What AI Actually Reads When It Flags an Indemnity Clause

Indemnification clauses are among the most negotiated provisions in any commercial agreement. They are also among the most misread, not because attorneys lack skill, but because the risk lives in structural relationships between phrases, not in the presence or absence of single words.

When we built Clausebeam's indemnity detection, the central question was not "does this clause contain indemnity language?" That part is straightforward. The harder question was: "what kind of indemnity obligation does this clause actually create, and who bears disproportionate risk under it?"

This post walks through how our clause-level model approaches that second question. It is not a product demo. It is an explanation of what we found when we tried to teach a model to distinguish broad from narrow indemnity, and why word order matters more than most people expect.

The anatomy of an indemnity clause

Any indemnification provision has three structural components that determine its practical risk profile: the triggering event (what must happen before indemnity obligations attach), the scope of covered losses (what the indemnitor must actually pay), and the carve-outs or limitations (what is excluded from coverage).

A clause that reads "Party A shall indemnify Party B from and against any and all losses, claims, damages, costs, and expenses arising out of or relating to Party A's breach of this Agreement" looks, at a glance, like a standard mutual indemnity. It is not. The triggering event is Party A's breach specifically, but "arising out of or relating to" is an extremely broad causal connector. Courts have repeatedly held that "relating to" extends indemnity well beyond the narrow but-for causation standard. The clause sounds bilateral but loads most risk onto one party.

Compare that to "Party A shall indemnify Party B from direct losses finally determined by a court of competent jurisdiction to have resulted solely from Party A's gross negligence or willful misconduct." Here the triggering event is narrower (gross negligence or willful misconduct, not mere breach), the causal standard is tighter ("resulted solely from"), and coverage is limited to direct losses as determined by a court. That is a meaningfully different risk allocation, even though both clauses nominally impose indemnity on Party A.

Our model reads those structural differences. It does not simply identify "indemnity language present." It maps the triggering condition, identifies the causal standard, and notes whether the loss scope is limited or broad.

Why word order changes the risk category

Consider two formulations:

Both versions say the same thing, roughly. But Version A includes "defend" as an explicit obligation alongside indemnify and hold harmless. That is not a stylistic difference. The duty to defend in most U.S. jurisdictions is broader than the duty to indemnify: it can be triggered by a mere allegation, before any determination of liability. Adding "defend" means the indemnitor may have to fund defense costs immediately, regardless of whether the underlying claim succeeds. Version A is substantially riskier for the indemnitor than Version B, and that risk signal lives in a single word in the middle of a list.

We encountered this pattern repeatedly in our review data. High-risk signals are often not in the operative verb ("indemnify") but in adjacent modifiers, conjunctions, and scope qualifiers. Our model flags the presence of "defend" in indemnity triads as a separate risk marker distinct from broad indemnity scope, because it implicates different legal obligations and different risk exposure windows.

The mutual vs. one-way asymmetry problem

One of the most common findings in our early review data was nominally mutual indemnity clauses that were functionally one-way. The structure: both parties have indemnity obligations in the same clause, but Party A's obligation is triggered by "any breach or alleged breach" while Party B's obligation is limited to "gross negligence or willful misconduct." In practice, Party A bears nearly all the indemnity risk, but the clause reads as if it is bilateral.

In a deal room review we ran last year for a mid-size technology company's vendor contract stack, seventeen of the company's thirty-one active vendor agreements contained this structure. The vendor's paper uniformly included "alleged breach" in the company's indemnity obligation and "gross negligence" in the vendor's. Not one attorney had flagged it as systematically asymmetric, because each agreement had been reviewed in isolation. When you read them together, the pattern is obvious. Clausebeam surfaces that cross-document pattern in a single read.

We want to be direct about what we are not saying here: we are not saying that one-way or vendor-favorable indemnity is always a problem. In some categories of vendor relationship, asymmetric indemnity is market-standard and reflects real differences in risk exposure between the parties. What we are saying is that you should know when you have it, so you can decide whether it is intentional and appropriate or something that slipped through.

Carve-outs and their placement

Indemnity carve-outs are the provisions that limit what the indemnitor must cover. Common carve-outs: losses caused by the indemnitee's own negligence, losses arising from the indemnitee's breach of its obligations, and consequential or punitive damages.

Where these carve-outs appear in the clause structure matters. A carve-out at the end of a long indemnity provision, following extensive scope language, can be read narrowly by courts that apply the principle that exceptions to broad obligations are construed against the party asserting them. A carve-out that explicitly says "notwithstanding the foregoing" and then specifies the exclusion with equal precision to the main obligation is more reliable.

Our model does not just detect the presence of carve-outs. It assesses their structural position relative to the scope language and flags carve-outs that appear likely to be read narrowly given their placement. That is a more useful output than "carve-out detected."

What the model does not do well yet

We want to be honest about the limits. Cross-reference analysis is genuinely hard. An indemnity clause in Section 12 that says "subject to the limitations in Section 15" requires reading Section 15 to understand what the indemnity actually covers. Our model handles common cross-references, but unusual or deeply nested cross-references still require attorney review. We flag those with a "cross-reference dependency" marker rather than providing a definitive risk assessment.

Jurisdiction-specific interpretation is another limitation. "Arising out of or relating to" in California commercial contracts has a different case law backdrop than in New York. Our risk flags are calibrated to general U.S. commercial standards. We do not yet provide jurisdiction-specific deviation scoring. That is on the roadmap.

And there is a category of indemnity risk that lives not in the language itself but in the factual context of the deal. An indemnity that looks standard for a services agreement may be extremely aggressive for a software license where consequential damages represent virtually all foreseeable loss. Context-dependent risk assessment requires attorney judgment. Clausebeam surfaces the clause; it does not tell you whether the risk is acceptable given your specific deal dynamics.

Reading indemnity clauses in volume

The practical use case that motivated this work was not any single high-stakes negotiation. It was the accumulation of vendor paper that every in-house legal team faces. NDAs, MSAs, service agreements, SaaS subscriptions: a mid-size company might receive fifty or more contracts per quarter, each with its own indemnity clause, each drafted on the counterparty's paper, each slightly different from the last.

Reading each of those clauses carefully enough to assess the triggering conditions, scope, and carve-outs takes time that most in-house teams do not have. The result, in practice, is that clauses get skimmed. The obvious things get caught. The structural problems, like the asymmetric mutual/one-way pattern or the presence of a broad "defend" obligation, get missed because they require more than a fast read.

Clausebeam was built to do that careful read on every clause, every time. Not to replace attorney judgment about what to do with what it finds, but to ensure that nothing structurally significant gets missed because a team is moving too fast through too much paper.