The liability cap is three fields, and only one of them is a number
In short
A carve-out is a category of liability the cap does not apply to — commonly fraud, breach of confidentiality, or an indemnity — so recovery for it is limited by what a tribunal awards rather than by the clause. A liability provision is therefore 3 separable facts: the cap amount, the basis that computes it, and the list that escapes it.
Key takeaways
- Cap amount, cap basis and carve-out list are 3 fields. A schema with 1 number cannot express the clause.
- Fees paid, fees payable and a fixed sum give 3 different answers from the same sentence.
- A carve-out removes a category from the ceiling entirely, so the honest summary is a cap plus a list.
- A super cap is a second, higher ceiling for one named category — not a carve-out, and stored as its own row.
- Excluding indirect loss is a separate mechanism from capping direct loss, and the record needs both fields.
A carve-out from a limitation of liability cap is a category of obligation the cap does not apply to. Whatever ceiling the clause sets, carved-out categories sit above it, and recovery for them is bounded by what a tribunal will award rather than by anything in the contract. So a record holding cap equals 100,000 has captured something true about part of the agreement and misleading about the part that matters most.
Read as data, the provision is 3 separable facts sharing a paragraph: the amount, the basis that computes the amount, and the list that escapes both. Store 1 and the other 2 go invisible to every query, dashboard and approval rule that runs on the record afterwards.
The amount, the thing that computes it, and the list that escapes it
- The cap amount. Sometimes a stated figure, more often the output of a formula the clause describes rather than states.
- The cap basis. Fees paid, fees payable, a fixed sum or a multiple of a period's charges — an enumerated field with a window attached, not free text.
- The carve-out list. Each entry pointing back to the clause creating the obligation, because a carve-out is a cross-reference and means nothing alone.
- Directionality, the field most often lost. A cap can be mutual or 1-way, identical words carry opposite risk by side, and on a counterparty's template the asymmetry is deliberate — part of the rebuilding job reviewing on the other side's paper sets out.
Fees paid and fees payable are not the same number
The basis is where most of the negotiation lives and where an extractor most often guesses. Four bases cover nearly everything: charges paid in a stated period before the claim, charges payable across the whole term, a fixed sum, or a multiple of an annual figure. On a 3-year subscription cancelled in month 4, paid and payable differ by roughly an order of magnitude.
The window is a separate value again: 12 months before the event giving rise to the claim, 12 months before notice of it, and 12 months before termination are 3 different windows, and a claim 2 years into a declining account lands on a different figure under each. That window usually sits in a defined term rather than in the sentence you extracted, which is why the resolution pass in defined terms and where their meanings move has to run first.
A cap of 50,000 with an unbounded list underneath it
Take annual charges of 50,000, a cap set at charges paid in the preceding 12 months, and 4 carve-outs: the intellectual property indemnity, breach of confidentiality, fraud, and the customer's duty to pay. The cap field reads 50,000, and any threshold rule keyed on that number calls the clause acceptable.
The exposure is not 50,000. An infringement claim runs to whatever the third party recovers plus the cost of defending it, and the confidentiality carve-out puts the whole value of the disclosed information above the ceiling. The honest summary is a sentence, not a figure: capped at about 50,000, except on intellectual property, confidentiality and fraud, where it is unbounded.
The cap says how bad an ordinary day is. The carve-outs say how bad the worst day gets.
Six carve-out families, and what each is really doing
| Family | What the clause typically does | What the record must store |
|---|---|---|
| Fraud and fraudulent misrepresentation | Excluded, often because many legal systems will not enforce a limit on it anyway | Whether it also reaches wilful misconduct and gross negligence |
| Death or personal injury from negligence | Excluded, and in several jurisdictions cannot be limited by agreement at all | Presence plus governing law, which decides enforceability |
| Payment obligations | The duty to pay charges is lifted out so a supplier is not capped out of its own invoice | Which party it protects — usually the 1 carve-out running one way |
| Indemnities | Intellectual property and third-party claims carved out, or given their own ceiling | The indemnifying party and the triggering events |
| Breach of confidentiality | Excluded entirely, narrowed to deliberate breach, or dropped to a super cap | Which of those 3 variants applies |
| Data protection and security incidents | Increasingly a separate higher ceiling rather than a true carve-out | Super-cap amount and basis, as their own row |
The indemnity row extracts badly most often: a system can be confident an indemnity exists and still record the wrong obligor, the failure taken apart in the indemnity extraction returns only half the obligation. A carve-out pointing at an indemnity facing the wrong way is worse than no carve-out — it flags risk on the wrong side of the deal.
The super cap and the exclusion of indirect loss are different instruments
A super cap is a second ceiling for a named subset, commonly data protection, set at a multiple of the general cap or a fixed sum. Liability under it is still limited, so treating it as unbounded overstates risk as badly as missing a real carve-out understates it. In templates that branch on deal size it is usually the switched element, which puts it with the rest of the conditional logic inside a legal template.
The exclusion of indirect and consequential loss works on another axis: the cap limits how much is recoverable, while the exclusion removes whole heads of loss — lost profit, lost revenue, lost data, wasted expenditure, goodwill — before the cap is reached at all. Watch the carve-back too. An exclusion that expressly preserves direct loss of profit is a materially different clause, and the difference is 1 subordinate clause long.
The fields a liability record has to carry
| Field | Shape | Why it exists |
|---|---|---|
| cap_amount | number or null | Null is legitimate while charges are unknown |
| cap_basis | enum: fees_paid, fees_payable, fixed_sum, multiple_of_period | Makes clauses comparable across a portfolio |
| basis_window | period plus the event it runs from | 12 months from claim and 12 from termination differ |
| mutuality | mutual, supplier_only, customer_only | The same text is an opposite position by side |
| super_caps[] | category, amount, basis | Keeps a bounded ceiling out of the carve-out list |
| carve_outs[] | category, clause reference, direction | Without the reference it cannot be verified |
| excluded_heads[] | heads of loss, plus any carve-back | The exclusion sits in a different sentence |
| source_citation | clause number and page | Checkable by whoever approves the deal |
What a complete liability record still will not decide
It does not say whether the clause is acceptable — that is a judgement against positions the firm holds, and where nobody wrote them down there is nothing to compare against, the elicitation problem in turning negotiation instinct into a scoring rubric. It does not say what to propose instead, which is the job of the fallback ladders inside a negotiation playbook. And it does not survive being read 1 file at a time: an order form capping liability over an agreement that caps it differently is a precedence question first, settled in which document wins when the master agreement and the order disagree.
All of it belongs to the contract review and redlining topic. The extraction and review systems we build for legal teams are scoped under AI agents and automation, against the operating constraints in AI agents in production.
Frequently asked questions
Short answers to the follow-ups this page tends to raise.
What does uncapped liability mean in a contract?
No contractual ceiling applies to that category of claim, so recovery is bounded only by what a claimant proves and a tribunal awards. Uncapped rarely describes a whole agreement — it describes the carved-out categories inside an otherwise capped clause.
Is a super cap the same thing as a carve-out?
No, and conflating them misstates risk in both directions. A carve-out removes a category from any contractual limit; a super cap gives that category its own higher limit. Data protection is the common case: often described loosely as carved out when the clause sets a separate ceiling at a multiple of the general cap.
Can a contract exclude liability for everything?
Rarely, and never uniformly across jurisdictions. Legal systems commonly refuse to enforce an exclusion for fraud, and many restrict limiting liability for death or personal injury caused by negligence, with further controls where terms are standard or the buyer is a consumer. Enforceability is decided by the governing law, so record that law beside the cap and have a lawyer confirm the position.
How should extraction handle a cap it cannot compute?
Record the basis and leave the amount empty rather than inventing a figure. A cap set at charges paid in the preceding 12 months has no amount until the charges are known, and a pipeline that fills the field with the contract value produces a record that looks complete and is wrong.
- liability
- clause extraction
- contract review
- data model
The work behind this page
Builds from our portfolio that this page draws on.
Brief Forge
Contract review AI for solo lawyers and small firms — extract, score, and redline contracts in minutes.
Legal TechAskVault
An AI internal knowledge-search platform that answers employee questions from your own docs — grounded in citations, with knowledge gaps surfaced and deflection tracked.
Productivity AIRead next
- Fallback positions: preferred, acceptable and walk-away, in one ladderA playbook that records only your ideal clause can say the counterparty deviated. It cannot say what to send back, which is the part a negotiator needs.definition
- Defined terms: no clause can be judged without its definitionA capitalised term carries whatever meaning the contract assigns it, wherever that assignment happens to live. Judge the clause without resolving it and you have judged a sentence you have not read.definition
- Order of precedence: the clause that decides which document winsA precedence clause ranks a stack of documents. It is also why a single file is an incomplete unit of review: the same words bind or not, depending on which instrument controls.definition
- Third-party paper: why their template is a different job from your ownOn your own template, review is a diff against known positions in known places. On the counterparty's, you have to rebuild the map before you can judge a single clause.definition
Related across the site
Working on something in this space?
Tell us where you are in a sentence or two. We'll tell you honestly whether we're the right team, and what a sensible first slice of the work looks like.
Start the conversation