AI data market
Qualify once, sell once: the data procurement platform business model
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Procurement platforms move trust upstream: they vet each supplier once, standardize what every listing must disclose, and let buyers order under terms agreed in advance. G-Cloud 13 asked suppliers 56 questions once and each cloud software listing 139 more. A data business can copy that structure but not the economics, because an exclusive dataset sells once and its per-listing diligence never spreads across repeat orders.
Table overview
Qualify once, sell once
Key takeaways
- Procurement platforms earn trust by checking suppliers once and letting many buyers rely on that check.
- G-Cloud 13 asked 56 supplier questions once and 139 questions per cloud software listing, nearly half of them on security.
- US schedules and EU frameworks pair one-time vetting with ordering vehicles that expire, typically within four to five years.
- A catalog listing spreads its diligence over every call-off; an exclusive dataset cannot, so per-dataset work sets a floor on cost per deal.
- The transferable lessons are a reusable master agreement, a short standard per-deal confirmation and disclosure fields buyers can compare.
What can a data business borrow from procurement platforms, and what can it not?
A data business can borrow the trust machinery: vet each supplier once, require every listing to disclose the same facts, and let buyers order under terms settled in advance. It cannot borrow the economics, because a procurement platform's listing is sold again and again while an exclusively licensed dataset is sold once.
Public procurement shows the machinery in its clearest form, because the rules are written down. The UK's G-Cloud framework, the US General Services Administration's schedules and the EU's framework agreements all separate three layers of work: checks on the supplier, disclosures about each offering, and the order itself. Each layer has a different owner and a different unit cost, which is exactly what an investor needs to see in a data company that calls itself a procurement platform.
This article counts those layers in one framework, compares four models on the same ledger, and shows with illustrative arithmetic where the analogy breaks for exclusive data licensing.
What G-Cloud 13 asked each supplier only once
G-Cloud 13 asked every supplier 56 questions, once, as part of its application, however many services it went on to list. The count comes from the framework content the UK's Digital Marketplace published as YAML files, read at the repository's latest commit, 11e5de7, dated 9 January 2023.
The 56 questions fall into ten sections. Twenty-two sit in the two exclusion sections, five mandatory and seventeen discretionary, and the declaration tells suppliers to check compliance with Regulation 57(1) of the Public Contracts Regulations 2015. The other 34 cover suitability of the services, acceptance of the framework terms, working with government, how the application works and its accuracy, subcontracting, contacts and modern slavery reporting. The declaration also warns that the Crown Commercial Service may ask for proof after an application is submitted.
The economic point is reuse. A supplier answers these questions once per framework, to the framework operator rather than to each buyer, which is what lets later buyers start from a vetted list. That is the first lesson for a data business: put rights authority, lawful collection and prior-licensing checks at the supplier level, where they can be reused across deals.
What every listing disclosed before a buyer could award
Each service listing carried far more questions than the supplier declaration. Across 34 sections the service submission listed 203 distinct questions; a cloud hosting listing answered 143 of them, a cloud software listing 139 and a cloud support listing 52. Every follow-up question is also listed in a multiquestion, so all are counted.
Security assurance dominates the hosted lots: 65 of 139 software questions (46.8%) and 71 of 143 hosting questions (49.7%). Use, data and exit questions make up 38 of the software listing (27.3%), including eight on onboarding and offboarding and five on importing and exporting data. Pricing takes five. A buyer comparing two listings therefore compares answers to the same questions, which is what makes a catalog searchable.
That standard disclosure is what makes award without a tender workable. The G-Cloud 15 agreement page describes two routes: award without competition, after using the platform's filters to shortlist services against a written statement of requirements, or a competitive process run by the buyer. Earlier G-Cloud 11 guidance told buyers that a supplier's catalogue information forms and defines the call-off contract, which is why the listing has to be right before award, and allowed call-offs of up to 24 months plus two 12-month extensions.
Repeat purchasing runs on vehicles with an end date
Repeat purchasing in public procurement happens inside ordering vehicles that are vetted once and then expire. The vehicle carries the trust; each order inside it relies on work already done.
In the US, FAR 8.404(d) states that "GSA has already determined the prices of supplies and fixed-price services, and rates for services offered at hourly rates, under schedule contracts to be fair and reasonable," so ordering activities need no separate fair-and-reasonable determination beyond a required price evaluation. FAR 8.405-3 adds blanket purchase agreements: "Ordering activities may establish BPAs under any schedule contract to fill repetitive needs for supplies or services." Multiple-award BPAs generally should not exceed five years, and a single-award BPA may not exceed one year, with up to four one-year options. The text here is the version in GSA's Acquisition-FAR repository at its 30 March 2026 commit.
The EU caps the vehicle too. Article 33 of Directive 2014/24/EU says: "The term of a framework agreement shall not exceed four years, save in exceptional cases duly justified, in particular by the subject of the framework agreement." Recital 62 adds that contracts awarded under a framework need not match its duration and may be shorter or longer.
For a data business, the lesson is that repeat buying needs a standing agreement with a renewal date, and that the standing agreement, not the individual order, is where legal review should concentrate.
The once, per listing, per order ledger
Put every model on one ledger: what is checked once per supplier, what is checked for each listing or dataset, what happens at each order, and the unit an investor should normalize by. The rows below use only the documents cited in this article; where a model's volumes are not public, the last column names the metric to request.
Read the rows for where the work sits. The procurement models push most diligence into the once and per-listing columns and leave the order light. A principal data licensor can keep the supplier column light on repeat deals, but its per-dataset column stays heavy, because each dataset is new and goes to one licensee.
| Model | Checked once | Checked per listing or dataset | At each order | Normalize by |
|---|---|---|---|---|
| UK G-Cloud 13 framework | 56 supplier questions, 22 of them in the exclusion sections | 52 to 143 service questions, by lot | Shortlist against requirements; award with or without competition | Questions per call-off, by lot |
| US GSA schedule with BPAs | Schedule award, with prices determined fair and reasonable | Schedule line items and rates | Order or BPA call; price evaluation where required | Orders per BPA over its term |
| EU framework agreement | Framework award, for up to four years in most cases | Lots and award criteria in the framework | Call-off contract under the framework terms | Call-offs per framework year |
| Principal data licensor (SourceX, by its own pages) | Master Supplier Licensing Agreement and supplier confirmations | Dataset Deal Confirmation, rights review and preparation | One exclusive AI training license per dataset | Labor hours and margin per licensed dataset |
Why an exclusive dataset cannot spread diligence like a catalog listing
A catalog listing spreads its diligence across every buyer who orders it; an exclusive dataset has one buyer, so its per-dataset diligence never shrinks. Simple arithmetic makes the gap concrete. Treat G-Cloud's question counts as stand-in units of diligence, 56 per supplier and 139 per listing, and divide by the number of orders they support.
These figures are illustrative arithmetic, not measurements of any company. They show a structural floor: with one licensee per dataset, cost per deal can never fall below the per-dataset layer, however many datasets a supplier brings. The only levers are making that layer smaller, through standard confirmations, reusable rights evidence and consistent data descriptions, and making sure buyers return.
The four cases below run the same arithmetic under different reuse:
- Catalog listing, one supplier with three listings, each called off by ten buyers: (56 + 3 x 139) / 30 = 15.8 diligence items per order.
- Exclusive datasets, the same supplier with three datasets, each licensed once: (56 + 3 x 139) / 3 = 157.7 items per deal.
- Exclusive datasets, a returning supplier with six datasets under the same master agreement: (56 + 6 x 139) / 6 = 148.3 items per deal, approaching the 139-item floor.
- Catalog listing with a single call-off: 56 + 139 = 195 items, which shows that a catalog with no repeat buyers is no cheaper than licensing.
Is a procurement platform a software business or a services business?
The best-known procurement platforms are mostly software, and that is the second place the analogy can mislead. In its 10-K for the fiscal year ended January 31, 2022, Coupa reported that subscription revenues were 87% of total revenues, with professional services and other revenue making up the rest. Most of what its customers paid for was software subscriptions, not hands-on work.
G-Cloud itself separates the two. Its cloud support lot covers planning, setup, migration, training and ongoing support, and 39 of that lot's 52 questions (75.0%) fall in the sections this count groups as support work offered rather than a hosted product. A data company that calls itself a procurement platform but has its own staff find, clean and prepare each requested dataset belongs in the services column, whatever the label.
To tell which you are looking at, ask for revenue split into subscriptions, per-transaction revenue and services; cost of revenue split into supplier payments, labor and tooling; and labor hours per completed order over the last four quarters. Rising hours per order as volume grows mark a services business.
Applying the ledger to SourceX
SourceX's published terms follow the procurement pattern at the supplier level and depart from it at the dataset level. Its how-it-works page describes one Master Supplier Licensing Agreement, signed once after qualification, and a short Dataset Deal Confirmation for each deal; returning suppliers reuse the Master and sign a new Confirmation. The supplier confirms that it has the rights to license the data, that personal data was collected lawfully, that the data has not been licensed for AI training before and that nobody generated or edited it with AI to prepare the sale, which works much like a framework declaration.
The dataset layer cannot be spread. Licenses are exclusive for AI training, datasets are found for each buyer request rather than kept on a shelf, and delivery waits for the supplier's signature on scope and terms. The investor page states the company's own view that one repeatable agreement shortens legal review for both sides on every deal; that addresses the supplier layer, and the per-dataset layer remains the cost to watch.
The metrics that would test this are not public: supplier return rate under the same Master, review hours per Confirmation, and margin per licensed dataset. The company discusses such numbers with investors rather than printing them. The raise it is conducting under Rule 506(c) is limited to accredited investors, who can ask for materials through its investors page.
Risks to the procurement-platform analogy
Four things could make the lessons here wrong or less useful for a particular company. Check each before relying on the ledger.
- Public frameworks borrow authority from procurement law. FAR 8.404 treats schedule orders and BPAs placed under its procedures as issued using full and open competition, which gives buyers legal cover for using them; a private data business has to earn the same reliance without that support.
- The vehicles expire. Frameworks capped at four years and BPAs that generally run no more than five end on schedule, and the next vehicle has to be won again; a data business that depends on a few standing buyer agreements carries the same renewal risk.
- Non-exclusive licensing changes the math. A company that licenses the same dataset to several buyers would amortize like a catalog, but it would give up the exclusivity that buyers of training data may pay for.
- The counts are a proxy. Question counts measure what a framework asks, not how long answers take or how often buyers rely on them, and G-Cloud content dates from 2023.
Questions to ask, and how the counts were made
Ask any company that calls itself a data procurement platform, SourceX included, to place its work on the three-column ledger and to back each column with figures.
Method: bin/gcloud_count.py read only the YAML content files in frameworks/g-cloud-13 of a clone of alphagov/digitalmarketplace-frameworks at commit 11e5de7. Supplier-level questions are those listed in manifests/declaration.yml; listing-level questions are those listed in manifests/edit_submission.yml. A multiquestion is expanded into its listed sub-questions and not counted itself. A question applies to a lot unless it, or a multiquestion above it, depends on a lot list that omits that lot. A question reached twice counts once and applies to every lot it was reached under; four did. Follow-up questions count only if a manifest or multiquestion lists them; six did not. Theme groups are an editorial mapping of the 34 section names, printed by the script. Percentages are rounded once, half up.
Not measured: how long suppliers took to answer, call-off volumes or values on G-Cloud, schedule or framework spending, or any figure for SourceX. The amortization cases use illustrative reuse counts.
Checklist · 5 items
- What does a supplier answer once, and what must be redone for every dataset?
- How many suppliers returned for a second deal under the same master agreement in the last four quarters?
- What are the median review hours per deal confirmation, and how has that changed?
- How many distinct buyers placed requests, and how many hold a standing agreement with a renewal date?
- What share of revenue is subscription, per-transaction and services, and what share of cost is your own labor?
0 of 5 ticked. Tick each item as you go.
Frequently asked questions
What is a data procurement platform business model?
It is a model in which a company vets data suppliers, standardizes what each dataset offer must disclose and lets buyers order under pre-agreed terms, earning from subscriptions, transactions or services. Whether it behaves like software or like services depends on how much per-order work its own staff do.
What can AI data companies learn from public procurement frameworks?
Frameworks such as G-Cloud check each supplier once and require every listing to answer the same questions, so buyers can compare offers and order without a full tender. G-Cloud 13 asked 56 supplier questions once and 139 per cloud software listing. The lesson is to front-load reusable checks and standardize disclosure.
Why doesn't exclusive data licensing scale like a procurement catalog?
A catalog listing is sold to many buyers, so its diligence is spread across every order. An exclusive dataset is licensed once, so its per-dataset diligence is never shared. Cost per deal therefore has a floor set by the per-dataset work.
How long do public framework agreements last?
Under Directive 2014/24/EU, a framework agreement may not exceed four years except in duly justified cases. Under FAR 8.405-3, multiple-award BPAs generally should not exceed five years, and a single-award BPA is limited to one year with up to four one-year options.
Are procurement platforms software businesses?
Coupa, a large enterprise example, mostly was. It reported that subscriptions were 87% of total revenues in its fiscal year ended January 31, 2022, with professional services and other revenue making up the rest. A data company doing hands-on sourcing for each order is closer to a services firm.
Sources
- G-Cloud 13 framework content at commit 11e5de7: 56 supplier declaration questions in 10 sections and 203 distinct service questions in 34 sections (143 hosting, 139 software, 52 support), as counted for this article. Source
- The G-Cloud 13 declaration tells suppliers to check compliance with Regulation 57(1) of the Public Contracts Regulations 2015 and says the Crown Commercial Service may ask for proof after submission. Source
- FAR 8.404 treats BPAs and orders placed under its procedures as issued using full and open competition, and 8.404(d) says GSA has already determined schedule prices and rates fair and reasonable, so ordering activities need no separate determination beyond a required price evaluation. Source
- FAR 8.405-3 lets ordering activities establish BPAs to fill repetitive needs and limits their duration: multiple-award generally five years, single-award one year plus up to four one-year options. Source
- Directive 2014/24/EU Article 33 limits a framework agreement to four years save in justified exceptional cases; recital 62 says contracts under a framework need not match its duration. Source
- The G-Cloud 15 agreement page describes award without competition after shortlisting services against a statement of requirements, or a competitive process. Source
- G-Cloud 11 customer guidance says a supplier's catalogue information forms and defines the call-off contract and allows call-offs of 24 months plus two 12-month extensions. Source
- Coupa's 10-K for the fiscal year ended January 31, 2022 reports subscription revenues at 87% of total revenues, with professional services and other making up the rest. Source
- SourceX's published terms: one Master Supplier Licensing Agreement signed once, a Dataset Deal Confirmation per deal reused by returning suppliers, exclusive AI training licenses, and the supplier's confirmations about rights and data. Source
- SourceX's investor page says one repeatable agreement shortens legal review for both sides on every deal, and that it shares company and commercial metrics with investors directly. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.