Workshare arguments on artificial intelligence task orders are rarely about percentages, even when percentages are what gets said out loud. The real question underneath is whether the prime ends up dependent on a specialist for something it cannot replace, and whether the customer starts believing the specialist is the one delivering. Both worries are legitimate. Both are addressable in the way the work is divided rather than in the way the percentage is negotiated. Once the boundary is drawn properly, the number settles in an afternoon.
This is written for the capture director or program manager on an indefinite-delivery vehicle who is deciding how much of an artificial intelligence or data scope to give a specialist partner, and what to keep. The answer is more specific than a percentage, and it holds through performance rather than only through award.
What the prime keeps, always
Four things belong to the prime and never move, regardless of how much technical content the specialist produces.
The customer. Every scheduled and unscheduled contact with the contracting officer, the contracting officer's representative, the program office and the end users runs through the prime. The specialist attends technical exchanges when the prime brings it, and speaks in the prime's presence or at the prime's direction. This is written into the teaming agreement as a communications clause and it is not a formality; it is the thing that makes the rest of the arrangement safe to sign.
The contract. The prime position, the deliverable schedule to the government, the invoicing, the reporting, the contract data requirements list items, and the accountability for all of it. The specialist delivers to the prime on the prime's schedule, ahead of the government date, with the prime reviewing before anything is transmitted.
The integration and the operations. The system the model lives inside, the interfaces to the rest of the program, the environment, the release process, the help desk and the sustainment. A specialist producing a model that the prime cannot deploy has produced nothing, and a prime that cannot deploy without the specialist has bought a dependency.
The right to take it in-house. Every option year is a decision point. If the code is assigned, the documentation delivered, the deployment defined as code and a handover rehearsed, the prime can staff the work internally at any renewal without a negotiation. A partner worth having writes that into the subcontract and means it.
What determines whether a specialist workshare is safe for the prime
Editorial weighting, illustrative rather than measured. The last row is deliberately low: the boundary protects the prime, the percentage does not.
What the specialist carries
The scope that suits a specialist on an artificial intelligence heavy task order is the part that is deep, narrow, and evaluated on evidence rather than on hours.
The data layer. Ingestion from the agency's systems of record, the schema work, the handling of late-arriving and superseding records, data quality checks that run before anything reaches a model, and the lineage that lets anyone reconstruct where a value came from. This is where most of the real engineering effort on an artificial intelligence program goes, and it is consistently underestimated by teams whose experience is with clean data.
The model and the reasoning behind it. Feature construction, model selection with a written rationale and a baseline comparison, training and versioning, and the packaging that lets the prime's platform team deploy a new version without understanding the training code.
The evaluation evidence. A program, not a spreadsheet: one command that rebuilds the split, runs the models, computes the metrics and writes a dated report with the failure cases attached. This artifact is what the prime shows the government at every program review, and it is the strongest single deliverable a specialist produces.
The model governance material. Documentation of intended use, training data provenance, known limitations, the monitoring plan, the retraining trigger and the audit record structure. Agencies increasingly ask for this and few teams produce it well.
The technical risks assigned to it. Drafted by the specialist, owned by the specialist on the program risk board, reported in the prime's format.
Drawing the line in engineering terms
The boundary that works is an interface, not a percentage. Define it as an artifact contract and the two organizations can work in parallel without weekly coordination overhead.
The specialist delivers a versioned artifact, not a running service. A container image or package with a stated input schema, a stated output schema, a version, and a manifest naming the training data version and the evaluation report that accompanies it. The prime's platform team deploys it exactly as it deploys anything else. Nobody needs to understand the other side's internals to do their job.
The input and output schemas are frozen early and versioned thereafter. Most of the friction on joint delivery comes from schema churn discovered at integration. Write the contract in the first two weeks, version it, and treat a change as a change request with a schedule impact rather than as a conversation.
The evaluation runs in the prime's pipeline, not the specialist's. The specialist writes the evaluation program; the prime's continuous integration runs it on every candidate version. That single arrangement removes the possibility of a number nobody can reproduce, and it means the prime always holds the current evidence.
Data access is the prime's to grant. The specialist works inside the accredited environment, with accounts the prime's program requested, under the program's rules. No copies leave. This matters for the security posture and it matters for the dependency question: the data never sits somewhere the prime cannot reach.
Monitoring belongs to the prime, instrumented by the specialist. The specialist specifies the statistics, thresholds and alerts; the prime's operations team owns the dashboards and the on-call rotation. A retraining trigger fires into the prime's ticket queue, not into the specialist's inbox.
A workshare model with a worked example
The following is illustrative rather than a claim about any specific procurement. Consider a five-year task order under an existing vehicle for an artificial intelligence assisted document review capability at a civilian agency. The scope contains platform engineering, integration with two systems of record, a model component, a reviewer interface, operations and sustainment, and program management.
A defensible division looks like this. The prime carries program management, platform and cloud engineering, systems integration, the reviewer interface, security and authorization work, operations, help desk and sustainment. The specialist carries the ingestion pipeline, the model component, the evaluation program and the governance documentation, plus its share of technical risk ownership.
Expressed as effort, the specialist's portion sits in the range where it is meaningful enough to be described in the management volume and small enough that the prime remains obviously the deliverer. In practice on this kind of scope that has meant somewhere between a fifth and a third of the technical labor, weighted toward the base period and the first option year, declining thereafter as the prime's team absorbs the sustainment. Both the shape and the decline are worth writing into the proposal, because an evaluator reading a flat subcontractor percentage across five years wonders whether the prime intends to learn anything.
| Function | Prime | Specialist | Interface between them |
|---|---|---|---|
| Data ingestion | Grants access, owns the environment and the accounts | Builds and operates the pipeline, owns data quality | Landing zone schema, versioned and frozen |
| Model component | Deploys the versioned artifact through its own release process | Trains, packages, documents, versions | Container with stated input and output schemas |
| Evaluation | Runs it in continuous integration, reports to the government | Writes the program, sets the thresholds | One command, dated report, failure cases attached |
| User interface | Designs, builds, tests for accessibility | Specifies what the reviewer must see when confidence is low | Written interaction requirements, reviewed jointly |
| Authorization | Owns the package and the assessment relationship | Supplies control evidence for its components | Control responsibility matrix, agreed in month one |
| Sustainment | Operates, on-call, absorbs the work by option year two | Retraining and model updates on a stated cadence | Ticket queue and a scheduled release train |
Pricing the specialist's share so the total stays competitive
The concern is that a specialist rate on a labor-hour basis prices the bid out of range. Usually it does not, and where it does the fix is structural rather than a rate argument.
Map to the contract's labor categories honestly. If the vehicle carries fixed rates, the specialist's people map to categories and the arithmetic is what it is. What moves the total is the mix: a specialist team that is small and senior often costs less in total than a larger team of mixed seniority doing the same scope, because the hours are fewer. Price the scope, not the rate.
Use fixed-price increments where the deliverable is well defined. A data pipeline against a described schema, a model component against a stated evaluation threshold, and a governance document set are all definable enough to price as firm increments. Fixed price transfers the estimating risk to the party that controls the work and usually lands below a time-and-materials estimate carrying the specialist's own contingency.
Shape the curve. Weight the specialist's hours to the base period where the hard engineering is, and reduce them across option years as the prime's team takes over sustainment. This lowers the total, reads well to an evaluator, and matches how the work actually behaves.
Count what the specialist removes. A partner who arrives with the pipeline patterns, the evaluation program structure and the governance templates already built is not billing for building them. Where the specialist brings background tooling, name it, license it into the deliverable, and price the increment rather than the reinvention.
Do not buy hours and expect an outcome. The most expensive version of this arrangement is one where the prime purchases labor hours and then holds the specialist accountable for a result nobody scoped. Decide which is being bought and write it that way.
Which levers actually lower the total price of a specialist workshare
Editorial weighting, illustrative rather than measured. The last row is deliberately low: the rate is the smallest term in the total.
What goes in the proposal
The workshare has to be visible to the evaluator, described in three places, consistently.
In the technical volume, the specialist's scope appears as the substance of the data and model sections, written in the prime's voice, with the specialist's name attached where the solicitation asks who performs what. In the management volume, it appears as an organization chart position with a percentage, a reporting line to the prime's program manager, and a described management approach for subcontractor performance. In the small business subcontracting plan, if one is required, it appears as a named firm with a scoped role and a dollar or percentage commitment rather than a placeholder.
Consistency across those three is checked by evaluators more often than teams expect. A specialist described as carrying the model in the technical volume and as providing general technical support in the management volume produces a question, and questions cost points.
Two questions the government will ask about the division
Evaluation boards and, later, contracting officers ask two questions about any meaningful subcontract on a technical scope. Answering them well in the proposal saves a clarification round.
Is the prime performing the work it is being paid to perform? On some contract types and set-aside categories there are limitations on how much of the effort may be passed through, and the prime's own compliance staff will know which apply. The proposal answer is not a percentage in isolation; it is a description showing the prime performing the integration, the operations, the interface and the program management, with the specialist performing a defined technical component. A division described that way reads as a design decision rather than as a pass-through.
What happens if the subcontractor goes away? The strong answer names the artifacts already in the prime's possession: the assigned source, the deployment definitions, the evaluation program running in the prime's own pipeline, the documentation, and the handover rehearsal completed in the current period. The weak answer is a sentence about a mitigation plan. Write the strong one, and make it true by structuring the subcontract to produce those artifacts continuously rather than at the end.
Failure modes worth avoiding
The boundary is a percentage instead of an interface. Twenty-five percent of what, measured how, delivered as what? Teams that negotiate only the number spend the first two months of performance discovering the boundary they should have drawn in week one.
The specialist is given a service to operate. Once a subcontractor runs something in production that the prime cannot restart, the dependency is real regardless of what the contract says. Take the artifact, deploy it yourself.
Schema churn goes unmanaged. The single most common source of joint-delivery friction. Freeze it, version it, treat changes as changes.
The evaluation lives on the specialist's side. Then the prime is reporting numbers it cannot reproduce, which is uncomfortable at a program review and dangerous during a protest or an audit.
The handover is a document rather than a rehearsal. Documentation nobody has followed is documentation nobody can follow. Once per option year, someone on the prime's team deploys the model component from a clean checkout while the specialist watches and says nothing.
How we take a workshare
Precision Federal is a small business engineering firm that builds artificial intelligence, data platforms, software, cloud and full-stack systems and delivers them into production inside federal agencies. On a prime's task order we take the data and model scope, and we take it in the shape described above because it is the shape that leaves the prime free.
In the first two weeks of performance we deliver the interface contract: the landing zone schema, the model artifact input and output schemas, the control responsibility matrix, and the evaluation program skeleton running in the prime's continuous integration. Those four artifacts are what let both teams work in parallel from week three.
Through the base period we deliver the pipeline, the model component as a versioned artifact the prime's platform team deploys, the dated evaluation reports, and the governance documentation. Our engineers attend the prime's technical reviews and the risk board, report in the prime's format, and deliver to the prime's internal dates rather than to the government's.
The prime keeps the customer relationship exclusively, the contract, the deployment, the operations and the past performance. Code, models, pipeline definitions and evaluation programs are assigned to the prime consistent with the data rights asserted to the government; our named background tooling is carved out and licensed back perpetually so nothing we build can ever be withheld. A handover rehearsal is priced as a deliverable in every period, whether or not anyone expects to use it.
Pricing takes one of two shapes and we are comfortable with either: firm fixed-price increments against numeric acceptance criteria, or a committed team at a stated allocation mapped to the contract's labor categories, weighted to the base period and declining as the prime absorbs sustainment.
The first step is one email with a one-page brief: the vehicle, the task order or the anticipated scope, the date that matters, and the technical problem as you see it. We return a scoped, priced statement of work.
Bottom line
The prime keeps the customer, the contract, the integration, the operations and the right to take the work in-house at any option year. The specialist carries the data pipeline, the model artifact, the evaluation evidence and the governance material, delivered across a versioned interface the prime's own team deploys. Draw that boundary as an artifact contract in the first two weeks, keep the evaluation running in the prime's pipeline, price the scope in fixed increments weighted to the base period, and describe the same division consistently in the technical volume, the management volume and the subcontracting plan. The percentage then becomes an arithmetic question rather than a negotiation.
Frequently asked questions
Enough that the scope can be described in the management volume and an evaluator can see what the specialist owns, and no more. The boundary matters far more than the number. A workable division gives the specialist the data pipeline, the model artifact, the evaluation program and the governance material, while the prime keeps the customer, the contract, the integration, operations and the option to take the work in-house at any renewal.
Take a versioned artifact rather than a running service, deploy it through the prime's own release process, run the evaluation program in the prime's continuous integration, keep all data inside the environment the prime controls, and require a handover rehearsal each period where the prime's team deploys from a clean checkout. With those five in place, the code assignment and the documentation make replacement an ordinary staffing decision.
Usually not. The hard engineering sits in the base period and the first option year, and sustainment is work a prime's own team should absorb. Weighting the specialist's hours to the early periods and declining thereafter lowers the total price, matches how the work behaves, and reads better to an evaluator than a flat percentage that suggests the prime does not intend to learn the system.
The prime, exclusively. Every scheduled and unscheduled contact with the contracting officer, the contracting officer's representative and the program office runs through the prime, and the specialist participates in technical exchanges only when the prime brings it and at the prime's direction. This is written as a communications clause in the teaming agreement, and it is what makes the rest of the arrangement comfortable to sign.
Price the scope rather than the rate. Fixed-price increments work wherever the deliverable is definable, such as a pipeline against a described schema or a model component against a stated evaluation threshold, and they usually land below a time-and-materials estimate carrying its own contingency. Map people honestly to the vehicle's labor categories, weight the hours to the base period, and count the work a partner's existing tooling removes.