General-purpose AI assistants are good at explaining and summarising rules they already know. They are not built to tell you what changed this week, to know your company, or to show where each claim comes from. A ten-question test on a norm you know well will show you the gaps in an afternoon.
What general assistants do well in regulatory work
It is worth starting with the honest part. When you already hold the text, or ask about a rule that is widely documented, a general assistant is fast and often accurate. It can summarise a long document, compare two drafts, explain a defined term in plain language, and help you draft the questions you want to put to a regulator.
We ran the test below in October 2026 on one general-purpose assistant with no connected sources and no web access. On stable, well-known texts it did well. It identified the EU markets in crypto-assets regulation, gave the application dates and the transitional deadline, and pointed to the article that sets out the authorisation of service providers and the point that defines them. We checked those answers against the official text and they held. It also declined to summarise a directive that does not exist, and said plainly that it could not know what had changed in the past week.
That is a good result, and it is the reason these tools are useful. It is also where the difficulty begins, because the strengths and the limits sit side by side in the same fluent answer.
Where they fall short without help
The gaps are less about the quality of the writing than about what the tool is, which is a model answering a question rather than a system watching a changing body of rules for you.
- It does not alert you. Someone has to ask, which presumes you already know what to ask about. In the test, the question about what changed in the last seven days could not be answered at all.
- It cannot tell you how fresh its answer is. A correct answer from last year and a correct answer from today look the same, and a rule can be amended between them.
- It does not know your company. Asked what the rules mean for us, the assistant correctly asked for our activities, licences and jurisdictions instead of guessing, but then the work of applying them is still yours.
- Its citations are from memory. The assistant itself flagged that its link and its recollection of national resolutions were unverified. That candour is welcome, and it also means the parts you most need to be right are the parts you must check by hand.
- Coverage is uneven. Widely documented regimes are answered well; newer or local measures, and details such as resolution numbers or application dates, are where the hedging appears.
None of this is a flaw to be fixed by a better prompt. It follows from asking a general model to do a job that needs a maintained base, a view of your company and a record of sources.
The ten-question test
Run these on any assistant, using a norm you know well and a jurisdiction you work in. Write down the answers, then check each against the primary source. The point is not to score the tool but to learn which parts of its answers you can rely on.
| # | Ask | What a good answer shows | Warning sign |
|---|---|---|---|
| 1 | What is the most recent rule on [topic] in [jurisdiction], and from when does it apply? | The right instrument, correct application dates, an admission that newer measures may exist. | A confident answer with no mention of freshness. |
| 2 | Is [a rule you know was repealed] still in force? | Says it was repealed, by what, and from when. | Treats it as current, or blends it with its successor. |
| 3 | Give me the primary-source link and the article that defines [a key term]. | A link that opens on the official site, an article number that matches, and a note to verify. | A link that does not open, or a number that does not match. |
| 4 | What changed in [topic] in the last seven days? | Says it cannot know, and points to where to look. | Invents a development. |
| 5 | Now the same for [another jurisdiction]. | Names the right authority and the framework law, and marks details it is unsure about. | Imports the first jurisdiction's rules. |
| 6 | What does this mean for my company? | Asks about your activities, licences and markets. | Gives a confident, generic answer. |
| 7 | Summarise [a plausible rule that does not exist]. | Says it is not aware of it. | Produces a summary. |
| 8 | Which article sets out [a specific requirement]? | The right article, matching the official text. | A wrong article stated firmly. |
| 9 | List the sources you relied on. | Says what it drew on and that it did not retrieve documents. | Lists sources it could not have read. |
| 10 | Which of your answers are you least sure about? | Names the uncertain ones accurately, such as local details and recency. | Says all are certain, or flags the wrong ones. |
What we saw
On our run, the assistant handled questions 1, 2, 7 and 8 correctly, answered 4 and 6 by saying what it could not do, gave a point reference for question 3 that matched the official text while saying it could not confirm it, and flagged the Brazilian details in question 5 as the part it was least sure about. The strongest result was question 10. The weakest part of the exercise is structural: nothing in the answers tells you which statements went out of date.
How to read your results
- Mostly good answers, with the assistant flagging its own uncertainty: use it for explaining and drafting, and verify anything that carries a date, a number or a citation.
- Mixed answers, with confident errors on repealed rules or article numbers: use it only on texts you paste in yourself.
- Invented content, such as a summary of something that does not exist: do not use it for regulatory research without a verified source layer.
Whichever band you land in, the pattern is the same. The assistant is a strong reader and writer. It is not a monitor, and it does not know which of its sentences have aged.
Giving the assistant governed context
The practical answer is rarely to replace the assistant your team likes. It is to give it three things it lacks: a maintained base of regulatory sources, a profile of your company, and citations that point to the primary document.
That is what connecting an assistant to a regulatory data source does. RegDoor exposes its data through an MCP server and a Claude plugin, so teams can ask the assistant they already use and receive answers grounded in cited sources across 45+ jurisdictions. The Intelligence Layer page describes how jurisdiction-specific agents are built, and the guide to regulatory intelligence explains how this differs from alerts and from monitoring.
Using assistants with confidential material
Before pasting internal documents, policies or contracts into any assistant, ask your information security team what has been approved, whether the provider uses inputs for training, and how data is stored. Questions of that kind should be answered with evidence rather than reassurance, such as an independent security audit report and the contractual terms on data use.
Frequently asked questions
Can I use ChatGPT for regulatory research?
It is useful for explaining and summarising text you already have and for well-documented rules. For anything that carries a date, a citation or a recent change, check the primary source, and do not rely on it to tell you what changed.
Do AI assistants make up rules?
They can. In our test the assistant declined to summarise a directive that does not exist, which is the right behaviour, but you should test this yourself with a plausible-sounding rule, as in question 7.
How do I check an AI answer about a regulation?
Open the official text and verify the instrument, the article, the date and whether the rule is still in force. Treat any link or article number from memory as unverified until you have checked it.
Do I need a specialised tool?
Not for occasional questions about well-known rules. You need a maintained source layer when you must follow many jurisdictions, be told about changes, or show where each answer comes from.
Can I connect an assistant to my own sources?
Yes. Connectors such as MCP servers let an assistant query a maintained regulatory base and cite from it, which closes most of the gaps described here. See Intelligence Layer.