Remedies When a Data Delivery Fails: Re-delivery, Replacement Records, Credits and Refunds

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Why remedies fail without a measurement behind them

A remedy clause is only enforceable in practice when the contract says how the triggering defect is measured. Dataset SLAs that work state, for each term, the metric, the measurement method and the remedy that follows a miss [1]. "Data will be of high quality" gives you nothing to invoke; "label accuracy of at least 95% on a 400-item stratified audit against a buyer-held gold set" does.

ISO/IEC 5259-2 gives a vocabulary of measurable data quality characteristics, such as completeness, accuracy and consistency, that you can borrow for clause definitions [2]. ISO/IEC 5259-4 frames the quality process for training and evaluation data, including labelling for supervised learning, which helps you agree on where in the pipeline each check happens [3]. Define the terms in your acceptance criteria for licensed training data first, then attach remedies to them here.

Classify the defect before choosing the remedy

Most disputes come from treating every failure the same way, so sort defects into five classes with different cures. A corrupt Parquet file and a withdrawn consent are both "bad data", but one is fixed by re-sending bytes and the other may require deletion and a price adjustment.

• Format and integrity defects: checksum mismatch, a Parquet file missing its trailing PAR1 magic bytes or footer metadata [8], JSON Lines with a byte order mark or blank lines [9], schema drift from the agreed data dictionary.
• Completeness and coverage defects: record counts short of the order, missing date ranges, empty required fields, class or segment quotas unmet.
• Accuracy and label defects: annotation error rate above threshold, inconsistent taxonomy use, wrong span boundaries, failed gold or honeypot items [5].
• Privacy and preparation defects: residual names, emails, phone or account numbers after de-identification, undisclosed duplicates, or test-set leakage.
• Rights defects: records outside the licensed scope, consent withdrawn after delivery, or a third-party claim against the content.

Timeliness sits across all five: a late delivery of perfect data is still a defect if your training run is scheduled around it.

The remedy ladder, from cheapest cure to exit

Remedies should escalate from the supplier's cheapest cure to your right to leave, with each rung triggered by a defined event. The ladder below is a starting structure; your counsel will set the exact language.

1. Correction in place. The supplier fixes the affected fields or files within a cure window, and you re-run the acceptance tests.
2. Re-annotation. For label defects, the supplier re-labels the failed batch at its own cost; re-annotation obligations are common in annotation contracts [1].
3. Replacement records. Rejected rows are swapped for new conforming rows, usually up to a stated percentage of the order [7].
4. Full re-delivery. Used when the defect is systemic (wrong schema, wrong export, corrupt archive), so a partial fix would leave you with mixed versions.
5. Service credits. A percentage of the delivery fee, credited against the next invoice, for lateness or missed SLA metrics that do not justify rejection.
6. Fee reduction or refund. A pro-rata reduction for permanently missing records, or a refund of the milestone if the batch fails acceptance after the allowed cures [7].
7. Termination. For repeated failure, an uncured material defect, or a rights defect that undermines the license.

Illustrative remedy schedule for a licensed dataset order

Illustrative example: invented to show structure; it does not describe an available dataset.

The numbers are placeholders. Set yours from the variance you saw in the paid pilot or sample, not from round numbers that sound strict.

How to set thresholds that survive a dispute

Thresholds hold up when both sides agreed on the sampling plan, the gold set and who runs the audit before the first delivery. Buyer guides suggest fixing completeness, validity and duplicate-rate limits and a replacement percentage before the order is placed [7]. Changing them after a disappointing batch reads as renegotiation.

For recurring deliveries, acceptance sampling borrowed from manufacturing helps. ASQ/ANSI Z1.4 defines normal, tightened and reduced inspection plans with switching rules for a continuing stream of lots at a specified acceptable quality level [4]. Applying it to label audits is an adaptation, not a data standard, but its switching idea (the standard moves from normal to tightened inspection when two of five consecutive lots are rejected) is a useful pattern to write in.

Agree on three operational details in writing:

• Gold set ownership. If the supplier builds the gold items it audits against, the audit tests consistency, not accuracy. Annotation platforms support gold (honeypot) items and consensus checks [5]; keep the gold set buyer-held.
• Who measures, and when. Specify the acceptance window (for example, 10 business days from receipt) and state whether silence after the window counts as acceptance, since deemed-acceptance wording favors the supplier.
• Tie-break. Name a joint re-audit or a neutral reviewer for disagreements over a sample, so a dispute does not stall the whole supply.

Late delivery remedies and service credits

Service credits suit lateness because the data may still be usable, just not on schedule. Express credits as a percentage of the delivery fee per period late, cap them (often at the fee for that delivery), and say whether they are your sole remedy for lateness. A "sole and exclusive remedy" phrase can quietly remove your right to terminate for chronic delays, so pair it with a termination trigger after a stated number of missed dates.

Credits only bite if delivery has a clear timestamp. Define "delivered" as the moment conforming files land in the agreed location with a valid manifest, not when the supplier sends an email saying an export started. Our guides to incremental vs full refresh deliveries and ongoing data supply agreements cover how delivery events are recorded across refresh cycles.

Rights defects are not quality defects

A rights defect, such as a record collected without the consent the license assumes, cannot be cured by re-annotation; it needs removal, replacement and protection against third-party claims. Practitioner guidance argues that a training data license needs an explicit grant plus a warranty package and indemnity, not a one-line grant [6]. Remedies here typically include:

• a deletion or quarantine procedure for affected records, including in derived training sets where your counsel considers it feasible;
• replacement records from within the licensed scope;
• an indemnity for claims arising from the supplier's rights warranties, with its own cap separate from quality remedies;
• a termination right if the defect affects a material share of the dataset.

Decide in advance what happens to models already trained on the affected records. Retraining cost can dwarf the data fee, so whether that is recoverable is a negotiating point, not a default.

Privacy defects need a fast lane

Residual personal data is the one quality defect where the remedy clock should be hours or days, not a normal cure window. Automated de-identification tools reduce but do not eliminate risk; Microsoft's Presidio project itself cautions that it may not find all sensitive information [10]. Write in an immediate notification duty, a stop-use instruction for the affected files, re-processing and re-delivery, and a written confirmation that superseded copies were deleted on both sides.

Caps, exclusivity of remedies and payment structure

Remedies are only as strong as the money still at stake, so structure payment to keep leverage until acceptance. Paying in milestones tied to acceptance means a failed batch is a withheld payment, not a refund you have to chase. Watch for three common drafting traps:

• Aggregate caps that include credits. If credits count toward a low liability cap, a few late deliveries can exhaust it before a rights claim arrives.
• Exclusive remedy language. It can turn re-delivery into your only option, even after the third failed attempt.
• No cure limit. Without "two cure attempts per batch", a supplier can keep re-delivering defective data indefinitely.

When you compare vendor quotes on cost per usable record, price in the replacement allowance: a 5% replacement right on a cheaper quote may beat a pricier quote with none. For how this fits the full contract, see the AI training data procurement hub, the data supplier SLAs guide and what happens if data contains errors.

Where SourceX fits

SourceX sources operational datasets from US companies on request and manages the commercial process, including licensing and ongoing purchases. Every dataset is rights-reviewed for ownership and consents and delivered under a license that defines the records, uses, term and delivery, and pricing and allowed uses are agreed per deal. Personal details are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect. If you are scoping remedies for a dataset you have not yet sourced, you can describe the data you need to SourceX.

Get licensed data with defined delivery terms

SourceX sources datasets on request from US businesses, and nothing is contracted until a supplier agrees to a license covering records, uses, term and delivery. Delivery runs through private, access-controlled workflows only after an executed agreement and supplier approval. Tell SourceX what data you need.