Federal agencies are large buyers of commercial data, and most commercial data companies misjudge what they are buying. The instinct is that the government wants a discounted version of the enterprise product. It usually does not. It wants a specific set of fields, delivered on a schedule it can rely on, with a licence that permits the uses a government agency actually makes of data, documented well enough that an analyst's conclusion can be defended two years later when someone asks how it was reached. A company that understands that difference can build a renewable federal revenue line out of a product it already ships. A company that does not spends two years selling a dashboard nobody asked for.
This is written for the product or sales leader at a commercial data company: what agencies buy, which agencies buy which categories, how the buying is organized, and the characteristics that make a dataset renew year after year instead of lapsing quietly at the end of its period.
The categories agencies pay for
Six broad families cover most of it. A given company usually sits in one and is adjacent to two.
Reference and entity data. Who a business is, who owns it, what it is registered as, where it operates, what it is related to. Agencies use this everywhere: verifying who they are contracting with, checking eligibility for programs, investigating fraud, understanding a supply chain. What agencies value here is coverage, currency and resolution quality, in that order. The hard problem is that agencies frequently have their own internal identifiers and need to join to them, which makes entity resolution against government-held identifiers more valuable than any additional field.
Risk and screening data. Sanctions, adverse media, enforcement actions, litigation, beneficial ownership, watchlist and politically exposed person data. Buyers are compliance, investigative and licensing functions. Here the valued properties are provenance and auditability: an agency needs to know where a listing came from and be able to show that. A screening result with no traceable source is unusable in a proceeding, so the source citation matters as much as the hit.
Market, economic and financial data. Prices, indices, company financials, credit and market structure data. Buyers include financial regulators, economic and statistical agencies, treasury and budget functions, and program offices doing cost analysis. The distinguishing requirement is point-in-time reconstruction: a regulator needs the value as it stood on a past date, not the current corrected value, because the analysis has to be reproducible.
Geospatial and imagery. Base maps, parcels, addresses, infrastructure, elevation, imagery and derived layers. Buyers span land management, transportation, emergency management, environmental and defense functions. Delivery format is unusually important because agencies have entrenched geospatial tooling, and a product that does not serve standard formats or standard web services adds an integration project to every purchase.
Identity, contact and demographic data. Used for outreach, benefit administration, service delivery and investigations. This category carries the heaviest privacy and legal review and the longest procurement, and the review is not a formality. A company entering it needs its data lineage and its collection basis documented before the first conversation, because that is what the review will ask for.
Analytics and scores built on the above. Risk scores, matching services, classification, forecasting. The important structural fact is that a score is treated differently from a field. A field is data. A score is a model output, and it inherits every question a federal reviewer asks about a model: what it was built on, how it is validated, how it fails, how a person overrides it, and whether its use is documented. Companies that package analytics as if they were data are surprised by this and lose a quarter to it.
What makes a commercial dataset renewable inside an agency
Editorial weighting, illustrative rather than measured. The last row is low because unused fields raise the price and the review burden without moving a renewal.
Who buys, and how the buying is organized
Two structural facts govern federal data purchasing and both matter more than the category.
The buyer is an office, not a department. Federal purchasing happens at the level of a program office, a bureau, a component, a division. A department may have twelve independent buyers of similar data with no shared contract and, quite often, no knowledge of each other. That is an obstacle and an opportunity: it means the first sale is smaller than you expected, and it means the second and third can come from the same department without a new competition.
Data is often bought as a subscription with a base period and options. A typical arrangement funds one year with several option years the government may exercise. The consequence is that renewal is a decision made annually by someone who has to justify it, and that decision turns on whether the data is embedded in a workflow. Data that flows into a system somebody uses every day renews. Data that arrives as a file into a shared drive does not survive the first budget pressure.
The contracting officer and the requiring official are different people with different questions. The requiring official cares about coverage, fitness and whether it solves the problem. The contracting officer cares about the licence terms, price reasonableness, data rights, and whether what is being bought is a commercial item under standard terms or something that needs a negotiated agreement. A company that answers only the first person stalls at the second.
The licence is the product
This is the single most common reason a good dataset fails to sell into government, and it is entirely within the company's control.
Standard commercial data licences are written for a corporate buyer with a defined user population, a single legal entity, and a set of uses that does not include publishing derived findings. Agencies do not fit that shape. An agency may need to share results with another agency, use the data in an enforcement proceeding, retain records for a statutory period, describe its methods in a public report, and place the data inside a system operated by a contractor on its behalf. A licence written without those uses in mind gets sent to a legal review that takes months and comes back with redlines the vendor's counsel has never seen.
The questions to settle before your first federal conversation are concrete. Who counts as a user, and does it include contractors supporting the agency. May results derived from the data be shared with other agencies, and under what conditions. May findings be published or entered into a proceeding. What happens to the data at the end of the term, and does the agency retain anything for its records obligations. May the data be hosted in a government cloud environment or a contractor-operated environment. May it be used to train or evaluate a model. What are the audit rights, and are they compatible with the agency's own security posture.
A company that prepares a government-specific licence addendum answering those questions in advance converts faster than one with better data and a commercial-only agreement. The addendum is a week of work with counsel. It removes a quarter from every deal.
Delivery: the difference between a renewal and a lapse
Delivery mechanics decide renewals more often than data quality does, because a dataset that is hard to consume never gets embedded, and one that never gets embedded has no defender at budget time.
| Delivery model | What it is good for | Why it fails to renew | What to do about it |
|---|---|---|---|
| Portal with logins | Ad hoc lookup by a small number of analysts | Usage depends on individuals; a staff rotation ends it | Pair it with an interface the agency's systems call |
| Bulk files on a schedule | Analytic work, joins against agency data | Ingestion breaks silently and nobody owns the pipeline | Publish a schema contract; version it; notify on change |
| Query interface | Embedding in an agency workflow or case system | Egress restrictions or authentication mismatch block it | Support federated authentication and document every endpoint |
| Deployed inside the boundary | Sensitive workflows and restricted networks | The vendor cannot package it for someone else to run | Build a single deployment artifact with an offline update path |
The pattern across the table is that renewal follows embedding. The strongest position a data company can reach inside an agency is that its feed is a dependency of a system the agency operates. Getting there means the delivery has to survive without the vendor's attention: a documented schema with versioning, a change notification path, an interface that authenticates the way the agency's other systems do, and a failure mode that is loud rather than silent. Silent ingestion failures are the most common cause of a dataset quietly ceasing to matter.
The engineering that makes a dataset federally saleable
Six things, and each is ordinary engineering with a definable scope.
Point-in-time reconstruction. Store data bitemporally: the time a fact was true and the time your system learned it. Agencies routinely need the value as it stood on a past date, for reproducing an analysis, defending a decision, or responding to a review. Retrofitting this is expensive because the history was not kept, so it is worth doing before the market requires it.
Provenance at the field level. For each value, the source, the date acquired, and the transformation applied. This is what makes a screening result usable in a proceeding and what a reviewer asks for when a conclusion is challenged. Storing it at the record level is cheap; reconstructing it later is not possible.
Entity resolution against government identifiers. Agencies hold their own identifiers, and the value of your data multiplies when it joins to them without manual work. Publish the match logic, the confidence score and the reason for a match, because agencies need to review borderline matches rather than accept them.
A schema contract with versioning. A published schema, a compatibility policy, a deprecation window, and notification when anything changes. Agency pipelines are maintained by people with many other responsibilities, and an unannounced field change is how a feed silently stops working.
Coverage and quality metrics published with the data. Completeness by field, freshness by source, known gaps. Agencies prefer a documented limitation to an undocumented surprise, and a company that publishes its own quality metrics is trusted more, not less. It also prevents the worst outcome, which is an agency discovering a gap during an analysis that mattered.
A deployable package. For sensitive workflows, the ability to run inside the agency's environment: a container set with a manifest, an offline update path, no requirement to reach the vendor for the system to function, and a software bill of materials generated by the build. Many of the most durable federal data relationships are in restricted environments, and they are only reachable by companies that can package their product for someone else to operate.
Where a data company's federal engineering effort pays back
Editorial weighting, illustrative rather than measured. The last row is low because extra fields raise price and review burden without changing a renewal.
How the first sale usually happens
Rarely through a competition, at least not the first one. The common sequence is that a program office has a specific problem, someone learns your data addresses it, a small purchase is made to try it, the data is embedded in a workflow, and a larger subscription follows the next year with option years attached.
That sequence has implications for how a company should aim. Target the office rather than the department. Aim at a stated obligation the office already reports on rather than a general capability. Make the first purchase small enough to be made without a lengthy process. Get the licence addendum settled before that first purchase so the second one is not the first legal review. And treat the integration into a workflow as the actual objective, because that is what makes the renewal defensible.
How we work with a data company entering this market
Precision Federal builds AI, data platforms, software and cloud systems and delivers them into production, including inside federal agencies. For a commercial data company, the work usually falls into three areas.
The first three weeks are an assessment with a written output. We read the data model, the delivery interfaces, the pipeline and the licence, and we return a gap list against what agencies require: what has to change to make the data joinable to government identifiers, what provenance is missing, whether point-in-time reconstruction is possible today, what the delivery interfaces need to support federated authentication, and what the licence has to say. Each item costs a stated number of engineering weeks. Companies use it to decide whether to enter, which is a legitimate use of it.
Then we build. Bitemporal storage and the query interface over it, field-level provenance, the entity resolution layer with published match confidence, the schema contract and versioning, quality metrics published alongside the data, the authenticated interface agency systems can call, and where it is needed, a deployment package that runs inside an agency boundary with an offline update path. Delivered against acceptance criteria written as tests: a query that returns state as of a past date and matches a manual reconstruction, a match rate measured against a held-out set, an installation that runs from a clean checkout.
And we build the integrations that create the embedding. Getting the feed into the agency's system is the step that converts a trial into a renewal, and it is usually an integration project the agency has no capacity to run. Doing it as part of the sale is what separates a subscription that renews for years from one that lapses.
What the company keeps. The code, assigned in writing and committed to the company's own repositories from day one. The data, entirely; we never hold it, and where our engineers touch agency data they do it inside the agency's environment under its access controls. The customer relationship, entirely. Our pre-existing tooling is named in the agreement, excluded from the assignment, and licensed to the company perpetually so nothing we bring can strand a future maintainer.
How it is priced. Fixed-price milestones fit most of this because the scope is definable: the assessment, the bitemporal layer, the resolution layer, the delivery interface, the deployment package, an integration. A committed team for a stated term fits where an agency relationship is live and the demands are still forming. The first step is one email with a one-page brief: what the data covers, how it is delivered today, which agencies or programs look like candidates, and the date that matters. We return a scoped, priced statement of work.
Bottom line
Agencies buy reference and entity data, risk and screening data, market and economic data, geospatial products, identity data, and the analytics built on them, and they buy them office by office as annual subscriptions with option years. What decides a renewal is not the size of the field list. It is whether the licence permits what the agency actually does, whether the data joins to identifiers the agency already holds, whether the delivery is a stable machine interface embedded in a working system, and whether every value can be traced and reconstructed as it stood on a past date. Those are engineering and legal properties, both buildable, and a data company that builds them turns a product it already ships into a revenue line that renews.
Frequently asked questions
Six broad families. Reference and entity data about businesses and their ownership. Risk and screening data including sanctions, enforcement, litigation and adverse media. Market, economic and financial data. Geospatial products and imagery. Identity, contact and demographic data. And analytics or scores built on any of those. Buyers are program offices in nearly every department, and the same department frequently has many independent buyers of similar data with no shared contract between them.
Because they are written for a corporate buyer with a defined user population and a narrow set of uses. Agencies need to share results across agencies, use data in proceedings, retain records for statutory periods, describe methods publicly, and let contractors operating systems on their behalf touch the data. A licence silent on those points goes to a legal review that takes months. A government-specific addendum answering them in advance is about a week of work with counsel and removes a quarter from every deal.
Embedding. Data that flows into a system somebody uses daily has a defender when the budget is reviewed. Data that arrives as files into a shared drive does not. That makes the delivery mechanics decisive: a documented and versioned schema, change notification, authentication that matches how the agency's other systems work, and loud rather than silent failure. Silent ingestion failures are the most common way a dataset quietly stops mattering before its option year comes up.
Because government analysis has to be reproducible and defensible later. A reviewer, an auditor or a court may ask what the data showed on a specific past date, not what the corrected current record says. Storing data bitemporally, with both the time a fact was true and the time the system learned it, makes that answerable. It is expensive to retrofit because the history was never retained, which is why it is worth building before a buyer requires it.
Yes. A field is data; a score is a model output and inherits every question a federal reviewer asks about a model. What it was trained or built on, how it is validated, how often performance is measured, what it does when uncertain, how a person overrides it, and what is logged so a past decision can be reconstructed. Companies that package analytics like data are frequently surprised by this and lose time producing documentation they could have prepared in advance.