Deal economics
AI data contract dependency: what binds after the buyer can leave
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
AI data contract dependency is the gap between what a large contract obliges the buyer to do and what it forbids the seller to do. Find the buyer's earliest exit date, the end of its notice period or committed minimum, then list every exclusivity, capacity and supply restriction still binding the seller after it. Restrictions that outlast the buyer's commitment are freedom sold without payment; price them before treating the contract as strength.
License terms
AI data contract dependency
Key takeaways
- A contract's size says little about its value to the seller; the balance of obligations on each side says much more.
- The exit-date test: on the first day the buyer can leave without paying, list which seller restrictions still bind and for how long.
- Filings show the asymmetry is common: Innodata says its customer agreements are generally terminable on 30 to 90 days' notice and generally do not oblige future purchases.
- Eight clauses carry most of the dependency: exclusivity, first rights, most-favored pricing, dedicated capacity, service levels, reuse limits on derived work, termination terms, and assignment or change of control.
- In the illustrative contract, the restriction tail is 22 months against a 2-month commitment, an 11-to-1 ratio before counting work assigned away for good.
- A restrictive contract can still be the right deal when the buyer pays for the restriction with a committed floor, a wind-down payment or a premium that covers the lost options.
What is AI data contract dependency?
AI data contract dependency is the condition in which one buyer's contract limits what a data company can sell, to whom and with which staff for longer than the buyer is committed to pay. It is measured clause by clause, by placing the seller's restrictions and the buyer's obligations on one timeline.
Large contracts are not the problem. A large contract can fund capability, set a specification and signal quality to other buyers. The problem is that size shows up on a slide, while the obligations that decide a contract's value sit in clauses investors seldom read: exclusivity, service commitments and limits on future supply.
This article sets out a two-sided clause ledger for those terms, an exit-date test that reduces the ledger to one comparison, four contract shapes to sort deals into, and the questions to send before a large contract is credited as strength.
Why can the largest contract leave the least strategic freedom?
Because the seller's obligations usually outlast the buyer's. A buyer that can leave on short notice holds an option; a seller that hired staff, turned away competitors or promised first rights to new data has already paid for that option.
Filings from AI data service providers describe the buyer's side plainly. Innodata's 10-K for 2025 says customers can generally end their agreements on 30 to 90 days' notice and are generally under no obligation to buy services in future periods. The same filing says the company normally invests in people and technology and incurs other costs in anticipation of revenues. Read together, those statements describe the asymmetry: costs committed in advance, revenue cancellable on a quarter's notice or less.
TaskUs's 10-K for 2025 describes most client agreements as terminable for convenience on advance notice of varying length, and puts its largest client, Meta, at 26% of 2025 revenue.
When a buyer uses that option, the reason may have nothing to do with the supplier. Appen disclosed in early 2024 that Google would end its global inbound services contract as part of a strategic review process, with all projects ceasing by 19 March 2024. Appen then announced cost measures worth $13.5 million in annualized savings, most of them tied to delivering Google projects: capacity built for one buyer became a cost to remove.
The two-sided clause ledger: who each term binds
For each of eight clause types, record what it requires of the seller, what it requires of the buyer, and whether the seller's side still binds after the buyer's earliest exit. A clause binds one-sidedly when only the seller's column has content.
These are the terms that most often turn a large data or data-services contract into dependence. The wording below is generic; real clauses vary, and counsel should read the actual text before any conclusion is drawn.
The ledger, with the document that proves each row:
| Clause | What it asks of the seller | What it asks of the buyer | Binds after the buyer can exit? | Evidence to request |
|---|---|---|---|---|
| Exclusivity | No licensing of the same records, field or use to others for the term | Often nothing beyond paying for orders it places | Yes, when the term runs past the notice period | The clause, its defined field and term, and which affiliates it covers |
| First refusal or first negotiation | Offer new datasets or capacity to this buyer first | Usually nothing; the buyer may decline | Often yes, including any tail after termination | Clause text, response windows and tail period |
| Most-favored pricing | Match any better price or term given to others | Nothing | Yes, for as long as the clause lasts | Clause text, comparison basis and audit rights |
| Dedicated capacity | Hire, train and reserve staff or systems for this buyer | Sometimes a forecast, rarely a paid floor | The cost does: staff stay hired after orders stop | Staffing plan, forecast terms, who pays for idle capacity |
| Service levels | Meet turnaround, quality and security targets, with credits | Timely specifications and feedback, if written | Ends with the orders but sets cost while they last | Service schedule, credit history and change-request log |
| Limits on reuse of derived work | Do not reuse labels, schemas or tooling for other buyers | Nothing | Often yes; ownership terms survive termination | Work-product and assignment clauses, survival clause |
| Notice and termination | Keep delivering through the notice period | Give notice and pay for work done | This clause sets the buyer's exit date | Termination-for-convenience terms and any wind-down payment |
| Assignment and change of control | Get consent before selling the company or moving the contract | Nothing | Yes, when it gives the buyer a veto or exit on a sale | Assignment clause and change-of-control triggers |
The exit-date test: what still binds after the buyer can leave?
Find the earliest date on which the buyer can stop paying, then list every seller restriction that still binds on that date and for how many months. If any restriction outlasts the buyer's commitment, the seller has sold strategic freedom the buyer did not pay for.
The buyer's exit date is the later of the end of its notice period and the end of any committed minimum. A termination right does not have to leave the seller empty-handed. The federal fixed-price clause, FAR 52.249-2, lets the government end work when the contracting officer determines that termination is in the government's interest, then sets out a settlement built from costs incurred and a profit on work done, with no profit allowed where the contractor would have lost money on the whole contract. Check whether a private data contract pairs its termination right with any comparable wind-down payment; where it does not, the seller carries the stranded cost.
An illustrative contract makes the test concrete. The buyer orders a large labeling and licensing program, may terminate on 60 days' notice and commits to no minimum. The seller grants 24 months of field exclusivity and a right of first negotiation on new datasets that runs 12 months past termination, assigns every label and schema to the buyer, and hires a dedicated team. If the buyer gives notice on day one, it can leave in month 2; exclusivity binds the seller until month 24 and first negotiation until month 14, and the assigned work is gone for good.
Call the months between the buyer's exit date and the end of the last time-limited restriction the restriction tail. Here it is 22 months against a 2-month commitment, a ratio of 11 to 1, before counting the permanent loss of reusable work. That ratio is the figure an investor can compare across contracts.
The timeline of the illustrative contract:
| Term (Illustrative) | Starts | Ends | Months past the buyer's exit date |
|---|---|---|---|
| Buyer commitment: 60 days' notice, no minimum | Month 0 | Month 2 | 0, this is the exit date |
| Field exclusivity | Month 0 | Month 24 | 22 |
| First negotiation on new datasets | Month 0 | 12 months after termination, month 14 | 12 |
| Assignment of labels and schemas to the buyer | Month 0 | No end date | Permanent |
| Dedicated team hired for the program | Month 0 | Until redeployed or released | Depends on redeployment |
Exclusivity, capacity or supply limits: which costs the most later?
Supply restrictions usually cost the most over time, because they shrink the pool of future buyers; capacity commitments cost the most in cash, because staff are paid whether or not orders arrive; exclusivity falls between the two, depending on how narrowly its field is drawn.
Exclusivity is cheapest when it is narrow and reciprocal. Sales law for goods offers a benchmark: under UCC section 2-306(2), a lawful exclusive-dealing agreement carries, unless the parties agree otherwise, a duty on the seller to use best efforts to supply the goods and a duty on the buyer to use best efforts to promote their sale. Article 2 governs sales of goods, not data licenses, so the rule does not apply directly, but it frames the right question for any data deal: what does the buyer owe in return for exclusivity?
Capacity is where cost lands first. Appen's response to the Google termination, cost measures mostly tied to delivering Google projects, shows dedicated capacity turning into a cost to remove once the buyer leaves. Ask who pays for capacity reserved against a forecast the buyer is not bound to meet.
Supply restrictions work like switching costs in reverse. The Federal Trade Commission's January 2025 staff report on partnerships between cloud providers and AI developers says exclusivity terms and other provisions in those agreements may prevent the AI developer partner from adding resources from a different cloud provider, or raise switching costs. Staff listed switching costs among its areas to watch, and the report concerns a different layer of the AI stack, but the mechanism is the one to look for in a data company's largest contract: terms that make adding a second partner harder.
In the EU, some one-sided data terms may not bind at all. Article 13 of the Data Act, Regulation (EU) 2023/2854, provides that a contractual term on access to and use of data that one enterprise unilaterally imposed on another does not bind the other party if it is unfair. It presumes some terms unfair, including terms letting the imposing party use the other's data in a way that significantly harms its legitimate interests, or stopping the other party from obtaining a copy of its data. The rule reaches only terms imposed without real negotiation, and counsel decides whether it applies to a given contract.
Four contract shapes: partnership, free option, clean order, spot work
Sort every large contract with two questions: does the buyer carry a committed obligation, and does the seller carry restrictions that outlast it? The answers give four shapes, and only one of them is the dependence this article warns about.
A contract can move between shapes at renewal, so classify the version in force and the one being negotiated. The four shapes and how to read them:
| Shape | Buyer commitment | Seller restrictions past the exit date | Investor read |
|---|---|---|---|
| Partnership | Committed floor or long fixed term | Yes, but no longer than the floor | Can be strength: both sides are bound and the restriction is paid for |
| Free option | Notice period only, no floor | Yes, for months or years past the exit date | Dependence: the buyer holds an option the seller wrote for nothing |
| Clean order | Committed order or floor | None past the exit date | Best revenue quality per contract; check renewal history |
| Spot work | Order by order, cancellable | None | Weak revenue quality but low dependence; the seller keeps its freedom |
When is a restrictive contract still worth signing?
When the buyer pays for the restriction in a form that survives its own exit: a committed floor at least as long as the restriction, a wind-down or settlement payment, or a premium large enough to cover the buyers the seller turns away. A famous name is not payment.
Close ties to one buyer can also change how other buyers see the seller. After Meta's June 2025 investment left it holding a minority of Scale AI's equity, Scale published a post titled Scale Remains Independent After Meta Investment, saying it was not integrating its business operations with Meta and that Meta's commercial relationship would carry the same protections and restrictions as any other customer's. The post shows a tie to one buyer becoming a question every other buyer asks.
Large announced budgets deserve the same discipline. When micro1 announced in October 2026 that it was committing $1 billion over 12 months to acquire and license operational data from enterprises, it published a stated plan. It tells no supplier what a contract with micro1 would commit, restrict or pay, and the exit-date test applies to whatever paper follows.
Even medium-term licensing deals carry renewal risk. Reddit's Form 10-Q for the quarter ended June 30, 2026 traces substantially all of its licensing contract value to two partners, and adds that those agreements might lapse unrenewed or be renewed on less favorable terms.
Risks that would make the exit-date test the wrong lens
The test simplifies on purpose, and five things can make it mislead:
- Unwritten commitment. Buyers sometimes keep buying long after their notice period because switching suppliers costs them too. The test counts written commitment only; renewal history can justify more credit, but it should be shown rather than asserted.
- Restrictions that cost nothing. Exclusivity in a field with no other buyers has little opportunity cost. Price each restriction against named alternative buyers, not in the abstract.
- Enforceability. Some restrictions may be unenforceable, or in the EU may not bind if unilaterally imposed and unfair under Article 13 of the Data Act. Record the clause as written, flag doubts and leave the conclusion to counsel.
- Missing paper. Side letters, statements of work and purchase-order terms can add or remove restrictions. Ask for every document in the contract stack, not just the master agreement.
- Concentration still counts. A clean order from a buyer that supplies most of the revenue is still concentrated exposure. Pair this test with a measure of buyer concentration.
Clause questions for the data room, SourceX included
Request the full contract stack for the three largest relationships and rebuild the ledger and the timeline yourself.
For SourceX, the published starting point is its standing terms: each deal's signed agreement sets its exclusivity and term, and that agreement is signed, on a scope and terms the supplying company has approved, before any data is delivered. Its investor materials, for a Rule 506(c) raise from accredited investors, can be requested from its investors page. Its contract terms with buyers are not public and none are described here, so these are the questions to send.
Questions that rebuild the test:
- For each of the three largest relationships, what is the buyer's earliest exit date, counting notice and any committed minimum?
- Which exclusivity, first-rights, pricing and reuse restrictions bind the company after that date, and for how long?
- Who owns the labels, schemas, tooling and methods built during each program, and what survives termination?
- How many people or systems are reserved for each buyer, and who pays if its forecast is not met?
- Does any contract include a wind-down or settlement payment on termination for convenience?
- Does any clause give a buyer consent rights or an exit on a change of control?
- Which other buyers has the company declined, or not approached, because of a restriction?
Method behind the clause ledger and exit-date test
The clause ledger, the exit-date test, the restriction-tail ratio and the four shapes are editorial frameworks written for this article on 2026-10-26. They draw on clause types common in data and data-services contracts and on the cited documents: Innodata's and TaskUs's 2025 Forms 10-K, Appen's February 2024 ASX announcement, Reddit's Form 10-Q for the quarter ended June 30, 2026, FAR 52.249-2, UCC section 2-306, the FTC's January 2025 staff report on AI partnerships, Article 13 of the EU Data Act, and Scale AI's and micro1's own announcements.
The illustrative contract and its timeline are invented. Not measured: how common each clause is in AI data contracts, what restrictions cost in lost sales, how often buyers stay past their notice period, and whether any particular clause is enforceable. Nothing here describes SourceX's contracts or any company's terms beyond what the cited documents state.
Frequently asked questions
What makes an AI data contract one-sided?
A contract is one-sided when the seller's obligations outlast the buyer's. Typical signs are exclusivity, first rights or reuse limits that keep binding the seller after the buyer can terminate on short notice. Compare the end date of each seller restriction with the buyer's earliest exit date to see the gap.
Is exclusivity in an AI data contract always bad for the seller?
No. Exclusivity that is narrow in field, limited in time and matched by a committed floor or a premium can be a fair trade. It becomes dependence when it runs well past the buyer's commitment and blocks buyers the seller could otherwise serve.
What is a termination for convenience clause in a data services contract?
It lets the buyer end work without cause, usually after a notice period. Innodata, for example, says its customer agreements are generally terminable on 30 to 90 days' notice. Federal contracts pair this right with a settlement covering costs and profit on work done, which is worth asking for in private deals.
How should investors value a large data contract with a short notice period?
Count committed revenue only up to the buyer's exit date and treat the rest as probable, supported by renewal history. Then subtract the cost of any restriction that outlasts the commitment, priced against buyers the seller cannot serve while it binds.
Do unfair data contract terms bind companies in the EU?
Under Article 13 of the EU Data Act, a term on access to and use of data that one enterprise unilaterally imposed on another does not bind the other party if it is unfair, and some terms are presumed unfair. The rule does not cover terms that were genuinely negotiated, and counsel should assess any specific contract.
Sources
- Innodata's 2025 Form 10-K says customer agreements are generally terminable on 30 to 90 days' notice, generally do not oblige future purchases, and that it invests in people and technology in anticipation of revenues. Source
- TaskUs's 2025 Form 10-K describes most client agreements as terminable for convenience on notice and puts Meta at 26% of 2025 revenue. Source
- Appen disclosed Google's termination of its global inbound services contract after a strategic review, with projects ceasing by 19 March 2024, and $13.5 million of annualized cost savings mostly tied to Google projects. Source
- FAR 52.249-2 lets the government terminate for convenience when in its interest and provides a settlement based on incurred costs and profit, with no profit where the contractor would have lost money. Source
- UCC 2-306(2): a lawful exclusive-dealing agreement imposes, unless otherwise agreed, best-efforts duties on the seller to supply and on the buyer to promote sale. Source
- The FTC's January 2025 staff report says exclusivity terms and other provisions in cloud-AI developer partnerships may prevent adding another cloud provider or raise switching costs. Source
- EU Data Act Article 13: unfair data access and use terms unilaterally imposed by one enterprise on another are not binding; some terms are presumed unfair. Source
- Scale AI's June 2025 announcement says Meta's investment leaves Meta holding a minority of Scale's outstanding equity. Source
- Scale's post Scale Remains Independent After Meta Investment says it is not integrating its business operations with Meta and applies the same customer protections and restrictions to Meta. Source
- micro1 announced a commitment to spend $1 billion over 12 months acquiring and licensing enterprise operational data (a stated plan). Source
- Reddit's Form 10-Q for the quarter ended June 30, 2026 says substantially all licensing contract value comes from two partners and warns the agreements may not be renewed or may be renewed on less favorable terms. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.