Three Of These Thirty Say “Monitoring”. The Knowledge Stops In June 2026.
Six coloured agents, five workflows each, thirty boxes. The subtitle promises “30 legal workflows across 10 Claude plugins” and the first number is right — we counted. The ten plugins appear nowhere on the card, which has six agents and names no plugin at all, but that is the smaller thing here.
The larger thing is three words that appear in three different boxes: monitoring.
- Regulatory change monitoring (agent 04)
- Privacy policy monitoring (agent 02)
- Policy monitoring (agent 04)
Monitoring means telling you what has changed. A model holds a snapshot, and the snapshot has a published date.
Source card: “Build Your Claude Legal Team” — BANI Academy, 21 October 2026. Six agents: contracts & corporate, privacy & data protection, litigation & disputes, regulatory & AI compliance, employment & product law, IP & legal education.
View the original post on Facebook
The date on the snapshot
Every current Claude model carries the same figure on its specification page:
Reliable knowledge cutoff: June 2026 — for Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 alike.
So a model asked to monitor regulatory change is being asked for the one thing a fixed snapshot cannot supply. It will answer. The answer will describe the position as of its cutoff, in confident present tense, with no marker separating “this was true in June” from “this is true now”. In a domain where the whole point of the word monitoring is the delta, that is the failure mode to design around.
None of this makes the card wrong about what Claude is good for. It makes the thirty boxes three different kinds of task wearing one visual style.
Diagram: the thirty, sorted by what each one needs
| What the task needs | How many | Which workflows |
|---|---|---|
| Only the documents you supply The material is in front of it. This is reading, comparing and drafting |
17 | Vendor agreement review · NDA triage · SaaS contract review · M&A due diligence · Board minutes & consents · Data processing agreement review · DSAR responses · Privacy issue triage · Case chronology building · Deposition preparation · Claim chart building · Legal hold management · Brief drafting · Employment agreement review · Termination review · Case briefing & IRAC practice · Law school exam preparation |
| Current law, in your jurisdiction The answer depends on rules that differ by country and change |
8 | Worker classification · Open-source license compliance · AI use-case assessments · AI vendor reviews · Marketing claims review · Product launch review · Privacy impact assessments · Compliance gap analysis |
| Continuous watching The deliverable is the change since last time |
3 | Regulatory change monitoring · Privacy policy monitoring · Policy monitoring |
| A search of a register Not a question to answer from memory |
2 | Trademark clearance · Patent claim analysis |
Trademark clearance is a lookup, not a recollection
The bottom row deserves singling out, because it fails differently from the others. Clearing a trademark is a search. You go to the registers, you look for conflicting marks in the relevant classes and territories, and the output depends entirely on what is on file this week.
Ask a model with no live lookup and you do not get a refusal. You get an answer: a plausible assessment, phrased the way a clearance opinion is phrased, assembled from whatever marks happened to be in the training data. It will be fluent. It will have searched nothing. And unlike a stale regulation, there is no version of this that was ever right — a register is not knowledge, it is a database, and you either queried it or you did not.
Patent claim analysis sits in the same place for the same reason: the analysis is only as good as the prior art in front of it, and prior art is a search result.
The fix is not to drop these two. It is to do the search yourself, paste the results in, and let the model do the comparison — which moves both workflows from the bottom row into the top one, where seventeen others already are.
Two lines worth adding to every one of the thirty
If a card hands you thirty legal prompts, these are the two clauses missing from all of them.
| Add this | Why | What it changes |
|---|---|---|
| “Today is [date]. Your knowledge has a cutoff. For anything that depends on current law, say so explicitly and tell me what to check.” | A model does not volunteer the boundary of its own knowledge | You get the answer plus a list of what to verify, instead of a confident snapshot with no date on it |
| “Jurisdiction: [country/state]. If a point differs by jurisdiction, name the difference rather than answering generally.” | Worker classification, privacy duties, marketing rules and licence enforcement all differ by territory | Stops a generic answer that happens to describe somewhere else |
| For the two register tasks: paste the search results | A register is a database, not knowledge | Turns a confabulation risk into a comparison task, which is what the model is actually good at |
| For the three monitoring tasks: supply the two versions | “What changed” needs a before and an after | The model is excellent at diffing two documents you give it. It cannot fetch the second one |
That last row is the useful reframe. “Regulatory change monitoring” is impossible as a standing request and straightforward as a diff. Paste last quarter’s text and this quarter’s, and ask what moved and what it means for you. The model does not need to watch anything; you did the watching, and it does the reading — which is the division of labour the top row of the table is built on.
Seventeen of the thirty are genuinely good uses
It would be a bad reading of this card to leave with the impression that the list is weak. More than half of it describes exactly what a long-context model is for.
| Group | Examples from the card | Why these work |
|---|---|---|
| Reading a document you have | Vendor agreement review, NDA triage, SaaS contract review, Employment agreement review | The whole text fits in context. The job is comparison against a standard you state — and you can state the standard |
| Building structure out of a pile | Case chronology building, Claim chart building, M&A due diligence, Board minutes & consents | Dull, high-volume, error-prone by hand, and entirely contained in the material you supply |
| Drafting from a template and facts | Brief drafting, DSAR responses, Deposition preparation | You hold the facts and the format. The model holds the patience |
| Practice and teaching | Case briefing & IRAC practice, Law school exam preparation | The safest two on the card. Nothing is relied on; being argued with is the point |
The best line on the card is the smallest
At the bottom, in grey:
“AI tools can assist legal work but are not a lawyer and do not give legal advice. Always have a qualified professional review important documents and decisions. Plugin names and features may change.”
That is the correct disclaimer, in the correct place, and it does something most fine print does not: it names who reviews (a qualified professional) and what (important documents and decisions), rather than saying “results may vary”. On a card that uses the words deposition, DSAR and trademark clearance, that sentence is doing more work than any of the thirty boxes.
It also means this piece does not need to make the obvious point. The card made it first. What the card does not say — and what the rest of this article is about — is that the thirty boxes fail in four different ways, and only one of those ways is the one the disclaimer covers.
Four honest caveats
- This is not legal advice and we are not lawyers. Everything here is about where a language model’s knowledge boundary falls, which is a technical question with a published answer. What any of these workflows is worth in practice is a question for someone qualified, as the card itself says.
- The four-way sort is our reading. A privacy lawyer might put “privacy impact assessments” in the top group, because the assessment is of your processing. We put it in the second because the criteria come from law that moves.
- A model with tools changes the picture. Give it web search and the three monitoring rows become partly feasible; give it a register’s API and trademark clearance changes category entirely. The card does not mention tools, so we read it as written: a model, and your documents.
- Cutoffs move. June 2026 is the reliable knowledge cutoff published for the current four models as of October 2026. Check the model page rather than this article, and check it again next quarter.
The one-line version
Seventeen of these thirty workflows need nothing but the file you paste in, and they are excellent; three of them have the word “monitoring” in the title and ask a June 2026 snapshot to tell you what changed since; and two are database searches dressed as questions — which is four different failure modes sharing one grid.
Paste the documents. Name the date and the jurisdiction. Do the register searches yourself and hand over the results. Then the model is doing the part it is good at.
Source card: “Build Your Claude Legal Team” — BANI Academy, 21 October 2026.
View the original post on Facebook · Related: why the model name on a card decides what it knows and where the facts in a prompt come from
