The decision in front of you
A capture manager rarely asks "should we add an AI sub" in the abstract. The question shows up attached to a specific requirement, a specific page count, and a specific date. The RFP has a task that reads like model development, evaluation, and sustainment. Your bench has systems engineers, cleared program people, and maybe two data scientists who are already committed through the option year. Somebody has to write four pages of technical approach that a government evaluator will believe. Somebody has to price it. And your subcontracting plan needs small business dollars that are real, not aspirational.
The honest answer is that a specialist sub is worth adding when it does at least two of three things: closes a scored gap, moves risk off your books at a price you can defend, and puts small business dollars against work the government actually wants done. One of the three is usually not enough to justify the coordination cost. Two is a clear yes. All three, and the only question left is how fast the paperwork can move.
This is written for the person doing that arithmetic: what the plan is scored on, what capability an AI line item needs that a broad systems shop rarely carries, what risk really transfers, and the terms that belong in the teaming agreement before your name and ours appear in the same volume.
The subcontracting plan is scored, and it is scored twice
Under FAR 19.702, a large business receiving a contract expected to exceed $750,000 with subcontracting possibilities has to submit an acceptable small business subcontracting plan, and FAR 52.219-9 puts that plan into the contract. That threshold rises to $1.5 million for construction. The statutory goals behind it live at 15 U.S.C. 644(g): five percent for small disadvantaged business, five percent for women-owned small business, three percent for HUBZone, three percent for service-disabled veteran-owned, with the small business category above all of them.
That is the compliance half. The evaluation half is the one capture managers underweight. FAR 15.304(c)(4) requires the extent of participation of small business concerns to be evaluated as a factor or subfactor in negotiated competitive acquisitions above $750,000. It is not a checkbox appended to the cost volume. It is a scored element in the same source selection that decides the award, and the scoring distinguishes between a plan that names firms with defined scope and a plan that promises percentages to be identified after award.
Reporting closes the loop. Individual Subcontract Reports and the Summary Subcontract Report go into eSRS on the schedule in FAR 52.219-9(d)(10), with ISRs due 30 days after the periods ending 31 March and 30 September. A plan that overpromises produces an ISR that underdelivers, and that record follows the prime into the next competition. The way to avoid it is to put a sub on a defined deliverable at a dollar value both parties intend to bill.
What Moves When an AI/ML Specialist Joins the Team
Editorial weighting drawn from public source-selection practice and capture experience. Illustrative, not a measured statistic.
The capability gap an AI line item actually opens
Most integrators can write a paragraph about machine learning. Very few can write the paragraph that survives a technical evaluator who has read forty of them that month. The gap is not vocabulary. It is the set of decisions an evaluator expects to see made in writing: what the model is being asked to decide, what the baseline is, what metric governs acceptance, what the false-positive cost looks like in the customer's workflow, how the evaluation set was built so it is not the training set with different filenames, and what happens the first month after fielding when the input distribution drifts.
Our engineers write those decisions down because we make them for a living. Our team is led by a former professor in technology who ranks in the top 200 of more than 200,000 on Kaggle, the top one tenth of one percent, and holds seven cloud certifications, with twenty years building production systems for federal agencies across five consulting firms, three of them federal. Behind that sits a standing bench of named engineers, licensed professional engineers, and domain specialists across defense, health, energy, transportation, and public-sector data. When your volume needs a named person against a task, we have one, with a resume the evaluator can check.
The second half of the gap is deployment. A model that runs in a notebook is not a deliverable. Federal AI work lands inside a security boundary, which means containers that pass STIG review, logging that an assessor can read, model documentation that satisfies the customer's governance office, and an inference path that behaves inside IL4, IL5, or an air-gapped enclave. That is cloud and platform engineering work sitting under the data science, and it is where AI line items slip when the sub is a research shop instead of an engineering shop.
What risk moves, and what stays with you
Be precise about this in your own head before you promise it upward. Privity of contract does not move. The government holds you, and FAR 42.1502 records performance against your prime contract, not against ours. A subcontractor does not receive its own CPARS record. If our work is late, your rating absorbs it.
What does move is technical execution risk, and it moves in proportion to how the line item is structured. A firm-fixed-price deliverable with written acceptance criteria transfers real exposure: if the model misses the agreed metric on the agreed evaluation set, the rework happens inside our price. A time-and-materials task order with a vague statement of work transfers almost nothing, because every hour of trouble bills back to you. The contract type you choose for the sub is the lever, and it is a lever most primes leave in the neutral position.
Staffing risk moves too. If a specialist sub owns the AI scope, the prime does not have to recruit, clear, and retain machine learning engineers for a base year of uncertain length. That cost avoided shows up in the bid as a lower and more defensible rate structure than standing up an internal practice for one contract.
An engineering subcontract and a staff-aug body are different purchases
Both arrive on your org chart in the same box. They are not the same instrument, and the difference shows up at closeout, when you find out whether you own anything.
| Question | Staff augmentation | Engineering subcontract |
|---|---|---|
| What you are buying | Hours billed against a labor category | A defined deliverable with written acceptance criteria |
| Who carries technical risk | The prime, entirely | Shared, and priced into the line item |
| What you hold at closeout | Timesheets and a person who leaves | Code, weights, evaluation suite, and documentation |
| Who answers the CO's technical question | Your engineer, reading someone else's work | The engineer who built it, in writing, on the record |
| If the metric is missed | More hours, billed to you | Rework inside the agreed price |
| Effect on the plan | Dollars only | Dollars plus a capability narrative that scores |
The practical test is the artifact list. Ask any candidate sub to name, before award, the exact objects they will hand you: repository, container image, model card, evaluation dataset and test suite, test report, deployment runbook. A shop that answers in ten seconds has done it. A shop that answers with a staffing plan is selling you people, which is a legitimate business and a different purchase from the one your technical volume is describing.
Past performance is the reason to bring us in early
FAR 15.305(a)(2)(ii) lets the source selection authority consider the past performance of subcontractors that will perform major or critical aspects of the requirement. That provision runs one direction only: a prime may cite a sub's record, and a sub may not cite a prime's. If the AI scope is a major or critical aspect, and it usually is when it appears as its own task, then the sub's relevant work becomes usable evidence in your volume.
This is the single most common reason primes call us three weeks before a due date, and the single most common reason it goes badly. Past-performance write-ups need a point of contact, a period of performance, a dollar value, and a description that maps to the RFP's relevancy definition. Assembling that under deadline pressure produces thin write-ups. Assembling it at the capture stage produces strong ones. Send the solicitation number when you start shaping the bid, not when you start pasting the volume together.
If you are the small business prime, the percentages bind you
On a small business set-aside for services, FAR 52.219-14 and 13 CFR 125.6 limit the prime to paying no more than 50 percent of the amount received from the government to firms that are not similarly situated. On an SBIR Phase I, at least two thirds of the work must be performed by the awardee, and at least half on Phase II. STTR splits differently: no less than 40 percent to the small business and no less than 30 percent to the research institution. Work share is a compliance number before it is a negotiating number, and the ostensible subcontractor rule at 13 CFR 121.103(h) means an unusually reliant prime can be found affiliated with its own sub for size purposes.
Defining work share so it survives an audit
Percentages alone are the weakest form of work share, because a percentage of an unknown total is not a commitment either party can plan against. Three definitions layered together hold up much better.
By deliverable. Name the objects, not the effort. "Sub delivers the evaluation suite, the trained model artifacts, the model card, and the deployment runbook for CLIN 0002" is enforceable. "Sub supports AI development" is not.
By dollars, with a floor. State the sub's value as a dollar figure with a minimum, and tie escalation to option exercise rather than to prime discretion. This is what makes the subcontracting plan real, and what makes the eSRS report at the end of the period match what you wrote at the start.
By named personnel and hours. Name the people, name their labor categories, and state the substitution rule: no substitution of key personnel without written consent and an equal-or-better qualification standard. Evaluators read key-personnel sections closely, and so do contracting officers when a resume quietly disappears in month four.
What belongs in the teaming agreement before the bid goes in
FAR 9.601 recognizes contractor team arrangements, and FAR 9.604 says the government will not normally interfere with the arrangement the parties reach. That freedom is useful and it means the paper is entirely on the two of you. Get it signed before your name and ours appear in the same volume.
- Scope, exclusivity, and term, with an end date tied to award or non-award, not left open
- Work share defined by deliverable, dollar floor, and named personnel, with the substitution rule written out
- Background IP schedule listing what each party brings and what stays theirs
- Data rights position stated before the proposal, including any SBIR or STTR assertions
- Flow-down exhibit listing the clauses by number, never "all applicable FAR clauses"
- Bid and proposal cost allocation, normally each party bearing its own
- Organizational conflict of interest representations under FAR subpart 9.5
- What happens on loss, including a clean release and a confidentiality tail
Two of those deserve a sentence more. Flow-downs first: sending your entire clause matrix downhill is the lazy default, and it costs you. A sub handling no covered defense information does not need the full weight of DFARS 252.204-7012 obligations, and a sub that does handle CUI needs it explicitly, with NIST SP 800-171 and the CMMC level from DFARS 252.204-7021 named rather than implied. Send what the data requires. Both parties price the truth better than they price the maximum.
Data rights second. On a DoD contract, DFARS 252.227-7013 and 252.227-7014 govern technical data and noncommercial computer software, and 252.227-7017 requires assertions to be identified before award. SBIR and STTR work sits under DFARS 252.227-7018 with its 20-year protection period. If the model weights, the training pipeline, and the evaluation suite were developed at private expense, that has to be asserted in the proposal. Nobody fixes an unasserted rights position in year five, and year five is exactly when the customer wants to recompete the sustainment.
How we run a first pass on your solicitation
From First Email to Signed Paper
Two notes on that sequence. We answer no when the fit is wrong, and we answer it fast, because a maybe that arrives at day ten is worse for your capture than a no that arrives at hour six. And step five is a commitment, not a courtesy: the people in the volume are the people on the kickoff call.
Where we fit, and where we are the wrong call
We fit where the requirement is production AI, ML, data, or cloud engineering that has to run inside a federal boundary and satisfy someone who will audit it. Model development with a defensible evaluation story. Document and records extraction where the error rate has to be measured, not asserted. Retrieval systems over data the customer cannot move. MLOps on GovCloud or a DoD impact level. Data engineering that turns a customer's messy operational record into something a model can be trusted on. We work as prime or sub, on federal, state, and commercial contracts.
We are the wrong call for hardware fabrication, for pure body-shop staffing where the ask is warm seats against a labor category rather than a deliverable we own, and for scope where the AI content is decorative rather than load-bearing. Saying that plainly saves both of us a week. It is also the same discipline we apply to your bid: a sub that will not tell you what it does badly is not a sub that will tell you when a metric is slipping.
Frequently asked questions
Under FAR 19.702, when the contract is expected to exceed $750,000, the awardee is a large business, and subcontracting possibilities exist. The threshold is $1.5 million for construction. FAR 52.219-9 puts the plan into the contract, and FAR 15.304(c)(4) makes small business participation a scored element in negotiated competitive acquisitions above the same dollar figure.
Yes, when the sub will perform major or critical aspects of the requirement, under FAR 15.305(a)(2)(ii). The reverse is not available: a sub cannot claim the prime contract as its own past performance, and CPARS under FAR 42.1502 records the evaluation against the prime only.
Three layers together. A named deliverable list with acceptance criteria, a dollar value with a stated floor tied to option exercise, and named key personnel with a written substitution rule. Percentages alone leave both parties guessing and produce eSRS reports that do not match the plan.
What the data and the work require, listed by clause number in an exhibit. If the sub handles covered defense information, DFARS 252.204-7012 with NIST SP 800-171 and the applicable CMMC level under DFARS 252.204-7021 belong in the list. If it does not, they do not. Blanket flow-down of an entire clause matrix inflates the sub's price and buys the prime nothing.
Technical text and a priced basis of estimate can be produced quickly. Past-performance write-ups and data rights assertions cannot be produced well under a week of pressure. If the close date is inside three weeks, send it anyway and we will tell you honestly what can and cannot be done in the window.