Procurement, samples and ongoing supply
Paid Data Pilots: Pricing, Credits and Conversion Terms
Quick answer
A paid data pilot should be priced and papered as the first step of the full license, not a standalone purchase. Agree five terms at pilot signature: a pilot fee credited against the full license if you sign within a fixed option window, a locked unit price and volume tiers for the conversion, evaluation-only rights on the pilot slice, written pass/fail criteria, and a deletion or retention rule if you walk away. Without these, a successful pilot reopens every commercial question.
By SourceX Editorial · Updated
This guide covers economics and conversion. For pilot scope, sample design and success measures, see the owner guide on how to run a data pilot with a supplier; for the full procurement path, start at the AI training data procurement hub.
This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.
Why a pilot should be paid, and what the fee buys
A paid pilot buys you a representative slice under real license terms, which a free sample rarely provides. Practitioner RFP guidance consistently says proof on your own data beats vendor demonstrations [1], and AI vendor RFP templates now carry proof-of-concept requirements as their own section before commitment [2]. For training data, "your own data" means running the pilot slice through your actual pipeline: your tokenizer, your SFT mixture, your eval harness.
The fee changes supplier behavior. A supplier preparing a free sample often pulls the cleanest records; a supplier delivering a paid slice under a defined record specification is accountable for representativeness. The fee also funds work the supplier must do anyway for the full deal, such as de-identification, schema mapping and export, so it is reasonable to treat it as a down payment rather than sunk cost. Teams on tight budgets can weigh pilot size against spend using data licensing for AI startups on a budget, and buyers who want operational records sourced on request can describe the data to SourceX.
Keep the pilot distinct from a no-fee look at records under NDA. That route is covered in evaluation licenses and NDAs for dataset samples and suits a schema check, not a training run.
Crediting the pilot fee against the full license
The cleanest structure credits 100% of the pilot fee against the first invoice of the full license, conditioned on signature within the option window. Write the credit as a fixed currency amount, not a percentage of an unknown future total, so finance can book it.
Points to settle in the credit clause:
- Scope of credit. Credit against license fees only, or also against delivery, preparation and refresh fees. Buyers should push for the broadest base.
- Expiry. The credit lapses at the end of the option window. When suppliers resist, one compromise is a sliding credit (for example, 100% in the first 30 days, 50% after).
- Partial conversion. If you license only some fields or record types, state whether the credit applies in full or pro rata.
- Records already delivered. Pilot records that fall inside the full license scope should not be re-billed. Name them by record ID range or manifest hash so both sides can count them.
Model the credit inside your total cost of ownership for licensed training data so the pilot fee is not double-counted against your budget.
Locking price and volume tiers at pilot signature
Lock the conversion price when you sign the pilot, because the pilot result gives the supplier leverage you should remove in advance. If the pilot shows the data lifts your eval scores, an unlocked supplier can reprice; a locked one cannot.
Fix these numbers in the pilot order:
- Unit basis (per record, per document, per hour of recording, per token) and the counting rule, including how duplicates and rejected records are excluded.
- Volume tiers with breakpoints, so the price for the second and third delivery is known.
- Any floor or minimum commitment for the full license.
- Refresh pricing if you expect ongoing supply; see ongoing data supply agreements.
Normalize the locked price to cost per usable record, as described in comparing data vendor quotes, using the rejection rate the pilot actually measured. The pricing structures comparison explains which unit fits which data type.
Setting the option window and conversion trigger
An option window gives you the right, not the obligation, to sign the full license on pre-agreed terms within a fixed period after pilot acceptance. Start the clock at written acceptance of the pilot delivery, not at pilot signature, so delays in delivery do not eat your evaluation time.
Size the window to your real evaluation cycle. An SFT ablation, a held-out eval run and a legal review of the full license often run in sequence, so a window shorter than that cycle forces a decision before results exist. Tie conversion to objective criteria written in advance: completeness and validity rates, a maximum duplicate rate and a replacement percentage are typical thresholds market guides recommend fixing before the order [3]. ISO/IEC 5259-4 gives a process framework for data quality in training and evaluation that both sides can reference for those criteria [5].
Say what happens when results are mixed. One option is a right to convert on a reduced scope, for example dropping a record type that failed, with the locked unit price unchanged. Align the thresholds with your acceptance criteria for licensed training data so the pilot and the full deal measure the same things.
Restricting pilot rights to evaluation
Pilot rights should be narrower than full license rights, and the order should say exactly how. Evaluation-only means you may load, inspect, run ablations and measure model behavior on the slice; it should state whether weights trained during the pilot may be kept, and whether they may ever be deployed.
Define these terms explicitly:
- Permitted activities. Exploratory analysis, fine-tuning experiments, eval runs. State whether training a production candidate is permitted.
- Model artifacts. Whether checkpoints trained on pilot data must be deleted on non-conversion, or may be retained for internal comparison only.
- Derived data. Embeddings, filtered subsets, synthetic examples generated from pilot records, and eval items built from them.
- Access. Named teams and environments; no transfer to affiliates or contractors without consent.
Rights obligations often survive termination and flow down to the supplier's own sources [4], so read the pilot rights against the supplier's upstream permissions, not just the order. The evaluation-only data license terms page covers these clauses in depth.
Handling pilot data if you do not convert
Agree the non-conversion path before any records move. Vague wording about what "return or destroy" covers is a frequent source of pilot disputes.
A workable non-conversion clause specifies a deletion deadline after the option window lapses, a written certificate of deletion signed by a named owner, and treatment of backups and logs. It should also carve out what you may keep: aggregate metrics, eval scores and your own written findings, without record-level content. If pilot checkpoints may be retained, the clause must say so, or the default reading is deletion.
Ask the supplier for per-slice documentation, such as a Data Card describing upstream sources, collection and annotation methods and intended use [6], so your evaluation records which slice was tested and why it may or may not generalize. NIST's voluntary AI RMF 1.0, organized around GOVERN, MAP, MEASURE and MANAGE, is a useful frame for recording this evidence; dataset documentation fits most naturally under MAP and MEASURE [7].
Pilot term sheet you can adapt
The term sheet below collects the commercial fields a procurement manager should fill before a pilot order is signed. Values are placeholders.
Illustrative example: invented to show structure; it does not describe an available dataset.
| Term | Example entry | Why it matters |
|---|---|---|
| Pilot slice | 5,000 closed support tickets, JSONL, fields: ticket_id, created_at, channel, product_area, messages[], resolution_code | Defines what "representative" means |
| Pilot fee | Fixed amount, invoiced on delivery | Funds preparation and de-identification |
| Fee credit | 100% against first license invoice if signed within window | Avoids paying twice for the same records |
| Option window | 60 days from written acceptance of pilot delivery | Matches the SFT and eval cycle |
| Locked price | Per-record rate, tiers at 100k and 500k records | Removes repricing after a good result |
| Conversion criteria | Completeness at or above threshold, duplicate rate at or below threshold, replacement for failed records | Objective conversion trigger |
| Pilot rights | Internal evaluation and fine-tuning experiments; no production deployment | Keeps pilot rights narrower than license rights |
| Pilot checkpoints | Retain for internal comparison only, or delete | Settles the most disputed artifact |
| Non-conversion | Deletion within 30 days of window lapse, signed certificate, metrics retained | Clean exit |
| Records carried over | Pilot record IDs count toward full license volume | No re-billing |
Worked credit example (illustrative). A pilot fee of 20 units is paid for 5,000 records. The full license is signed on day 45 of a 60-day window at a locked rate covering 200,000 records. The first invoice is reduced by 20 units, and the 5,000 pilot records count toward the 200,000, so the supplier delivers 195,000 new records.
Sourcing operational data to license after a pilot
SourceX sources operational datasets from US companies on request and manages the commercial process, including licensing agreements and ongoing purchases; it does not hold inventory, and a request does not guarantee a match. Each dataset is rights-reviewed and delivered under a license defining records, uses, term and delivery, with pricing and allowed uses agreed per deal, and nothing is contracted until the supplier agrees. If you need support histories, engineering records or document workflows, describe the data you need.
Frequently asked questions
Should the pilot fee be refundable if the data fails?
Pilot terms more often use credits than refunds, but you can tie a partial refund or replacement records to failed conversion criteria. Writing refund and replacement terms into the pilot order is recommended market practice [3]. The remedies menu is covered in remedies when a data delivery fails.
Can the supplier change the slice between pilot and full delivery?
It should not without notice. Specify that the full delivery uses the same source systems, field definitions and preparation method as the pilot, and that changes trigger re-testing against the conversion criteria.
How do I keep the pilot from becoming a free training run?
Limit permitted activities, define checkpoint treatment, and require deletion certificates. Suppliers are more willing to price a modest pilot fee when these limits are explicit.
Sources
- Amit Koth, "AI RFP Template". https://amitkoth.com/ai-rfp-template/
- Dan Cumberland Labs, "AI Vendor RFP Template". https://dancumberlandlabs.com/blog/ai-vendor-rfp-template/
- CloudPano, "How to Evaluate an AI Training Data Provider: 7 Questions to Ask". https://www.cloudpano.com/blog/how-to-evaluate-ai-training-data-provider
- Morgan Lewis, "Key Concepts in AI Contracting: Data Rights and Restrictions" (2025). https://www.morganlewis.com/blogs/sourcingatmorganlewis/2025/12/key-concepts-in-ai-contracting-data-rights-and-restrictions
- ISO/IEC JTC 1/SC 42, "ISO/IEC 5259-4:2024 Artificial intelligence - Data quality for analytics and machine learning (ML) - Part 4: Data quality process framework" (2024). https://www.iso.org/standard/81093.html
- Google Research (Pushkarna, Zaldivar, Kjartansson), "Data Cards: Purposeful and Transparent Dataset Documentation for Responsible AI" (2022). https://arxiv.org/pdf/2204.01075
- National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1" (2023). https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf
Tell us what your models need
Share scope, volume, language, format, timing and licensing requirements.