Continuity and escrow when a data supplier stops supplying
You already hold the data. What you lose is the records, the recipe and the ability to prove where it came from. What to escrow, and what is cheaper than escrow.
What actually breaks when a supplier fails
Not the dataset. It sits on your storage, and a vendor's insolvency does not erase it. What breaks is everything around it.
- The records that prove provenance. If the consent forms, the roster and the provenance log live only with the vendor, you lose the ability to answer an audit, a regulator or a customer's diligence request about material you are still training on.
- The ability to extend. A refresh needs the guideline, the screening criteria and the recruitment channel, and those belong to the vendor, not to you.
- The ability to re-annotate. Revising an existing corpus requires the original conventions, the tooling settings and annotators trained on them.
- Support for anything they built. If your pipeline depends on a converter, a validation script or a labelling tool the vendor wrote, you lose maintenance along with the vendor.
Escrow the records first, the code second
The instinct is to escrow source code. For data supply the priority runs the other way: the records package is the asset you cannot recreate, and the software is usually the smaller problem.
Define the records package explicitly: signed speaker forms in the version that applied to each batch, the roster with dates and locations, the consent scope matrix, the provenance log, the guideline revisions, the quality control records, and the deletion confirmations. Deliver it with every delivery rather than depositing it once — the cheapest continuity measure in this whole area is that you already hold the records for what you have already received.
Escrow then covers the gap: the material for sessions not yet delivered, plus whatever records the vendor generates between deliveries.
Release conditions, and the one people forget
The standard triggers are insolvency, administration, cessation of business, and failure to deliver after notice and a cure period. Add two more that matter in this market: a change of control that puts the vendor inside a competitor, and a refusal to continue supplying the refresh line covered by the agreement.
That last one is the trigger buyers forget. A vendor can stay solvent, keep trading and simply decline to produce the next version, and a release condition that only responds to insolvency leaves you with an escrow you cannot open and no remedy.
Define who declares the release and on what evidence. A trigger that requires the buyer to prove insolvency in a foreign court is not a trigger. Name events that can be verified from public records or from a written admission, and set a short clock for the agent to act once the condition is met.
An unverified escrow, and five cheaper alternatives
Deposits are made once and never examined, and by the time they are needed they contain the wrong version, an incomplete export, or files in a format nobody can read. The fix is a verification right: on request, and at least once during the term, the buyer may inspect the deposit's manifest and confirm the contents and the format.
Go further if the deal justifies it and run a release drill. Request a copy under a temporary release, restore it in a test environment, and confirm the records are complete and the tooling works. A drill that succeeds in peacetime is the only evidence the escrow functions.
Choose the agent carefully. A deposit held by the vendor is not an escrow, it is a promise, and promises are exactly what fail at the moment they are needed. A deposit held by the vendor's law firm is better, but not independent if that firm also advises the vendor on the dispute that triggers the release.
None of the following replaces an escrow, but each one removes part of the risk at a fraction of the effort, and together they cover most of what actually goes wrong.
- Records at delivery. If the consent forms and the provenance log are delivery items from day one, the escrow only has to cover what has not yet arrived.
- Versioned documents in your own hands. Keep the specification and the guideline revisions you were given under your own version control. When a vendor disappears, the specification you wrote is the starting point for a re-tender, and reconstructing it from memory costs weeks.
- A licence on failure. For tooling, a clause granting you the right to use, maintain and modify the vendor's software if they cease to support it is often more practical than depositing code, because it removes the release process entirely.
- A key-person clause with a named deputy. Recruitment channels and language expertise often sit with one individual. Requiring the vendor to document the channel and name a successor costs nothing and covers a risk escrow does not reach.
- A second source, priced in. Two suppliers for the same data type is the real continuity answer, and its cost should be compared against the cost of a stalled programme rather than against zero.
The continuity schedule to attach
These items turn continuity from a hope into a clause. Most of them are cheap at signature and unavailable afterwards.
- The records package, as a delivery item at every delivery.
- The escrow contents, defined by category, with a format specification.
- Release conditions, including change of control and refusal to refresh.
- Who declares the release, and on what evidence.
- Verification rights, and a release drill once during the term.
- The escrow agent, and limits on who may act in that role.
- A licence on failure for tooling, and a data-return obligation.
- A key-person clause with a documented deputy.
- A flow-down obligation reaching any field partner who holds the original records.
The test to apply
Imagine the vendor stops answering email tomorrow. Ask which of your obligations you could still meet: a customer diligence request, a regulator inquiry, a refresh for the next model release, a re-annotation after a guideline change. Whatever you could not do is the gap, and it is worth closing the cheapest way available rather than the most elaborate way.
This is an operational outline rather than legal advice. Escrow, release and licensing-on-failure provisions are drafted differently in different jurisdictions, and counsel should review the wording before it is signed.