Obligations, Risk Terms, and Diligence at Volume
Obligation Extraction Across a Portfolio
A recurring commercial question is simple to ask and painful to answer: across our four hundred active contracts, what are we actually obliged to do, by when, and to whom? Answering it manually is a project. Extraction models handle it reasonably well because obligations have recognisable linguistic form — a party, a modal verb, an action, often a deadline. The output is best treated as a structured register with a link to the source clause for each entry, which makes spot-checking practical and lets the register be maintained rather than rebuilt. The known weaknesses are conditional obligations that only trigger on some event, obligations created by reference to an incorporated schedule, and duties expressed as a negative restriction rather than a positive act.
- Obligations have recognisable form, which makes them a good extraction target across a portfolio
- Output should be a structured register with a source link per entry, not prose
- Weak spots: conditional triggers, incorporated schedules, and negatively-framed restrictions
- Spot-check by sampling clauses the register does not mention, not only entries it produced
Risk Terms and Change-of-Control Sweeps
Transactional work generates repeated sweeps for the same categories: change-of-control provisions, assignment restrictions, exclusivity, most-favoured-nation clauses, unusual termination rights, uncapped indemnities. These are well-defined targets and a model can surface candidates across a large set far faster than sequential human reading. The correct posture is recall-oriented: configure for over-inclusion, accept the false positives, and use human review to discard them. That is the opposite of how these tools are often demonstrated, where a clean short list looks impressive. A short list is impressive precisely because something was filtered out, and on a diligence sweep the thing filtered out is what you were looking for.
- Configure sweeps for recall and accept false positives — a clean short list means something was dropped
- Well-defined targets work best: change of control, assignment, exclusivity, uncapped indemnity
- Impressive-looking precision in a demo is usually recall being traded away invisibly
- Record the categories swept and the criteria used — the sweep definition is part of the work product
Diligence at Volume Without Losing the Thread
Volume is where these tools earn their place, and also where the review protocol matters most. A diligence exercise across thousands of documents cannot be fully read by anyone, which was true before AI and is why sampling and prioritisation were always part of the method. What changes is that the prioritisation is now performed by a system whose criteria are opaque and whose errors are unpatterned. The professional obligation is unchanged: the practitioner must be able to explain what was reviewed, how, and with what limitations. That means documenting the tool, the configuration, the sampling approach, and the known gaps — and being able to state those limitations to a client or a tribunal without having to reconstruct them afterwards.
- Prioritisation was always part of large-scale diligence; what changed is who performs it and how visibly
- Document tool, configuration, sampling method, and known gaps as you go, not retrospectively
- You must be able to explain the method and its limits to a client or a court in plain terms
- Opaque criteria are acceptable only if the resulting coverage is measured and disclosed
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.