Three different reasons, one line on the org chart
When a small AI firm shows up on a prime's team, it is there for one of three reasons — and capture managers get into trouble when they treat the three as interchangeable. The first is arithmetic: a plan has to be filled. The second is eligibility: a vehicle or a program will not accept the team without a qualifying participant. The third is capability: there is a piece of the technical volume that no one at the prime can write, and no one at the prime can deliver. Each reason produces a different subcontract, a different workshare, and a different definition of success.
The arithmetic is well defined. Under FAR 19.702(a)(1)(i), a solicitation expected to exceed $900,000 ($2 million for construction) that has subcontracting possibilities requires the apparently successful offeror to submit an acceptable subcontracting plan, administered through FAR 52.219-9. The government-wide target behind that machinery is set by statute at 15 U.S.C. § 644(g)(1)(A)(i): not less than 23 percent of the total value of all prime contract awards each fiscal year, with separate goals for several socioeconomic categories. Those are dollar mechanisms and they are satisfied with dollars.
Eligibility is a harder gate, because dollars do not open it. For prototype other transactions, 10 U.S.C. § 4022(d)(1) permits award only when one of four conditions holds: a nontraditional defense contractor or nonprofit research institution participates to a significant extent; all significant non-government participants are small businesses or nontraditional defense contractors; at least one third of total project cost comes from non-federal sources; or a senior procurement executive determines in writing that exceptional circumstances justify the arrangement. On SBIR, the small business holds the award outright, and SBA's SBIR/STTR Policy Directive sets a minimum of two-thirds of the research or analytical effort in Phase I and one-half in Phase II to be performed by the awardee, with deviations requiring written approval from the funding agreement officer. A large firm cannot subcontract its way around either constraint; it can only be on the other side of it.
Public-record piece. Every rule cited here is a statute, regulation, published policy directive, or reported decision, named so you can go read it. Nothing here describes a specific customer, program, or engagement.
Capability is the reason that actually decides whether the teaming was worth doing, and it is the one most often assumed rather than tested. A prime that needs the first two things and buys the third by accident will get a sub who invoices correctly and contributes nothing. A prime that needs the third and scopes it like the first will hand real technical risk to a firm selected on socioeconomic status.
| Why the sub is on the team | What it actually obligates | Failure mode |
|---|---|---|
| Subcontracting plan arithmetic | Dollars placed with a qualifying firm, reported and auditable against the plan (FAR 52.219-9). | Scope invented after award to justify the dollars, then absorbed back by the prime when it slips. |
| Eligibility for a vehicle | Genuine, significant participation — the standard in 10 U.S.C. § 4022(d)(1) is participation "to a significant extent," not presence on a chart. | Nominal participation that will not survive a look at the actual labor distribution. |
| Capability the prime lacks | A named deliverable, an interface, an acceptance test, and an engineer whose name is on it. | Buying it on a time-and-materials basis and hoping expertise arrives with the hours. |
What a specialist can move on the bid
The first thing a sub contributes is evaluable past performance in a place the prime has none. FAR 15.305(a)(2)(iii) states that a past-performance evaluation "should take into account past performance information regarding predecessor companies, key personnel who have relevant experience, or subcontractors that will perform major or critical aspects of the requirement when such information is relevant to the instant acquisition." That clause is the reason the sub's record is worth something on the bid — but only for the aspects the sub will actually perform. Scope the sub to a peripheral task and its record stops being relevant to the evaluation.
The corollary is worth knowing before it disappoints someone: FAR 15.305(a)(2)(iv) provides that an offeror without a record of relevant past performance "may not be evaluated favorably or unfavorably on past performance." A young specialist is neutral, not a liability. What it cannot do is manufacture a favorable record it does not have, and a prime that leans on a sub's thin record to carry a past-performance volume has built the weakest part of the bid out of the least substantiated material.
The second contribution is specificity in the technical volume. Evaluators read specificity as evidence of having done the thing. "We will apply machine learning to the sensor feed" and "we will run a quantized detector at a stated latency budget on named hardware, evaluated against a held-out set with a published false-positive rate" are the same claim at two different levels of credibility, and only the second one can be written by someone who has built it. That is the part a specialist writes.
And here is our boundary on that. We write the technical description of the scope we would perform, under our own name, as a member of the team. We do not write proposals for other firms. A shop that writes volumes for work it will never execute is producing text that no one on the delivery side has to answer for — and the resulting technical approach is exactly the one that cannot be built after award.
What a specialist owes on delivery
On the bid, a sub sells a résumé. On delivery, a sub owes an interface. The single best predictor of whether a small technical sub helps or hurts a program is whether its scope was written as a deliverable with an acceptance test, or as a headcount with a labor category. Headcount scopes make the prime's program manager responsible for supervising engineers they did not hire and cannot evaluate. Deliverable scopes make the sub responsible for producing something that either passes or does not.
The work we carry on a prime team is usually the plumbing that everything else stands on: the golden paths and container governance and controls-inheritance documentation described on our platform engineering page, plus the inference component that runs on top of it. Both are deliverable-shaped. A golden path either scaffolds a compliant service or it does not. An inference component either produces the same output twice from the same input or it does not, and a reviewer will check that before anything else.
The seam is where programs actually fail. Decide, in writing and before award, who owns integration: whose repository the artifact lands in, whose pipeline builds it, who runs the acceptance test, and what the escalation path is when the sub's component passes its own tests and the system still fails. Unowned seams do not announce themselves; they surface in month seven as a scheduling problem.
Telling an engineering shop from a pass-through
Every capability statement says the same things. The distinguishing signal is not in the marketing, it is in what a firm can produce on request and what it declines to promise. Six questions separate the two populations faster than any amount of reference-checking.
| Ask | Pass-through answer | Engineering answer |
|---|---|---|
| "Show me something your team built." | A capability statement, a logo wall, a list of certifications. | A repository, a container image referenced by digest, an interface schema, an offline run manual. |
| "What will you not do on this program?" | "We're flexible." Nothing is out of scope. | A specific list, with the reason for each item. |
| "Who signs the technical approach?" | Business development, or an unnamed "technical team." | The named engineer who will perform the work and will be available when it starts. |
| "How will you know the model is wrong?" | An accuracy number on a public benchmark. | An evaluation harness, an error rate the firm considers unacceptable, and a stated behavior on inputs it cannot handle. |
| "What do you need from us before you can build?" | "Just send the statement of work." | Runtime and version, data path, service identity, memory and disk ceilings, who authorizes the transfer. |
| "What happens if the customer descopes your piece?" | Silence, or an assurance that it will not happen. | A termination-for-convenience mechanism and a rate basis already written into the subcontract. |
One more test that costs nothing: ask what the firm thinks the hardest part of the work is. A shop that has built the thing will name a specific unglamorous obstacle — driver ABI mismatches inside a closed enclave, label noise in the training data, the fact that the customer's ground truth is three inconsistent spreadsheets. A shop that has not will describe the opportunity.
The teaming agreement, and why "we'll paper it later" is a real risk
A contractor team arrangement is defined at FAR 9.601, and the government generally recognizes the integrity of such arrangements (FAR 9.603). What the government does not do is enforce them between the parties. That happens in state court, and the case every capture manager should know is Cyberlock Consulting, Inc. v. Information Experts, Inc., 939 F. Supp. 2d 572 (E.D. Va. 2013), aff'd, 549 F. App'x 211 (4th Cir. 2014). Applying Virginia law, the court held that a teaming agreement whose operative promise was to negotiate a subcontract in the future was an unenforceable agreement to agree — leaving the prospective subcontractor with no way to compel the workshare the parties had discussed.
The practical lesson runs both directions. A sub that has done bid-and-proposal work on a promise has no remedy if the promise was drafted loosely. A prime that intended the promise to be real can find itself in litigation over what "approximately 49 percent" meant. Neither outcome is good for a program that has to start on time.
- The subcontract statement of work, attached as an exhibit — not described, attached, so there is nothing left to negotiate later.
- The workshare basis, stated in one unit — percentage of what, measured how, and what happens if the government's scope changes.
- Exclusivity that is bounded — this solicitation, this customer, this period. Blanket exclusivity across a portfolio is a reason to decline.
- Rights in technical data and software, asserted before submission rather than discovered at delivery.
- Named key personnel and a substitution rule — including what happens if a named person becomes unavailable between proposal and award.
- Flow-downs identified by clause number, so the sub can price them instead of absorbing them.
- A termination path that covers descope, customer direction, and non-award, with the rate basis already agreed.
- Which party carries which organizational-conflict restriction under FAR Subpart 9.5 — biased ground rules, unequal access to information, or impaired objectivity — and what future work each party is giving up.
Data rights decide who owns the thing in year five
For a software or AI scope, the data-rights clause matters more than the labor rate and is negotiated with a fraction of the attention. DFARS 252.227-7013 and 252.227-7014 govern rights in noncommercial technical data and computer software, and the categories — unlimited, government purpose, limited or restricted — turn on who funded development. Work developed exclusively at private expense, asserted properly and on time, does not become the government's to hand a competitor.
SBIR and STTR work carries its own regime. DFARS 252.227-7018 defines an SBIR/STTR data protection period that "begins on the date of award of the contract under which the SBIR/STTR data are developed or generated and ends 20 years after that date," absent a negotiated alternative. If a small sub is bringing SBIR-developed software onto a prime's program, both parties need to know which of it is protected, for how long, and what the prime may and may not do with it downstream. Our longer treatment is in SBIR data rights and march-in rights.
Flow down what the data requires, not the maximum
Blanket flow-down is a habit, not a policy, and it costs primes good subs. Send down the clauses the work actually triggers.
When covered defense information is involved, DFARS 252.204-7012 requires the prime to include the clause "including this paragraph (m), in subcontracts... without alteration, except to identify the parties," and it defines rapid reporting as within 72 hours of discovery of a cyber incident, submitted to DoD at dibnet.dod.mil. That one is not optional and not negotiable. CMMC arrives through DFARS 252.204-7021: the program itself is codified at 32 CFR Part 170, and DoD's acquisition rule implementing it in contracts was published on 10 September 2025 with an effective date of 10 November 2025, phased in over subsequent years. The required level flows down according to the type of information the subcontractor will actually handle — which means a prime who flows Level 2 to a sub that will never touch controlled unclassified information has purchased an assessment nobody needed. Background reading: NIST SP 800-171 and CMMC for AI firms.
Export-controlled technical data has its own gate. Receiving militarily critical technical data from DoD requires certification under the Joint Certification Program via DD Form 2345; Precision Federal holds that certification under CAGE 1AYQ0. That is a distinct thing from a facility clearance, which we do not hold today, and which no amount of teaming enthusiasm substitutes for.
Size, affiliation, and the arithmetic behind workshare
If the prime is itself a small business on a set-aside, workshare percentages stop being a courtesy and become a compliance line. Under 13 CFR 125.6 — implemented through FAR 52.219-14 — a small business prime on a services set-aside above the simplified acquisition threshold will not pay more than 50 percent of the amount paid by the government to firms that are not similarly situated entities. Work performed by a similarly situated subcontractor does not count against the limit, but anything that sub further subcontracts outside the class does.
The sharper risk is affiliation. SBA's ostensible subcontractor rule at 13 CFR 121.103(h) treats a prime and a subcontractor as joint venturers — and therefore as affiliates for size purposes — where the subcontractor is not a similarly situated entity and performs the primary and vital requirements of the contract, or where the prime is unusually reliant on it. A small prime that hands the technical core of an AI program to a larger specialist and keeps the program management can lose a size protest over exactly that arrangement. The fix is unglamorous: keep the primary and vital work inside, document the management and technical capability that makes that credible, and scope the sub to a defined component.
Where a small AI sub is the wrong call, and where we draw our own line
Honest scope is worth more to a capture manager than enthusiasm, so here is ours, stated plainly. Precision Federal is a small Iowa limited liability company — Precision Delivery Federal LLC, UEI Y2JVCZXT9HP5, CAGE 1AYQ0. We build platform and inference engineering: internal developer platforms, golden paths and container governance, and small models that read a body of data and produce a written conclusion with every statement traced back to the record it came from. Several things follow from that, and a few of them are refusals.
We do not issue authorizations, and nothing we build arrives pre-accredited
An authorization to operate is granted by an agency authorizing official for a system in a boundary. A component inherits the customer's authorization; it does not create one. Any vendor telling a prime that their software "is IL5" without naming whose boundary it sits inside is describing a hope. We will write the controls-inheritance and boundary documentation an assessor needs, and we will not tell a capture team the ATO is a formality.
We are not a C3PAO and cannot assess anyone for CMMC
We can build to NIST SP 800-171 practices and produce the evidence, but certification assessments are performed by authorized third-party assessment organizations. We are not one, and we will not sit on both sides of that line.
We will not perform your independent assessment and then remediate our own findings
That is textbook impaired objectivity under FAR Subpart 9.5. If we build the system, someone else evaluates it. If we evaluate it, we are out of the running to fix it. This costs us work and we would rather lose it than be the reason a program has an OCI problem.
We hold no facility clearance today and do not perform classified work on classified networks
We build to run inside environments we cannot see, packaged and documented so cleared personnel can execute without us. That is a different capability from performing on classified systems, and we will not blur the two on a bid.
We are not a training vendor and do not write proposals for other firms
We will write the technical description of scope we intend to execute, and we will explain our work to your engineers as part of delivery. We do not sell curriculum, and we do not staff proposal shops.
We decline pass-through arrangements
If the scope exists so a number lands in a subcontracting report, the right answer is a different sub, not a quiet arrangement. Fronting is a False Claims Act exposure for the prime and a reputational end for the sub.
And the honest negative case: if the requirement is forty cleared engineers on site, or twenty-four-hour tier-one operations, or a body shop that will absorb whatever scope the customer invents, a small specialist is the wrong instrument. Say so early. A capture manager who hears "that is not us, and here is the kind of firm you want" from a sub has learned something useful about how that firm will behave in month nine.
What a first conversation looks like
Short. Send the solicitation reference or a description of the requirement, the piece of scope you are considering handing over, and the response deadline. We come back within one business day with a fit assessment — including a plain "no" when it is a no — a proposed workshare and deliverable list, and the questions we would need answered before we could price it. If the scope is sensitive, we sign an NDA first; that takes hours, not weeks. If it fits, the next artifact is a teaming agreement with the statement of work attached, not a promise to negotiate one. Our standing terms are on the teaming page, and the full scope boundary is written out in what we build and what we decline.
Frequently asked questions
FAR 19.702(a)(1)(i) requires the apparently successful offeror to submit an acceptable subcontracting plan for solicitations expected to exceed $900,000 — $2 million for construction — where subcontracting possibilities exist. The plan is administered under FAR 52.219-9. The government-wide small business goal behind it is set at not less than 23 percent of total prime contract dollars by 15 U.S.C. § 644(g)(1)(A)(i).
It can. FAR 15.305(a)(2)(iii) says the evaluation should take into account past performance of subcontractors that will perform major or critical aspects of the requirement, when relevant to the acquisition. The qualifier matters: a sub scoped to a peripheral task contributes little. FAR 15.305(a)(2)(iv) also provides that an offeror with no relevant record may not be rated favorably or unfavorably on past performance.
Not automatically. FAR 9.601 defines contractor team arrangements and FAR 9.603 says the government recognizes their integrity, but enforcement between the parties is a matter of state law. In Cyberlock Consulting, Inc. v. Information Experts, Inc. (E.D. Va. 2013, aff'd 4th Cir. 2014), a teaming agreement that promised to negotiate a subcontract later was held to be an unenforceable agreement to agree. Attaching the statement of work and stating the workshare basis is the practical protection.
Ask for an artifact rather than a capability statement — a repository, an image digest, an interface schema, a run manual. Ask what the firm will not do, and whether it can name the hardest unglamorous part of the work. Ask who signs the technical approach and whether that person will be available at start. Firms that build things answer those quickly; firms that broker résumés change the subject.
Under 13 CFR 121.103(h), SBA may treat a prime and a sub as joint venturers — and therefore affiliates for size purposes — where the sub is not a similarly situated entity and performs the primary and vital requirements of the contract, or where the prime is unusually reliant on it. On a set-aside, that can cost a small prime its eligibility. Related: 13 CFR 125.6 caps what a small services prime may pay firms that are not similarly situated at 50 percent of the amount the government pays it.