A work share is a scope, not a percentage
Most teaming conversations open with a number. The prime offers fifteen percent, the small business asks for thirty, and both sides spend two weeks arguing about a figure neither has yet attached to any actual work. Evaluators do not score percentages. They score described technical work: which contract line item it sits under, which deliverables come out of it, which labor categories produce them, how many hours each one takes, and which named people sign up for it. A percentage is a negotiating position. A scope is an offer.
This matters more for AI and ML than for almost any other work share, because AI scope is unusually easy to write badly. "Machine learning support," "data science services," and "AI/ML expertise as required" are all phrases that have appeared in real teaming agreements, and all three tell an evaluator nothing except that the team has not decided who is doing what. When the technical evaluator cannot tell what the sub is actually building, the sub's hours get read as overhead attached to the prime's labor. That is the moment the work share stops helping the bid.
The fix is mechanical. Decide what the AI component produces, put those products on the line-item structure, price the hours that produce them, and name the people. Every section below is one step of that.
Start at Section B, not the org chart
Federal contracts are organized by line item, and the government's uniform line item rules in FAR subpart 4.10 are stricter than most teams realize. A CLIN has to have a single, defined deliverable or service, one price arrangement, and one accounting classification. You cannot mix a firm-fixed-price deliverable and a cost-reimbursement level of effort in the same line item. In the uniform contract format described at FAR 15.204-2, all of this lives in Section B, and Section B is the first place a contracting officer looks when deciding whether the proposed structure is even administrable.
So the first scoping question is not "what percentage does the sub get." It is "which line items does the AI work produce, and are they separable?" If the AI component produces a trained model, an evaluation report, an inference service, and a set of documented interfaces, those are candidates for their own CLIN or for sub-line items under a larger one. If the AI component only produces a portion of a single integrated system that the prime delivers as one article, then the work share lives inside that CLIN and has to be made visible some other way: through the work breakdown structure, the basis of estimate, and the staffing table.
Data deliverables get their own treatment. On DoD work, technical data and software deliverables arrive through the Contract Data Requirements List on DD Form 1423, attached as an exhibit and cross-referenced to the CLIN that generates them. An AI work share that produces model cards, test reports, training-data documentation, or a software design description should have those CDRL items identified by name in the teaming agreement. This is the cheapest possible way to make a work share look real, and teams skip it constantly.
Where an AI work share belongs, by contract type
The right carve depends on the price arrangement of the parent contract or order. The sub's subcontract does not have to match the prime contract type, but any mismatch has to be deliberate and priced for the risk it moves.
| Parent arrangement | Where the AI work share sits | How the sub is priced |
|---|---|---|
| Firm-fixed-price, completion | Its own CLIN or SLIN when the AI output is a separable article; otherwise a WBS element under the integration CLIN | Fixed price against written acceptance criteria. Vague criteria here is how subs lose money |
| Cost-plus-fixed-fee, completion form (FAR 16.306(d)) | A cost CLIN tied to the end item, with the AI scope carried as a named WBS element and its own basis of estimate | Cost reimbursement with fixed fee. Requires an accounting system the government will accept |
| Term form CPFF or FFP level of effort (FAR 16.207) | A level-of-effort CLIN stated in hours by labor category | Hours at burdened rates. The hour count is the deliverable, so the labor mix carries the technical argument |
| Time-and-materials or labor-hour (FAR 16.601) | Labor-category CLINs, often with separate ceiling amounts per category | Loaded hourly rates against qualified categories. Category qualifications get checked |
| IDIQ with task orders | Nothing until the task order. The base award carries scope areas and rates only | Rates on the schedule, then a per-order carve. Write the teaming agreement to cover order-level allocation |
| SBIR or STTR Phase I | Usually one cost line with an explicit subcontract cost element and a performance-of-work statement | Cost or fixed price by component practice, with a hard ceiling on total subcontracted effort |
What "substantive" looks like on the page
Evaluators are reading for one thing when they get to the teaming section: is the small business doing engineering, or is it doing invoicing? The signals that answer that question are consistent across agencies, and none of them require a security clearance or a past award to produce.
What Carries Weight in a Subcontracted AI Scope
Editorial weighting from public source-selection practice and practitioner reading. Illustrative, not a measured statistic.
Write the AI scope in the same voice the rest of the technical volume uses. If the prime's sections say "the system shall ingest," the sub's section should say what it ingests, what it produces, and what has to be true for the output to be accepted. A model evaluation deliverable is stronger when it names the metric and the threshold: precision at a fixed recall, false-alarm rate per hour of operation, calibration error on held-out data. Reviewers who cannot check a claim discount it.
The limitations on subcontracting, in plain arithmetic
On a small business set-aside, the prime carries a performance obligation set by SBA regulation at 13 CFR 125.6 and implemented in the contract by the limitations on subcontracting clause. For a services contract, the rule is that the prime will not pay more than 50 percent of the amount paid by the government to it to firms that are not similarly situated entities. For supplies other than from a regular dealer, the same 50 percent test applies to the amount paid, excluding the cost of materials. General construction runs at 85 percent excluding materials, and special trade construction at 75 percent.
Three details change how teams should read that. First, the measure is money paid, not hours worked and not scope described, so a work share expressed only in percent of effort does not answer the compliance question. Second, the calculation is applied to the contract as a whole for its full period unless the contracting officer states otherwise, which means an option year with a heavier subcontract load is not automatically a violation. Third, the clause governs set-asides. On a full and open competition, the limitation does not apply, and the compliance conversation shifts to the large business subcontracting plan required by FAR 19.702 for contracts expected to exceed $750,000 with subcontracting opportunities.
Beyond the arithmetic sits a separate risk that ends more teams than 125.6 ever has: the ostensible subcontractor rule at 13 CFR 121.103(h). If the subcontractor performs the primary and vital contract requirements, or if the prime is unusually reliant on it, SBA can treat the two firms as affiliated for size purposes, which can disqualify the offer. The practical guidance is symmetric. An AI work share should be substantial and clearly bounded, and the prime should still own the contract's primary requirement. Both firms want a scope that is deep in its lane and does not swallow the effort.
Similarly situated entities change the math
13 CFR 125.6 excludes payments to similarly situated entities from the amount counted against the limit. A similarly situated entity is a first-tier subcontractor that holds the same small business program status the set-aside was made under, and that is small under the NAICS code the prime assigns to the subcontract. On a plain small business set-aside, that means a small subcontractor under the assigned code. On an 8(a), HUBZone, SDVOSB, or WOSB set-aside, the sub has to carry that same designation.
Two mistakes recur here. The first is assuming the prime's NAICS code carries down automatically. The prime assigns a code to the subcontract, and it should be the code that actually describes the subcontracted work. AI and data engineering work most often lands in 541715 for research and development in the physical, engineering and life sciences, or 541512 for computer systems design services, and the size standards behind those two are different in kind, one measured in employees and one in receipts. The second mistake is checking status once. Status is verified at the time of the subcontract, so a sub that has grown past a size standard between the bid and the award is a live risk on a long capture.
SBIR and STTR run on a different rule entirely
Teams that carry the 50 percent number into an SBIR bid get it wrong, because the SBIR and STTR performance-of-work requirements are stricter and are set by the SBIR/STTR Policy Directive rather than 13 CFR 125.6. On an SBIR Phase I, the awardee must perform at least two-thirds of the research and analytical effort. On Phase II, at least one-half. That leaves at most one-third of a Phase I for everyone else combined, consultants and subcontractors and universities together.
STTR inverts the structure. The small business performs at least 40 percent of the work, the single partnering research institution performs at least 30 percent, and the remaining 30 percent can go anywhere including additional subcontractors. That 30 percent band is where a specialist AI or software firm fits cleanly on an STTR, and it is a materially easier lane than the SBIR Phase I one-third.
Two more SBIR specifics matter to an AI sub. Effort is measured against the research and analytical effort, not against dollars, and the two do not track when one party's labor is far more expensive than another's. And SBIR data rights flow down. On DoD work, DFARS 252.227-7018 protects SBIR technical data and computer software for a defined protection period, and the subcontract has to flow that clause down and identify any software the sub brings in with restrictions, or the protection can be lost on delivery.
Sizing the hours so cost realism does not eat you
Once the scope is carved, the hours have to be real. Cost realism analysis on a cost-type acquisition asks whether the proposed effort is what it would actually take to do the described work, and an unrealistically thin AI work share gets adjusted upward or read as a misunderstanding of the requirement. On the prime side, FAR 15.404-3 requires the prime to conduct its own cost or price analysis of subcontractor proposals before agreeing to them, so the sub's build-up gets read by two audiences.
Use a productive-hour convention and state it. A full-time engineer year is 2,080 clock hours, and after holidays and leave the planning convention most federal cost volumes use lands near 1,860 productive direct hours. Half an engineer for a twelve-month base is about 930 hours. A six-month period of performance halves it again.
Work an example. Take a $250,000 Phase I with a 7 percent fee, $6,000 of travel, and $9,000 of cloud and data ODCs. Fee comes off first, leaving roughly $233,600 of cost. Subtract travel and ODCs and about $218,600 remains for burdened labor. At a blended fully burdened rate of $185 per hour, that buys about 1,180 direct labor hours. Across a six-month base, 1,180 hours is roughly 1.27 full-time engineers. Any technical narrative that promises four workstreams and a fielded prototype on 1.27 engineers is telling the evaluator something unflattering about the team's estimating.
The same arithmetic protects a sub on the receiving end of a work share. If a prime offers "about 20 percent" of a $1.5 million task order, that is roughly $300,000 of subcontract value, which at $185 per hour is about 1,620 hours, which is 0.87 of one engineer for a year. Decide whether the described AI scope fits in 0.87 of an engineer before signing anything. Most do not, and the honest conversation is far cheaper before the bid than after the award.
Build the estimate bottom-up by labor category and defend each one. Fully burdened rates are direct labor plus fringe, overhead, general and administrative expense, and fee, computed under the FAR part 31 cost principles. If the subcontract is cost-reimbursement, the sub needs an accounting system adequate for accumulating costs by contract, and the government's pre-award accounting system survey on SF 1408 is the checklist that gets used. Small businesses are exempt from Cost Accounting Standards coverage under 48 CFR 9903.201-1, which removes one layer of burden, and certified cost or pricing data under FAR 15.403-4 is not required below the $2 million truthful cost or pricing data threshold, which removes another.
The six tells that make an AI scope read as pass-through
Contracting officers and technical evaluators develop a fast read for work shares that exist to satisfy a goal rather than to do work. These are the recurring markers.
- Scope stated as a capability, not a product. "provides AI/ML expertise" with no artifact anyone can accept or reject.
- Labor mirrored across both firms. the same category doing the same task on both sides of the seam, which reads as staff augmentation.
- No deliverable in the CDRL or WBS. the sub's name appears in the management volume and nowhere in the technical structure.
- Hours that fall on a round percentage. a build-up that lands on exactly 20.0 percent was written backwards from the target.
- Unnamed personnel. "a senior data scientist to be assigned" carries a fraction of the weight a named person with committed hours carries.
- No interface definition. no statement of what data crosses, in what format, on what schedule, and who accepts it.
Each of these has a direct fix, and the fix costs a page. Name the artifact. Split the labor by function rather than by title. Put the sub's deliverables in the data requirements list. Build the hours bottom-up and let the percentage be whatever it turns out to be. Name the people. Write one paragraph describing the interface.
The seam: data, interfaces, and who owns what
The technical risk in a split AI scope concentrates at the boundary between the two firms, so the proposal should describe that boundary explicitly. Three questions cover most of it. What data crosses the seam, in what schema, at what cadence, and with what quality gate? What runs where, meaning which side hosts training, which side hosts inference, and in whose accreditation boundary? And what gets accepted, meaning what artifact the prime accepts from the sub and against which written criteria.
The compliance flowdowns sit on top of those answers. If government data crossing the seam is controlled unclassified information, DFARS 252.204-7012 flows down to the subcontractor, which brings the NIST SP 800-171 control set and the 72-hour cyber incident reporting obligation with it. DFARS 252.204-7020 brings the assessment and Supplier Performance Risk System posting requirement. If the sub is receiving export-controlled technical data, DD Form 2345 registration through the Joint Certification Program is what allows it to receive militarily critical technical data from DoD distribution channels. A sub that already holds these is a shorter path to award than a sub who will start the paperwork after the contract is signed.
Intellectual property gets settled in the teaming agreement, before the bid, in writing. Identify background IP each side brings, state what happens to foreground IP developed under the contract, and list any open-source or third-party components with their licenses. A model trained on government-furnished data under a contract with SBIR data rights is a different asset from a model the sub brings in pre-trained, and mixing them without documentation creates a rights problem that surfaces at delivery, which is the worst possible moment.
What we send back when you send a SOW
Our engineers scope AI and ML work shares across defense, health, energy, transportation, and public-sector data programs, as prime and as subcontractor. The process below is what a capture manager gets from us on a live pursuit, and it runs on the bid's clock, not ours.
From SOW to a Priced Work Share
What comes back is short and usable: a one-page scope, an hour table, a named team, and a plain statement of whether our hours count toward your side of the limit or against it. If the honest answer is that the scope does not need us, we say that in step one and you lose nothing but an email.
Common questions on the carve
Should the work share be a separate CLIN or buried in the WBS?
Separate when the AI output is a separable article the government would accept on its own, such as a model, an evaluation report, or a deployed service. Inside the WBS when the AI component is one contributor to a single integrated end item. Forcing a separate CLIN onto integrated work creates administrative friction the contracting officer will push back on, and the visibility can be recovered through the work breakdown structure and the staffing table instead.
Can we set the percentage first and write the scope to match?
It is common and it is visible. Build-ups that land on exactly 20.0 or 25.0 percent read as reverse-engineered, and a cost realism analyst who pulls the hour table finds round numbers with no basis behind them. Build bottom-up, then check the result against the limitation and against the goal. If the number lands short of what the prime committed to, adjust the scope rather than the arithmetic.
How does an AI work share count toward a subcontracting goal?
Under a large business subcontracting plan required by FAR 19.702, subcontract dollars to small business concerns count toward the plan's goals, reported in the Electronic Subcontracting Reporting System. On a set-aside, the frame is inverted: payments to a similarly situated small subcontractor are excluded from the amount counted against the limitations on subcontracting at 13 CFR 125.6. Same firm, two different tests, and both should be checked before the teaming agreement is signed.
What if the solicitation is an IDIQ where scope is decided later?
Write the teaming agreement to allocate at the order level, with a defined scope area rather than a fixed percentage of the base. Name the labor categories and rates that carry forward, state how order-level carves get proposed, and set an exclusivity term with an end date. A percentage promise against an IDIQ with no guaranteed volume beyond the minimum is close to unenforceable, and both parties are better served by a clear scope area.
Frequently asked questions
On a small business set-aside for services, 13 CFR 125.6 provides that the prime will not pay more than 50 percent of the amount paid by the government to it to firms that are not similarly situated entities. Supplies use the same 50 percent test excluding materials. General construction is 85 percent and special trade construction is 75 percent, both excluding materials.
No. SBIR and STTR carry their own performance-of-work requirements from the SBIR/STTR Policy Directive. SBIR Phase I requires the awardee to perform at least two-thirds of the research and analytical effort and Phase II at least one-half. STTR requires at least 40 percent from the small business and at least 30 percent from the research institution.
Divide by the blended fully burdened rate after removing fee, travel, and other direct costs. At $185 per hour, $300,000 of subcontract value is roughly 1,620 hours, which is about 0.87 of a full-time engineer for a year against a 1,860 productive-hour planning year. Run that division before agreeing to a scope, not after.
Scope stated as a capability instead of a product, labor categories mirrored on both sides of the team, no deliverable in the data requirements list or work breakdown structure, hours that land on a round percentage, unnamed personnel, and no interface definition. Each has a one-page fix.
On DoD work with controlled unclassified information, DFARS 252.204-7012 and its NIST SP 800-171 control set and 72-hour incident reporting flow down, along with DFARS 252.204-7020 for the assessment and SPRS posting. SBIR technical data and software protections flow down through DFARS 252.227-7018. Export-controlled technical data requires DD Form 2345 registration under the Joint Certification Program.