AI data market
Procurement negotiation data investment: five segments and their gaps
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Procurement negotiation data is investable only where a company's records link each round of bargaining to the award and to what the supplier then delivered, under terms that allow licensing. Public corpora do not cover this: the two we measured hold 6,838 dialogues of study participants dividing campsite supplies or books, hats and balls. Competitive sourcing events score best; pooled price terms carry the largest competition risk.
License terms
Procurement negotiation data
Key takeaways
- The scarce asset is a linked chain from request to bids, approvals, award and supplier performance, not the volume of procurement activity.
- CaSiNo (1,030 dialogues) and Deal or No Deal (5,808 dialogues) teach bargaining over toy goods with no approvals, rounds over weeks or post-deal performance.
- Competitive sourcing events and freight rate quotes rate strongest on recorded outcomes; renewals hold rich reasoning that mostly lives in email.
- Supplier-provided terms set the confidentiality boundary; even SEC filers may redact confidential, immaterial contract terms under Item 601(b)(10)(iv).
- Pooling negotiated prices across competitors is the largest structural risk after the Justice Department withdrew its information-sharing safety zones in 2023.
What would procurement negotiation data sell to an agent developer?
It would sell decision histories with consequences attached: what the company asked for, which suppliers bid, how offers moved over the rounds, who approved the concessions, what was awarded, and whether the supplier then delivered on the agreed terms. An enterprise agent that drafts a request for quotation, compares bids, proposes a counteroffer within its authority or escalates to a human needs exactly that chain, and almost none of it is public.
The work itself is large and documented. The Bureau of Labor Statistics counts 604,400 jobs for purchasing managers, buyers and purchasing agents in 2025, 516,100 of them buyers and purchasing agents, and describes buyers as people who evaluate suppliers, negotiate contracts and review the quality of products. The same Occupational Outlook Handbook page notes that organizations are automating some procurement tasks and integrating AI. That points to employers wanting the work automated, not to anyone's records being licensable.
So the investable claim is narrow. A company whose procurement records link each negotiation step to an award and to what the supplier later delivered holds evidence public data lacks, provided it can license that evidence without breaching what it owes its suppliers or what competition law requires. The rest of this page tests each half of that sentence.
What do public negotiation corpora teach, and what do they leave out?
They teach turn-taking bargaining over toy goods between study participants, and they leave out the business outcome. We read two widely cited corpora directly from their public repositories to see what an agent could learn from them.
CaSiNo holds 1,030 dialogues in which two participants play campsite neighbors dividing packages of food, water and firewood according to private priorities. The data labels participants as Mechanical Turk workers; dialogues average 11.6 utterances, 1,005 of the 1,030 include an accepted deal and 25 end in a walk-away, and the recorded outcomes are points scored, satisfaction and how much each side liked the other.
Deal or No Deal, released with Facebook Research's end-to-end negotiator, contains 5,808 dialogues on 2,236 unique scenarios in which two players divide books, hats and balls with private values. Of the 12,234 training lines in its data file, each a dialogue seen from one player's side, 78.5% end with both players' selections matching, 10.7% with selections that conflict and 10.7% with no deal, after about five messages on average. CraigslistBargain, introduced by He and colleagues, adds price bargaining over items listed on Craigslist; it is consumer haggling, not supplier management.
Models also look weakest on the side procurement needs. In the bargaining benchmark from Xia and colleagues, which treats bargaining as an asymmetric incomplete-information game, seller results were mostly positive while most tested models recorded negative surplus when playing the buyer. One benchmark is not a market, but it is a dated, public sign that buyer-side negotiation is an open capability gap.
Set against what a procurement agent needs, the two corpora we measured look like this:
| What a procurement agent needs | CaSiNo (1,030 dialogues) | Deal or No Deal (5,808 dialogues) | Where it sits in a company |
|---|---|---|---|
| Goods or services a business actually buys | No: campsite food, water and firewood | No: books, hats and balls | Requisitions and item masters in the ERP |
| Real counterparties | No: paired study participants | No: paired study players | Supplier master and sourcing event records |
| Several issues traded at once | Yes: three item types with private priorities | Yes: three item types with private values | Price, payment terms, lead time and quality clauses in quotes |
| Rounds spread over days or weeks | No: one chat averaging 11.6 utterances | No: about five messages before the deal | Email threads and sourcing event rounds |
| Internal approvals for concessions | No | No | Approval workflows and delegation records |
| Recorded final terms | Yes: 1,005 of 1,030 reach an accepted deal | Yes: 78.5% of lines end with matching selections | Award records and signed contracts |
| What happened after the deal | Only satisfaction and liking ratings | No | Receipts, delivery dates, quality and invoice records |
The procurement negotiation scorecard: five segments rated on four tests
Competitive sourcing events and freight rate quotes lead the scorecard, renewals hold valuable reasoning that is hard to link, and tail-spend quotes teach little negotiation. The ratings are editorial judgments of the records each segment usually produces, not measured model utility.
Each segment is rated on four tests: whether the agent decision is specific enough to grade, whether the outcome is recorded in a system rather than remembered, how heavy the confidentiality burden is, and what evidence is still missing before an investor should believe the segment is worth supplying. Strong, partial and weak mark the outcome test because it decides whether the records can train or evaluate anything at all.
The five segments, rated and with their largest gaps named:
| Segment | Agent decision to test | Outcome evidence in records | Confidentiality boundary | Biggest evidence gap |
|---|---|---|---|---|
| Competitive sourcing events (RFQs, RFPs, reverse auctions) | Shortlist bidders, normalize bids, recommend an award or another round | Strong: bids, rounds and award sit in one sourcing tool; delivery data in the ERP | Supplier bids and pricing under NDA or event terms | No public benchmark of multi-round B2B bid events located; utility unmeasured |
| Renewals and supplier price-increase requests | Accept, counter or re-source when a supplier raises price | Partial: final terms in the contract system, reasoning mostly in email | Contract terms companies treat as confidential | Whether email threads can be linked to final terms at scale |
| Direct-materials term negotiation | Trade price against lead time, payment terms and quality clauses | Strong to partial: orders, receipts and quality records show whether terms held | Drawings and specifications may belong to customers or suppliers | Preparation cost of joining specifications, quotes and quality outcomes |
| Freight and logistics rate quotes | Accept or counter a carrier's quoted rate on a lane | Strong: quoted, accepted and paid rates in the transportation system | Carrier rates; pooled rates across shippers draw competition scrutiny | Whether one company's lanes generalize to another's network |
| Tail-spend quick quotes | Pick among a few quotes for a low-value purchase | Weak: the outcome is mostly the price paid | Lighter, but quotes still carry supplier terms | Little negotiation in the record to learn from |
Where does the confidentiality boundary run through a negotiation record?
It runs around whatever the counterparty gave you. The company's own requests, decisions, approvals and stated reasons are usually its own to license once rights review clears them; quotes, price lists and proposals came from suppliers, often under a nondisclosure agreement or the terms of a sourcing event, and supplier representatives' names and contact details add personal data on top.
Even mandatory public disclosure respects that line. Under Item 601(b)(10)(iv) of Regulation S-K, a company filing a material contract with the SEC may redact specific provisions or terms if it customarily and actually treats them as private or confidential and they are not material; it must bracket the omissions, state on the first page why they were excluded, and supply an unredacted copy if the SEC staff ask. Negotiated terms are the kind of information companies keep out of view even when regulation requires the contract itself to be filed.
A permissible decision history therefore keeps the company's decisions and the outcomes, such as which bid won and whether delivery was on time, and treats supplier-provided terms one supplier at a time: remove them, band them or leave that supplier out. Our guide to supplier pricing inside procurement records walks through those treatments field by field.
Can pooled negotiation terms become a competition problem?
Yes, it can, and an investor should ask how any negotiation data business keeps current price terms from passing between competitors. Pooling negotiated terms from many companies is valuable precisely because it reveals what others pay, which is also what makes it sensitive.
The U.S. Justice Department removed a long-standing comfort zone on this point. On February 3, 2023, its Antitrust Division withdrew three health care antitrust policy statements from 1993, 1996 and 2011, saying they were overly permissive on certain subjects, such as information sharing. In remarks explaining the move, Principal Deputy Assistant Attorney General Doha Mekki described the old safety zones as circumstances in which the Division would decline to challenge exchanges of competitively sensitive information, and said exchanges run through intermediaries can have the same effect as direct exchanges between competitors (paraphrased from the Department's published remarks).
A data company that aggregates negotiated prices is such an intermediary. Licensing one company's history at a time, aging or aggregating price terms, and limiting use to model training rather than price benchmarking are design choices worth asking about; whether any of them is enough is a question for antitrust counsel on the specific facts, not for a diligence memo.
Does a large purchasing workforce mean large licensable supply?
No. A workforce of 604,400 purchasing professionals shows how much negotiating happens, not how much of it survives as licensable, linked records. This is the strongest alternative explanation for any sector thesis built on headcount.
Five filters stand between the work and the supply. The negotiation has to be recorded in a system rather than in a buyer's personal inbox or phone calls; the steps have to link to an award and to what happened after it; supplier terms have to allow use, or be removable without gutting the record; the dataset has to be designed so it does not pass current prices between competitors; and an AI developer has to want that task. Any one filter can remove most of a company's archive.
Volume still matters once the filters pass. As a general guide that is assessed deal by deal, SourceX notes that buyers look for years of consistent history, typically two or more, and tens of thousands of records or more in one system, while smaller sets can qualify when the work is rare. Three years of sourcing events in one tool fits that shape far better than a decade of scattered email.
What could make the procurement negotiation thesis wrong?
Five developments would weaken the case for procurement negotiation records; each can be tracked in public.
- Simulation substitutes for records. Deal or No Deal's repository ships self-play tooling, and negotiation skill can be practiced between models; if simulated suppliers prove good enough, real histories lose their premium.
- Supply proves thinner than the workforce suggests. If most reasoning lives in email and most terms sit under supplier confidentiality, few archives pass the five filters above.
- Competition enforcement tightens further. The 2023 withdrawal of the information-sharing safety zones signals scrutiny of pooled price data, and a data business built on pooling would carry that risk directly.
- Demand stays unproven. We found no dated public statement from an AI developer seeking procurement negotiation records specifically, and no published study measuring what such records do for a procurement agent.
- General models close the buyer-side gap. The weakness Xia and colleagues measured is on one benchmark; broader reasoning gains could narrow it without any procurement data.
Questions for a procurement data supplier, SourceX included
Ask any supplier of negotiation data to show the chain from request to delivered outcome on a sample, and to explain who signed off on every supplier term it contains. SourceX should meet the same standard.
With SourceX, the business that owns the procurement archive approves the scope and terms and signs the agreement before any records are delivered, and SourceX may then use, sublicense and distribute the de-identified dataset only as that agreement allows. SourceX keeps company and commercial metrics off its public pages and provides them to investors on request. Its raise is open only to accredited investors under Rule 506(c), and materials can be requested at sourcex.si/investors along with any of these questions:
- Which procurement segment does each dataset cover, and which agent decision will a buyer grade with it?
- What share of negotiation steps link to an award and to post-award delivery or quality records?
- For each supplier in the data, what term governs use of its quotes, and how were its prices treated?
- How does the design keep current, company-specific price terms from reaching competitors?
- Which AI developer has asked for this segment in writing, and for training, evaluation or both?
- What would show that these records improve a procurement agent beyond what simulated negotiation achieves?
Methods note: corpora counted, segments rated and what was not measured
Corpora: CaSiNo from github.com/kushalchawla/CaSiNo at commit 2f6ed4a (September 7, 2021) and Deal or No Deal from github.com/facebookresearch/end-to-end-negotiator at commit bbb93bb (May 3, 2020). For CaSiNo we counted dialogues in data/casino.json, utterances excluding deal-action markers, and dialogues containing an accepted deal or a walk-away. For Deal or No Deal we parsed the outcome after the reward field on each line of src/data/negotiate/data.txt; the dialogue and scenario counts come from the repository README. The script, negotiation_corpora.py, is kept with this article's review files.
Scorecard: segments and ratings were drawn up on October 27, 2026 from the record types each segment usually produces and the sources cited here; they are editorial judgments. Workforce figures come from the Bureau of Labor Statistics page as checked on that date.
Not measured: the volume or quality of any company's procurement records, the share that would pass rights review, buyer demand for each segment, and the effect of procurement records on any model. CraigslistBargain was described from its paper, not counted.
Frequently asked questions
Is there a public dataset of real B2B procurement negotiations?
We did not find one. The public corpora we measured, CaSiNo and Deal or No Deal, pair study participants who divide campsite supplies or books, hats and balls, and CraigslistBargain covers consumer price bargaining over Craigslist items. None records internal approvals, rounds over weeks or what a supplier delivered after the award.
Why would an AI developer want procurement negotiation records?
An agent that drafts requests for quotation, compares bids or proposes counteroffers needs examples of real offers moving toward an award, with the outcome attached. Benchmarks such as the one from Xia and colleagues suggest models are weaker on the buyer side of a bargain, which is the side procurement agents play.
Can a company license negotiation records that contain supplier prices?
Sometimes, but supplier prices usually need their own review because quotes and price lists often arrive under nondisclosure agreements or sourcing event terms. Common treatments are removing prices, banding them or excluding a supplier, while the company's own decisions and outcomes stay. Counsel should confirm what each supplier agreement allows.
Does pooling negotiated prices for AI training raise antitrust questions?
It can. The Justice Department withdrew its health care information-sharing safety zones in February 2023 as overly permissive, and its officials have said exchanges through intermediaries can harm competition like direct ones. Ask whether a dataset is built from one company at a time and how its price terms are aged or aggregated, and have antitrust counsel review the specific design.
Does SourceX publish how many procurement datasets it has licensed?
No. SourceX does not publish deal, supplier or commercial figures; it shares company and commercial metrics with investors directly. Investors can request materials at sourcex.si/investors, and any offering is made only to verified accredited investors through SourceX's offering documents.
Sources
- BLS counts 604,400 jobs for purchasing managers, buyers and purchasing agents in 2025, says buyers evaluate suppliers and negotiate contracts, and notes organizations are automating procurement tasks and integrating AI. Source
- CaSiNo contains 1,030 campsite negotiation dialogues over food, water and firewood, with outcomes recorded as points scored, satisfaction and opponent likeness; this article's counts come from its data file. Source
- The end-to-end negotiator repository releases the Deal or No Deal dataset of 5,808 dialogues on 2,236 scenarios, each converted into two training examples; this article's outcome shares come from its data file. Source
- He and colleagues (2018) separate negotiation strategy from generation and test on DealOrNoDeal and a new dataset built from real Craigslist items. Source
- Xia and colleagues frame bargaining as an asymmetric incomplete-information game; seller results were mostly positive while most models recorded negative surplus as buyers. Source
- On February 3, 2023 the Antitrust Division withdrew its 1993, 1996 and 2011 health care policy statements as overly permissive on subjects such as information sharing. Source
- Doha Mekki's remarks explained that the withdrawn safety zones covered exchanges of competitively sensitive information and that exchanges through intermediaries can have the same effect as direct ones. Source
- Regulation S-K Item 601(b)(10)(iv) lets registrants redact contract terms they customarily and actually treat as private or confidential and that are not material, with brackets and a first-page statement. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.