The gap between winning the work and building it
A federal practice inside a large consultancy sells trust. The relationship with the agency CIO, the modernization roadmap, the assessment that told the program office where its data actually lives. Then the task order lands and somebody has to write code that runs inside the agency's boundary, passes an authorization review, and still works when the contract vehicle rolls over. That is a different business with a different cost structure, and the practices that struggle are not the ones with weak advisory skills. They are the ones that sold a build and staffed it out of a bench sized for slide decks.
This is written for the practice lead who has to decide, this quarter, whether to hire against a pipeline that is not yet awarded. The choice is usually framed as build versus buy. That framing hides the real variable, which is how much of the demand repeats. A practice that will sell the same document-processing build to four agencies over three years should own that capability. A practice that will sell it once, on a five-year program, should not carry the payroll for it between awards.
Why the capacity problem is structural and not a hiring failure
Consulting economics run on utilization. A federal practice targets a billable percentage, staffs to a forecast, and manages the gap with a mix of overtime and bench time. Engineering demand does not fit that model, for three reasons that have nothing to do with recruiting.
Engineering demand is lumpy and award-gated. The practice does not know in March whether it will need eleven engineers in September, because the source selection has not happened. Payroll commitments are made in March. When the award slips two quarters, which is ordinary, the practice carries a fully loaded engineering team against advisory revenue.
Senior individual contributors sit badly in a pyramid. The consulting model prices a ratio: one partner, several managers, many analysts. A principal engineer who has taken three systems through an agency authorization process is not a manager and does not want to become one. The compensation band that keeps that person does not sit anywhere natural on the pyramid, and the ones who are hired anyway tend to leave inside eighteen months for a firm where the career ladder points at engineering.
Skills that are not exercised go stale fast. An engineer who deployed a containerized service into a FedRAMP-authorized environment two years ago and has been writing assessment reports since is not the person who deployed it. The tooling moved. The agency's boundary moved. That decay is invisible on a resume and very visible in week three of a delivery.
Put together, these mean the practice either overbuilds the bench and carries cost, or underbuilds and subcontracts under time pressure at whatever price the market gives that week. Both are expensive. The third option, a standing specialist partner engaged before the pursuit rather than after the award, changes the timing of the commitment rather than the amount of the work.
Conditions that argue for a standing engineering partner over practice headcount
Editorial weighting, illustrative rather than measured. The last row is deliberately low: repeat resale is the strongest case for owning the bench outright.
What a specialist partner actually does inside a federal practice
The word "partner" covers too many arrangements to be useful on its own. In practice there are four distinct things a consultancy buys, and a given program needs two of them rather than all four.
Proposal-stage technical depth. Somebody has to write the technical volume section that describes the data pipeline, the model evaluation approach, the authorization path, and the migration sequence, in enough detail that an evaluator with engineering background scores it above "understands the requirement." That writing is done by people who have built the thing, not by people who have read about it. This work happens before award, at risk, and it is where the specialist earns the right to the workshare.
A delivery team that answers for a result. Not hours. A defined increment with written acceptance criteria: an ingestion pipeline that processes a named corpus at a stated throughput, an evaluation suite that produces a specified report, a service deployed into the agency environment and running from a clean checkout. The practice keeps program management and the client relationship; the partner owns the technical result and is accountable when it misses.
The first month and the last month. Environment standup, identity integration, data access negotiation, pipeline plumbing at the front. Security review, accessibility conformance, documentation, and handover at the back. Both stretches consume calendar out of proportion to their apparent difficulty, and both use skills an advisory bench does not keep current.
Reusable assets the practice keeps. An evaluation pipeline, a document-processing pipeline, an infrastructure-as-code baseline that has already passed a review at one agency. Written correctly into the agreement, these become the practice's property and the second engagement starts at week six instead of week one.
Engineering depth, specifically: what a federal AI or data build actually contains
Practice leads are sometimes sold "AI capability" and receive a demonstration on clean sample data. The distance between that and a production system inside an agency is most of the work, and naming the parts is how a proposal gets scored well and how a delivery gets estimated honestly.
Ingestion and the document reality
Federal document corpora are heterogeneous in ways that break naive pipelines. Scanned filings from the 1990s next to native PDFs next to structured extracts from a mainframe. Multi-column layouts, tables that span pages, form overlays, handwritten annotations, redaction blocks. A pipeline that assumes clean text extraction fails silently: it produces output, the output is wrong, and nobody notices until an analyst disputes a result. The engineering answer is layout-aware extraction with a confidence signal per element, a routing rule that sends low-confidence pages to a different path, and a stored provenance record for every extracted field that names the source document, page, and bounding region. That provenance record is what lets a program answer an oversight question six months later.
Entity resolution and the identity problem
Most federal data problems are joins the source systems were never designed to support. The same organization appears as four spellings across three systems, with different identifiers, none authoritative. Resolution is not a model choice; it is a pipeline with blocking, candidate generation, pairwise scoring, clustering, and a human adjudication queue for the ambiguous middle. The design decisions that matter are the blocking key, which sets the recall ceiling, and the threshold policy, which sets what the program is willing to be wrong about in each direction. A program that has not written down its tolerance for a false merge versus a missed match has not specified the system.
Evaluation as infrastructure, not a report
The difference between a demonstration and a system is that the system knows whether it is still working. That means a held-out set with labels the agency accepts, a scoring script that runs in the deployment pipeline, thresholds that fail a build, and drift monitoring on the input distribution rather than only the output. It also means the evaluation set is versioned and stored alongside the code, because the first question a reviewer asks is what the numbers were measured on. Building this in week two costs a fortnight. Retrofitting it after an oversight inquiry costs a program.
The authorization path, designed in from the first sprint
A system that cannot be authorized cannot be deployed, and authorization is a design constraint rather than a final review. Control inheritance from the hosting platform, boundary definition, data classification and where each class may live, logging and audit trails, encryption at rest and in transit, and the specific controls in the NIST SP 800-53 catalog that the system must satisfy on its own rather than inherit. Section 508 conformance for anything with a user interface. These decisions are cheap in the architecture and expensive in a rewrite. A partner who has been through this arrives with the boundary diagram already sketched.
Where the calendar actually goes on a federal AI build
Editorial weighting, illustrative rather than measured. The last row is deliberately low: model selection is the smallest part of the schedule.
How it appears in a proposal without weakening the practice's position
A capture lead's real worry is not capability. It is that naming a technical subcontractor invites the evaluator to wonder who is really doing the work. That worry is manageable, and the way through it is structure rather than concealment.
Name the specialist with a defined, bounded technical scope and a percentage of the work, and describe the practice's role as program management, agency relationship, mission understanding, and integration accountability. Evaluators read that as a team that knows what it is good at. The alternative, an anonymous resource pool behind a general capability claim, reads worse, especially where the solicitation scores key personnel and asks for named individuals with relevant experience.
Where the solicitation carries small business subcontracting goals, a named specialist with a real technical scope is stronger content for the subcontracting plan than a percentage with nobody attached. The plan is written before award and reviewed after it, and a plan that describes actual scope for an actual firm survives that review in a way a placeholder does not.
Key personnel deserve a specific decision. If the solicitation requires named individuals in technical roles and the specialist supplies them, put them in the proposal as key personnel with the specialist's affiliation stated. Substitution rules then apply to them, which is a constraint worth accepting because the alternative is a proposal that names practice staff who will not do the work. Evaluators who have seen that pattern price it into their confidence rating.
Independence and conflict rules, handled before the pursuit rather than after
Federal advisory work carries constraints that commercial practices sometimes discover late. FAR Subpart 9.5 governs organizational and consultant conflicts of interest, and the two situations that most often catch a consultancy are stated plainly in it. Under FAR 9.505-2, a contractor that prepares or assists in preparing a work statement for a competitive acquisition generally may not supply that system or those services, subject to stated exceptions. Under FAR 9.505-1(a), a contractor providing systems engineering and technical direction for a system without overall contractual responsibility for its development may not be awarded a contract to supply the system or its major components, and may not be a subcontractor or consultant to a supplier of it. The second clause closes the workaround of quietly subcontracting to whoever wins.
FAR 9.504(a) directs the contracting officer to identify and evaluate potential conflicts as early in the acquisition process as possible and to avoid, neutralize, or mitigate significant ones before award. That timing is the practical lesson. A conflict found while the pursuit is being shaped is a teaming decision. The same conflict found after the proposal is submitted is a withdrawal.
A firm whose network includes an assurance practice carries a second constraint that has nothing to do with procurement. Where the build touches systems or numbers the network audits, the arrangement has to be checked against the firm's own independence policy before commitments are made. In both regimes the answer is often that the build sits with a party outside the network, which is an ordinary teaming outcome when it is decided in week one.
How we work inside a federal practice
Precision Federal builds AI systems, data platforms, cloud infrastructure, and full-stack web and mobile software, and delivers them into production inside federal agencies. We work as a specialist subcontractor and teaming partner to large primes, integrators and consultancies. We are a small business, which matters for a practice's subcontracting plan, and the more useful fact is that our engineers have taken systems through agency environments and know what the review boards ask.
The engagement shape is consistent. In the first week we read what exists, talk to the practice's technical lead and the person who owns the agency relationship, and write down the architecture we would build, the data we need, and the risks we see. In the first two to four weeks we deliver something working against real data in a real environment, sized so that stopping costs a few weeks and nothing else. From there the work runs as fixed-price increments against written acceptance criteria, or as a committed team at a stated allocation when the program needs continuity rather than a defined result.
The practice keeps everything that matters. The client relationship is the practice's, always. The code, the pipelines, the infrastructure definitions, the evaluation suites and the documentation are assigned to the practice by present written assignment, with our pre-existing tooling named, carved out, and licensed back perpetually so a future maintainer is never blocked. The data stays where the practice and the agency say it stays. On a proposal we take the posture the practice wants, named with a scored scope or behind the practice's brand, decided before the volume goes out.
The first step is one email with a one-page brief: what the agency has decided and funded, what system the result lives inside, what data exists and who grants access, the security destination, the date that matters, and the contract instrument. We return a scoped, priced statement of work. No call is needed and none is asked for.
Three arrangements, compared
The comparison below is the decision in one view. Most practices need more than one of these live at a time, because the right answer differs by program.
| Dimension | Practice headcount | Staff augmentation | Standing specialist partner |
|---|---|---|---|
| When the commitment is made | Before award, against a forecast | After award, under time pressure | Before the pursuit, with cost at award |
| Who answers for the result | The practice, through its management line | The practice. The supplier owns attendance | The partner, against written acceptance criteria |
| Cost between programs | Continues; utilization decides the margin | Stops; the knowledge leaves with the person | Stops, or a small retained increment for continuity |
| Proposal contribution | Real, if the people are current | None. Suppliers are found after award | Technical volume content and a named scope |
| Subcontracting plan value | None; the work is self-performed | Depends on the supplier's status | A named firm with real scope and dollars |
| Typical failure mode | Award slips and the bench is carried unsold | Nobody is accountable when the result misses | Bought as hours, then asked for an outcome |
The agreement terms that decide how this ends
Most of what goes wrong in a prime-specialist relationship was decided in the master agreement and discovered in delivery. Five terms carry the weight.
- A present assignment of intellectual property rather than a work-made-for-hire recital alone. Under 17 U.S.C. § 101 a commissioned work qualifies as a work made for hire only if there is a written agreement and the work falls within one of nine enumerated categories. Software is a literary work and is not among them. A written present assignment of copyright is what transfers title.
- Background intellectual property named, excluded, and licensed back. Any specialist arrives with existing tooling. List it, keep it out of the assignment, and take a perpetual license to use it in the delivered system so maintenance is never hostage to a relationship.
- Acceptance criteria written as measurements. A throughput number on a named corpus, an evaluation score on a named held-out set, a deployment that runs from a clean checkout in the target environment. Adjectives make acceptance a negotiation.
- Named people, committed allocations, and a substitution path. Ask who does the work, at what percentage, and what happens if award slips a quarter. A stated constraint can be planned around; an optimistic resume cannot.
- Flow-downs stated once, in the master agreement. Data handling, incident reporting, supply chain representations, insurance, and where each class of data may live. Negotiating these per task order is how a two-week turnaround becomes a four-week one.
Five ways this goes wrong
The partner is engaged after award instead of before the pursuit. The proposal then describes a technical approach nobody committed to, and the specialist inherits a scope written by someone who did not have to build it. Engage during capture, at the specialist's risk, and the technical volume improves and the estimate is real.
The partner is bought as hours and asked for an outcome. Hours were priced, so the partner staffed to hours, and the practice expected accountability it did not purchase. Decide which arrangement is being bought and price it that way.
The pilot is built where it can never be deployed. A demonstration on a commercial cloud account with a copy of agency data is a demonstration that dies at the boundary review. Build in the target environment, or against a written authorization path, from the first sprint.
Conflict and independence checks run after the proposal is out. Both are cheap in week one and very expensive in week ten. Run them while the pursuit is being shaped.
Nobody wrote the exit, so knowledge leaves with the partner. A handover is a rehearsal, not a document. Someone on the receiving side should deploy the system while the partner watches, before the last invoice, and the runbook should have been tested by the person who will use it.
What to do between now and the next award
The value of a standing partner is mostly created before there is any work to do. Pick one program in the pipeline where the technical content is the weak part of the story, and use it as the trial. Have the specialist write the architecture section and a two-page technical approach at risk, and read it next to what the practice would have written. That comparison answers the capability question honestly and costs nothing but a few weeks of somebody's attention.
In parallel, put a master subcontract in place with the flow-downs, the intellectual property terms, the rate structure and the insurance already negotiated, so a task order can be responded to in the two or three weeks it actually gives. Practices that skip this step spend the first of those weeks in legal review and submit a weaker technical approach because there was no time left to think about the technology.
Bottom line
A federal practice does not have a hiring problem; it has a timing problem. The commitment for engineering capacity comes due before the award that pays for it, and the decision is usually made under a forecast nobody believes. A standing specialist partner moves the commitment to the point where the revenue is known, supplies technical depth to the proposal at risk, and leaves the practice owning the code, the assets, the relationship and the program. Hire the bench when the same build sells across many accounts and the fixed cost has a variable revenue stream underneath it. Until then, put the master agreement in place, use the next pursuit as the trial, and keep the practice's capital in the relationships that only the practice can hold.
Frequently asked questions
It depends on whether the demand repeats. An engineering bench is a fixed cost against award-gated demand, and the hiring commitment usually comes due before the source selection decision. Hire when the same build sells across several agency accounts, when the practice intends to hold a product, or when the work is continuous. Use a standing specialist partner when the build is program-specific, when the award timing is uncertain, or when the technical depth is needed in the proposal before any revenue exists.
Named, with a defined technical scope and a stated share of the work, while the prime describes its own role as program management, mission understanding, agency relationship and integration accountability. Evaluators read a team that knows its division of labor as lower risk than a general capability claim behind an anonymous resource pool. Where the solicitation scores key personnel, name the individuals who will actually do the technical work with their affiliation stated, and accept the substitution rules that follow.
Often not. FAR 9.505-2 provides that a contractor that prepares or assists in preparing a work statement for a competitive acquisition generally may not supply that system or those services, subject to stated exceptions. FAR 9.505-1(a) bars a contractor providing systems engineering and technical direction without overall contractual responsibility from supplying the system or its major components, and from being a subcontractor or consultant to a supplier. FAR 9.504(a) requires the contracting officer to identify and evaluate these conflicts as early in the acquisition process as possible.
Whoever the assignment clause names, which is why it should be written before work starts. Under 17 U.S.C. § 101 a commissioned software work does not automatically qualify as a work made for hire, so a written present assignment of copyright is what transfers title. Pair it with a named carve-out for the specialist's pre-existing tooling and a perpetual license back to use that tooling in the delivered system, so future maintenance is never blocked by the relationship.
A one-page brief: what the agency has decided and funded, the target system named by product and version, what data exists and who can grant access to it, the security or authorization destination, the date that matters, and the contract instrument. With those, a scoped and priced statement of work comes back quickly. Without them, the first two weeks are spent reconstructing the requirement from a presentation.