team disagrees with? Ask it as a scenario. Your VP of Product wants a feature the vendor's engineer thinks will not scale past 50,000 records. Walk me through the next two weeks. Staffing vendors build it. Partners write down the objection, price both paths, and make you decide with the cost in front of you.
What do we get if we stop after 90 days? The answer separates firms that produce systems from firms that produce activity. A defensible answer includes a tagged repository, a runbook someone else can follow, the evaluation notebook behind whatever performance was claimed, and a data dictionary. If the answer is "the code so far," the knowledge lives in the contractors, and it leaves when they do.
Where each model fits the work — our rating
Our judgment of the models rather than a measurement of any firm. The ordering is the useful part; a strong staffing team beats a weak partner in every row.
The contract shape tells you more than the proposal
Read the contract structure before the narrative. Time-and-materials with a not-to-exceed ceiling is a staffing instrument. It prices hours, it makes you the party who decides when to stop, and it carries no commitment about the result. Firm-fixed-price against defined acceptance criteria is a partner instrument. It prices a result and shifts the estimating risk to the seller.
In federal work these have specific homes. FAR 16.601 governs time-and-materials and requires a determination that no other contract type is suitable, precisely because the government understands that T&M puts the schedule risk on the buyer. FAR 16.202 covers firm-fixed-price. Commercial buyers are not bound by either, but the logic transfers exactly, and the cleanest commercial arrangement is a short paid discovery under T&M that converts into fixed-price milestones once the shape is known.
A useful diagnostic: ask a firm that quoted T&M whether they would convert to fixed-price after a four-week discovery. A partner will say yes with conditions and name the conditions. A staffing vendor will say no, honestly, because they have no mechanism for absorbing an estimating miss and no basis for pricing one.
| Signal in the proposal | Staffing vendor | Engineering partner |
|---|---|---|
| Pricing unit | Hourly rate per labor category, NTE ceiling | Fixed price per milestone with named deliverables |
| Acceptance criteria | Written by you, or absent entirely | Written by the vendor, with metric, instrument and approver |
| Who runs standups | Your project manager | Their technical lead, reporting out to you |
| Rework from a design mistake | Billed as additional hours | Absorbed inside the fixed price |
| Compliance work | Out of scope unless a seat is added for it | Priced into the build as a control-mapped deliverable |
| Team continuity | Replaceable by an equivalent résumé | Named people, with a named backup |
| What ends the engagement | The ceiling, or your decision to stop | Acceptance of the final milestone |
Where this decision gets expensive: anything touching government
If your build touches federal data, a federal customer, or a prime who has one, the model choice stops being a preference and starts being a cost driver. Compliance work does not decompose into tickets. NIST SP 800-171 has 110 controls across 14 families, and roughly a third of them are architectural: boundary definition, flow enforcement, session handling, audit record content. Deciding where the CUI boundary sits is a design decision that changes the system. It cannot be assigned to a contractor who was hired to close tickets, because there is no ticket that says "decide what our enclave is."
The same holds for FedRAMP inheritance. A team that does not know which controls a cloud service provider already satisfies will write compensating controls for things the platform handles, or assume inheritance that does not exist. NIST SP 800-53 Rev 5 carries over a thousand controls; deciding which apply at which baseline has a several-hundred-thousand-dollar spread between doing it well and doing it by default.
For AI systems the pattern repeats with newer frameworks. The NIST AI Risk Management Framework's Map function requires that intended use, data provenance, and out-of-scope uses be documented before the system is fielded, which means somebody has to decide them. The OWASP LLM Top 10 and MITRE ATLAS describe attack classes that a design either anticipates or does not. Financial institutions carry SR 11-7 model risk obligations that require independent validation, which is structurally impossible when the only people who understand the model are the ones who built it under your direction. ISO/IEC 42001 asks for an AI management system, not a set of features.
How much of the work is a design decision, not a ticket
Our estimate of how much of each task is judgment that cannot be assigned as a ticket. Illustrative, not a measured statistic.
The failure mode is specific and common. Six months of good engineering, then a security review, then a finding that the data flow crosses a boundary it should not, then three months of rework nobody budgeted. The engineers did nothing wrong. They built what the tickets said. The tickets did not say where the boundary was because no one on either side owned that question.
The math on total cost
Build the comparison on total quarterly cost, including your own people, over the realistic life of the work. Four numbers matter and only one appears on the quote.
The billed amount at realistic hours. Realistic means the estimate plus the change orders you can already predict. If a vendor's estimate has no contingency, add 15 to 25 percent yourself before comparing anything.
Your internal load. A staffing arrangement consumes senior internal time at a rate people consistently underestimate. Six contractors typically consume 40 to 60 percent of one senior architect. At a fully loaded $220,000, that is $88,000 to $132,000 a year of your best engineer, spent on specification and review. It belongs in the comparison because it is real money spent on the arrangement.
Rework you will pay for twice. Work built correctly against a specification you wrote incorrectly is not the vendor's fault and is still your money. Models that put a senior person between your intent and the code reduce this. Models that faithfully execute your ticket do not.
Exit cost. What it costs to operate this system next year with different people. Undocumented, single-author, untested code has an exit cost that can exceed the original build, appears on no quote, and lands entirely on you.
The hybrid that usually works
Most companies with real budget do not choose one. They buy design responsibility from a small senior firm and capacity from a staffing vendor, and they define the seam explicitly. The partner owns architecture, acceptance criteria, compliance boundary, and code review. The staffing team builds inside that frame.
It works when three things are written down. The partner has explicit authority to reject work below standard. The staffing contract names the partner's technical lead as approving authority. And the partner's scope includes review time, priced as a line item rather than assumed as goodwill. Skip any of the three and the seam becomes where accountability disappears.
Ratio matters. One senior design owner can hold a frame for four to six builders in a shared codebase. Past that the reviews become rubber stamps and you are paying for a signature rather than for judgment.
How to run the selection
Give both bidders the same real problem and a small budget. Two to four weeks, a fixed fee somewhere between $15,000 and $40,000 depending on scale, and a written deliverable you own regardless of who wins the build. Ask for an architecture recommendation with the tradeoffs named, an acceptance criteria set, a risk list with the top three technical unknowns, and a range with the assumptions that would move it.
What comes back is the actual product. A staffing vendor returns a good staffing plan, because that is their competence and they are being honest about it. A partner returns a position: here is what we would build, here is what we would not, here is where our estimate is soft and why. Notice which one told you something you did not already know. That is what you are considering paying for.
The worst outcome of this process is a small invoice and a document you keep. The alternative is a seven-figure build aimed at a target nobody validated.
Bottom line
Staffing vendors and engineering partners are different businesses that quote on the same sheet. A staffing vendor sells recruiting and availability, prices hours, and leaves design ownership with you. That is correct and cost-effective when your architecture is settled and your technical lead has capacity. An engineering partner sells a design position and absorbs the cost of being wrong about it, which is what you need when the right answer is not yet known or when compliance obligations make the boundary decisions the hard part. Decide which problem you have before comparing rates. The rate difference is small and the consequence of the mismatch is not.
Frequently asked questions
Per hour, usually yes, often 25 to 50 percent higher. Per outcome it depends entirely on whether your organization already owns the design capacity. If your architect has time to specify and review, staffing is cheaper on both axes. If that time comes out of work only they can do, the hourly saving is an accounting artifact.
Look at three things: whether the pricing unit is an hour or a milestone, whether acceptance criteria appear and who authored them, and whether individuals are named with their other commitments. Proposals that price hours, omit criteria, and describe roles instead of people are staffing proposals regardless of the language.
They can supply engineers who know the control families. What they do not supply is the person who decides where your CUI boundary sits and which controls you inherit from a platform. Those are architecture decisions with large cost consequences, and if nobody on your side owns them the answer arrives during a security review instead.
A paid discovery of two to four weeks against a real problem, typically $15,000 to $40,000, with a written deliverable you keep either way. Ask for an architecture recommendation, acceptance criteria, a ranked risk list, and a range with its assumptions. The quality of that document predicts the engagement better than any reference call.
Yes, and it is often the best structure. The partner owns architecture, acceptance criteria and code review; the staffing team builds inside that frame. It requires three things in writing: the partner's authority to reject work, the staffing contract naming the partner's lead as approver, and review time priced as a line item.