The comparison everyone makes, and the one that decides it
A leader with an unfilled machine-learning role usually compares two numbers: a salary divided by 2,080, and an outside firm's hourly rate. The rate is higher. It is supposed to be higher, and a comparison built on that surprise answers a question nobody asked. Two numbers are missing. The first is the fully loaded cost of the employee, which is not the salary. The second is the cost of the work not happening while the seat stays open, which is running right now and appears on no one's budget line. Put all four numbers on one page and the decision usually stops being close.
The deeper problem is that the two paths do not buy the same good. A hire buys a person who accumulates context nobody wrote down, who is still there in three years, and whose cost continues whether or not there is work in front of them. A contract buys a defined result on a schedule, with the estimate risk sitting somewhere other than your budget, and it stops when you stop it. Pricing those against each other per hour is like comparing a mortgage payment to a hotel bill.
Name what you are short of before you price anything
Three shortages look identical on an org chart and have different fixes. Getting this wrong is how a company hires three people and still ships nothing.
Hands. The work is defined, the approach is settled, and there are not enough people. This is the only shortage headcount cleanly solves, and also the one most easily contracted, because a defined scope transfers to an outside team without much loss.
Judgment. Nobody in the building can say whether the approach is right. Retrieval or fine-tuning, whether the labeled data is sufficient, whether the evaluation will survive contact with a regulator. Adding hands to a judgment shortage multiplies the wrong direction faster. What is needed is a small number of senior hours, early, on the design.
A decision. The work is blocked on a choice that has an owner but no answer: build or buy, one vendor or another, whether the data supports the claim at all. Engineering capacity does not move this. A bounded assessment with a written recommendation and an accountable author does.
Judgment and decision shortages get mislabeled as hands shortages because a requisition is the tool the organization knows how to reach for. If you cannot say in one sentence what the new person would do in their first month, the shortage is not hands.
Why the role is hard to fill in the first place
Open a typical AI job description and count the jobs inside it. Four or five is normal. Each is a full career track with its own labor market, and asking for all of them in one person shrinks the candidate pool enough to explain the vacancy by itself.
- Applied modeling — choosing and adapting model architectures, running experiments, knowing which published result will survive contact with your data.
- Data engineering — access, lineage, joins nobody documented, and the unglamorous work that usually consumes most of the schedule.
- Platform and deployment — containers, inference cost, monitoring, rollback, and the accreditation path if the system lands in a regulated environment.
- Evaluation — building the measurement that tells you whether the thing works, and defending it to someone who did not build it.
- Domain translation — turning what an underwriter, a clinician, or a program manager does into a specification an engineer can implement.
- Documentation and governance — model cards, data handling, audit trail, and the paperwork that decides whether a working system may run.
Split the description and the market reappears. A data engineer who can also stand up inference is a common and hirable profile. Someone who is simultaneously a research-grade modeler, a platform engineer, and a domain expert in your field has several offers already, and that search takes as long as it takes.
What the hire costs, in numbers you can check
Start with the part that is arithmetic rather than argument. The Bureau of Labor Statistics measures employer compensation costs directly. In its Employer Costs for Employee Compensation series for March 2026, employer costs for civilian workers averaged $49.32 per hour worked, of which $15.60 was benefits. Benefits were 31.6 percent of total compensation and wages and salaries were 68.4 percent.
The dollar level in that series is an economy-wide average and has nothing to say about what an AI engineer earns. The ratio is the transferable part. If you are budgeting a role at a stated salary and have not added roughly a third again for the employer's side of benefits, you are budgeting about three quarters of the real number.
Inside that benefits figure, the mandatory pieces are exact. For 2026 the employer pays Social Security tax of 6.2 percent on wages up to the taxable maximum of $184,500 and Medicare tax of 1.45 percent on all covered wages with no ceiling, and the employee owes a matching 6.2 and 1.45 percent. An Additional Medicare Tax of 0.9 percent applies to wages above $200,000, which the employer withholds and does not match. Federal and state unemployment taxes sit on top, varying by state and by experience rating.
Then come the lines no honest model reduces to a single multiplier: health and retirement contributions, paid leave, workers' compensation, hardware, software seats, recruiter fees, and the hours senior people spend interviewing and onboarding instead of building. All real, and none of them in a salary-divided-by-2,080 comparison.
The open seat has a price, and almost nobody computes it
Here is the number that changes the decision. The work waiting on the hire is worth something per week, and that value is being forgone every week. It is measurable in at least one of three ways, and you only need one.
The line the work moves. If the model is meant to cut a manual review queue, price the queue: people, hours, error rate. If it is meant to win business, price the pipeline it unblocks. If it supports a milestone, price the milestone.
The people already idled behind it. A stalled data platform usually has several analysts partially blocked on it. Their loaded cost multiplied by the fraction of their time spent waiting is a defensible weekly number, and usually a larger one than expected.
The date that does not move. Some work has a fixed deadline: a fiscal year end, a regulatory submission, a customer commitment, a proposal due date. When the deadline is inside the hiring cycle, the hire cannot solve the problem no matter how good the candidate is. That is not a cost comparison at all. It is a feasibility test, and it is the single most common reason the answer comes out contract.
Side by side, on the dimensions that actually differ
Price per hour is the least informative column, so it is not here. What separates the two paths is when work starts, who absorbs a bad estimate, what happens when the need ends, and where the knowledge sits afterward.
| Dimension | Filling the role | Contracting the work |
|---|---|---|
| What you are buying | Capacity and accumulated context, indefinitely | A defined result, on a schedule, with acceptance criteria |
| When useful work starts | After search, offer, notice period, onboarding, and access provisioning | After contracting and access provisioning; access is usually the longer pole |
| Who carries a bad estimate | You do, entirely | The firm, to the extent the price is fixed and the scope held |
| When priorities change | A conversation, once the ramp is already paid for | A change order, or a renegotiated scope |
| What you own at the end | Everything, by operation of employment | Whatever the assignment clause says, so write it down |
| When the need ends | Severance, redeployment, or a role that quietly drifts | Notice period, then the meter stops |
| Where the knowledge lives | In the person, until they leave | In the documents you required as deliverables |
Two rows deserve emphasis. The knowledge row is the strongest argument for hiring and the classic failure of contracting done badly: a working system arrives, the firm leaves, and nobody can run the pipeline. The prevention is boring and it works. Require the runbook, the dependency manifest, and a rehearsed handoff as named deliverables rather than courtesies, and hold the rehearsal while the engagement is still healthy.
The estimate row is the strongest argument for contracting and the one buyers most often give away. If the price is fixed and the scope then moves every two weeks, the risk has walked back to your side of the table without anyone lowering the price.
Conditions that push the answer toward contracting
The bars below are the relative weight we give each condition when the answer comes out contract rather than hire. They are an ordering, not a measurement: a condition at the top usually settles the question alone, one near the bottom rarely does.
Weight of each condition when the answer is contract, not hire
Editorial weighting, presented as an ordering rather than a measurement. Not a statistic.
The inverse test is shorter. If the system being built is the thing your organization sells or the thing it will run for the next five years, hire. Employment is the cheapest way to hold context over that horizon, and no contract structure changes that. The fourth condition deserves a note of its own: a company that cannot evaluate the specialty cannot interview for it either, and a short paid engagement with a written output shows what good looks like in this field faster than a reference call does.
Where the research happens changes the tax treatment
This one moved recently and rarely reaches the people making the staffing call. Public Law 119-21, enacted July 4, 2025, added section 174A to the Internal Revenue Code. Under 26 U.S.C. 174A, domestic research or experimental expenditures may be deducted in the year paid or incurred, with an election instead to capitalize and amortize over not less than 60 months. It applies to amounts paid or incurred in taxable years beginning after December 31, 2024.
Section 174 now carries the title "Amortization of research and experimental expenditures" and applies to foreign research or experimental expenditures, which must be charged to capital account and amortized ratably over 15 years. The statute keys on where the research is performed, not on whether the person doing it is an employee or a contractor.
The consequence is direct. Qualifying research performed in the United States is currently deductible whether you staff it with employees or engage a domestic firm, while the same work performed abroad is recovered over fifteen years. That is a cash-flow difference large enough to reorder two otherwise comparable bids. It is a conversation for your tax advisor rather than a rule of thumb; the point here is that it belongs on the sheet.
The supply side has policy inputs that changed
If part of the hiring plan involves a visa, the ground has shifted and is still moving. Proclamation 10973, "Restriction on Entry of Certain Nonimmigrant Workers," was signed September 19, 2025 and published September 24, 2025. It requires a $100,000 payment in connection with H-1B petitions for workers who are outside the United States, took effect at 12:01 a.m. eastern daylight time on September 21, 2025, and by its own terms runs for 12 months following the effective date unless extended. The Secretary of Homeland Security may grant exceptions for individuals, companies, or entire industries where hiring is determined to be in the national interest.
Two things follow. The twelve-month term reaches its scheduled end in September 2026, and as of publication no later Federal Register document extends, modifies, or rescinds it, so a hiring plan that depends on what happens next carries an unpriced risk. And the surrounding rules are in motion: the Department of Labor published a proposed rule on March 27, 2026 to revise prevailing wage computations for H-1B, H-1B1, and E-3 employment, and the Department of Homeland Security published a rule on August 10, 2026 amending the 9-11 Response and Biometric Entry-Exit Fee for certain H-1B and L-1 visas. Check the rules the week you file, not the week you plan.
On federal work, the rules decide before preference does
If the work sits under a federal contract, several clauses constrain the choice, and reading them first avoids an expensive reversal later.
FAR 52.222-46, Evaluation of Compensation for Professional Employees. Where it applies, offerors submit a total compensation plan covering salaries and fringe benefits for the professional employees who will work under the contract. The government evaluates it for sound management, for the ability to provide uninterrupted high-quality work, and for realism. Compensation lower than a predecessor contractor's draws heightened scrutiny on continuity and competence, and the provision states that failure to comply may be sufficient cause to reject a proposal. Bidding a bargain hire is a proposal risk, not a savings.
FAR 31.205-6, Compensation for personal services. Compensation must be reasonable for the work performed, judged against the practices of firms of the same size, in the same industry, in the same geographic area, engaged in similar non-government work. For costs incurred on or after June 24, 2014, compensation above the benchmark compensation amount set by the Administrator of the Office of Federal Procurement Policy is unallowable, under 10 U.S.C. 3744(a)(16) and 41 U.S.C. 4304(a)(16). The ceiling is on cost allowability, not on what you may pay.
FAR 52.219-14, Limitations on Subcontracting. On a services contract set aside for small business, the prime may not pay more than 50 percent of the amount paid by the government to firms that are not similarly situated entities. Contracting out the scarce skill is exactly the move that can breach this, and work a similarly situated subcontractor further subcontracts counts against the same limit. Run the math before promising a workshare.
FAR 52.244-2, Subcontracts. Depending on contract type and whether the contractor's purchasing system is approved, the contracting officer's consent is required before certain subcontracts are placed, on notification reasonably in advance naming the work, the subcontract type, the proposed subcontractor, and the price. Consent is not a determination that the subcontract is acceptable or its costs allowable, and it does not relieve the prime of performance responsibility.
FAR 52.237-10, Identification of Uncompensated Overtime. Uncompensated overtime is hours worked without additional compensation in excess of an average of 40 hours per week by direct-charge employees who are exempt from the Fair Labor Standards Act. Where it is proposed, the adjusted hourly rate applies to all proposed hours, labor must be identified as regular or overtime by category, and the contractor's uncompensated overtime policy goes in with the proposal. A staffing plan quietly built on fifty-hour weeks from salaried engineers has to be priced honestly.
FAR 37.104, Personal services contracts. A personal services contract is one characterized by the employer-employee relationship it creates between the government and the contractor's personnel, and agencies may not award one unless specifically authorized by statute. The test is facts and circumstances, turning on whether the government will exercise relatively continuous supervision and control. Among the descriptive elements the rule lists are performance on a government site, principal tools furnished by the government, and a need expected to run beyond a year. None alone converts a contract. Together, with a government manager assigning daily tasks, they describe the arrangement the rule forbids.
Key personnel obligations are the other half of this, and they are contract-specific rather than one uniform clause, which is why they surprise people at substitution time. Our piece on what key personnel clauses actually bind covers the mechanics.
Classification is a facts test, and the federal rule is in rulemaking
If the contract path means engaging an individual rather than a firm, classification becomes your problem, and it is decided by facts rather than by the label on the agreement. The IRS applies three common-law categories: behavioral control, meaning whether the payer controls what the worker does and how; financial control, covering how the worker is paid, whether expenses are reimbursed, and who supplies the tools; and the type of relationship, including written contracts and employee-type benefits such as pension, insurance, or vacation pay. Either party may file Form SS-8 for a determination, which the IRS cautions may take at least six months. An employer that classified an employee as a contractor without a reasonable basis may be liable for employment taxes, with Section 530 relief and the Voluntary Classification Settlement Program available in defined circumstances.
At the federal wage-and-hour level the analysis is unsettled and saying otherwise would be misleading. The Department of Labor's 2024 final rule on employee or independent contractor classification under the Fair Labor Standards Act, published January 10, 2024, remains the rule on the books. The Department published a proposed rule on February 27, 2026 to rescind that analysis and replace it with the framework adopted in the January 7, 2021 rule with modifications, extended to the Family and Medical Leave Act and the Migrant and Seasonal Agricultural Worker Protection Act. That is a proposal, not law. State tests are separate and in several states stricter.
The durable answer does not depend on which version prevails. Engaging a firm under a written services agreement, where the firm directs the manner and means and you direct the objectives, is a materially different posture from engaging an individual and treating them like staff. Our article on an embedded engineer inside your team works through that arrangement in detail, including the copyright assignment that commissioned software requires.
The sequence that usually beats choosing
Treating this as an either-or is the most expensive framing available, because it forces a permanent decision at the moment of least information. The sequence below costs less and produces a better role description as a side effect.
Contract a bounded first increment with a written definition of done, sized in weeks rather than quarters. Use it to learn the shape of the standing function: what the work requires day to day, which of the four or five jobs in the description dominates, and what a strong performer looks like in your environment. Write the requisition from that instead of from a template, run the search in parallel from day one, and price conversion now, while everyone is friendly, so that hiring the engineer who did the work is a stated buyout rather than a dispute.
Done this way the contract is not a substitute for the hire. It is what makes the hire land correctly, and it keeps the work moving through months that would otherwise be dead time. Building in house versus bringing in an engineering partner goes further into how the two paths hand off.
Settle these before you choose
- The weekly value of the work that is waiting, computed one of the three ways above, written down where the budget conversation can see it.
- The horizon of the need in months. Under six months rarely justifies a hire; past two years rarely justifies anything else.
- The definition of done for the first increment, in a form someone outside the team can verify.
- Access, named and dated. Repository, data, environments, badging. On regulated work this is the schedule, not a formality.
- Ownership of the artifacts that are not source code — prompts, evaluation sets, labeled data, fine-tuning outputs, model weights. Silence is where disputes start.
- Exit artifacts as deliverables: runbook, dependency manifest, credential rotation list, and a handoff someone on your team runs unaided.
- Conversion terms, priced in advance, if hiring the person who did the work is a plausible ending.
- The federal clauses that apply, read before the staffing plan is promised.
Bottom line
The salary-versus-rate comparison is the wrong instrument. Compare total cost over the real horizon of the need, with the loaded employer cost on one side, the value of the delay on the other, and the ramp counted on both. Where a fixed date sits inside the hiring cycle, the question is feasibility rather than cost. Where the system is the product and the horizon runs for years, hire and stop debating. Everywhere in between, contract a bounded increment, learn what the standing role actually is, and hire against that. On federal work, read FAR 52.222-46, 31.205-6, 52.219-14, 52.244-2, 52.237-10, and 37.104 first, because those clauses settle parts of this question regardless of preference.
Frequently asked questions
Benefits were 31.6 percent of total compensation for civilian workers in the Bureau of Labor Statistics Employer Costs for Employee Compensation series for March 2026, with wages and salaries at 68.4 percent. That ratio is the transferable part; the dollar level is an economy-wide average, not a technical-role benchmark. On top of benefits sit recruiting, equipment, onboarding time, and the weeks the seat stays empty.
When a fixed deadline falls inside the hiring cycle, when the scope is bounded and someone can write down what done means, when the skill is needed once rather than continuously, or when you cannot yet evaluate candidates in the specialty well enough to interview for it. The first is a feasibility test, not a cost comparison.
Only within limits. On a services contract set aside for small business, FAR 52.219-14 bars paying more than 50 percent of the amount received from the government to firms that are not similarly situated entities. FAR 52.244-2 may require the contracting officer's consent before the subcontract is placed, with advance notification. And FAR 37.104 forbids arrangements in which the government exercises relatively continuous supervision and control over contractor personnel absent statutory authority.
It matters where the research is performed. Under 26 U.S.C. 174A, added by Public Law 119-21 and effective for taxable years beginning after December 31, 2024, domestic research or experimental expenditures may be deducted in the year paid or incurred. Section 174 now covers foreign research or experimental expenditures, capitalized and amortized over 15 years. Treat it as a question for your tax advisor and put it on the comparison sheet.
For tax purposes the IRS applies three common-law categories: behavioral control, financial control, and the type of relationship, with Form SS-8 available for a determination that may take at least six months. For wage and hour purposes the Department of Labor's January 10, 2024 final rule remains on the books, and the Department published a proposed rule on February 27, 2026 to rescind that analysis and restore the January 7, 2021 framework with modifications. That rescission is a proposal, not settled law, and state tests apply separately.