Excellent reasoning, still a generalist

It is tempting to frame every AI comparison as which model is smarter. On professional questions that is rarely the constraint. Claude will reason about a tenancy dispute perfectly well; what it will not do is assume you are the tenant, apply your state's statute, and produce the letter.

That is a routing and configuration problem, not an intelligence one. A specialist starts from a narrower brief and can therefore skip straight to the specific.

If your work is reading and writing long documents, Claude is superb and this page will not talk you out of it. Large context, careful reading, and prose that does not sound generated are real advantages.

A contract you want summarised and explained is a good Claude task. A contract you want checked against your state's requirements, with the problematic clause flagged and a redline drafted is a specialist task.

Claude's context window is one of the largest available and it is not a marketing number - it changes what is possible on document work in a way worth being direct about.

If you have a 200-page agreement, a full deposition transcript or a year of correspondence and you want it read end to end with cross-references held in mind, Claude does that better than almost anything else, and better than Copilotly. Document analysis at that scale is genuinely its strength and no amount of domain configuration substitutes for the ability to hold the whole thing at once.

It is also unusually good at careful, structured reasoning over several steps, and at saying when it is uncertain rather than producing confident filler. Those are real properties and they matter on exactly the kind of problem this site is about.

So the honest division is by task rather than by quality. For "read this entire thing and tell me what is in it", Claude. For "this clause, in my state, with this deadline - what do I do about it", a copilot configured for that field, because the constraint there is not context length. It is knowing which three questions to ask before answering.

Several people use both, and that is a reasonable outcome rather than an admission. If your work is mostly long-document analysis, Claude alone is the right answer.

Anthropic's usage policies, like OpenAI's, restrict tailored professional advice in regulated areas without appropriate professional involvement. That is deliberate, it is reasonable, and it means the constraint on a legal or medical answer is not purely about what the model could say.

The practical effect is the same shape as with any general assistant. You get the rule, you get an accurate general explanation, and you get a recommendation to consult a professional - which is correct and is not what you asked.

Copilotly is not exempt from the underlying reality. It is not a law firm or a medical provider either, and it says so in real type on every regulated page. What differs is that the boundary is stated precisely rather than applied as a general caution: the legal copilot will tell you which specific thing needs an attorney and why, and the wellness copilot routes anything involving self-harm to a crisis line rather than attempting to help.

A precise boundary is more useful than a broad one, because it tells you what you can still do yourself. "Consult a professional" tells you nothing about which part.

Claude is one model, well tuned, applied to everything. Copilotly is a set of configurations, each aimed at one field, and part of what varies between them is which model is used underneath - including Anthropic's, for the domains where careful multi-step reasoning matters most.

That is not a competitive claim so much as an architectural one. Different jobs suit different models: some need sustained reasoning, some need speed, some need literal fidelity when reading a document. Fixing one model for all of them means being wrong for most of them, and being able to change it per domain is also what stops the product ageing badly when something better ships.

The other half of the difference is what the configuration encodes. A copilot for tax knows to ask which notice code is in the corner before answering, because CP504 and LT11 carry different rights and different windows. That is not intelligence, it is domain knowledge baked into the questions - and it is the sort of thing a general model produces only if you already knew to ask for it.

If you want the full mechanics, how it works covers what a copilot is at the configuration level and where the approach stops helping.

Abstract comparisons are easy to agree with and hard to act on, so here is the concrete difference on one question.

Ask a general assistant about a flagged thyroid result and you get a good explanation of what TSH measures, the general reference range, and a recommendation to discuss it with your doctor. All accurate. None of it tells you whether yours is marginal or meaningful, what else moves it, or what to ask at an appointment three weeks out.

A copilot configured for health asks first: what was the actual number, was free T4 run alongside it, is this the first time or is there a trend, what medication are you on. Those are the questions a clinician asks, and asking them turns a frightening number into a picture.

What comes back is different in kind. Not a diagnosis - it will say plainly that it cannot give you one - but what the marker measures, what commonly moves it, why a single point matters less than a trend, and the three questions worth taking into the appointment so that fifteen minutes goes further.

The same pattern holds across domains. The general answer is the rule; the specialist answer is the rule plus the questions that make it apply to you, plus an explicit statement of where it stops. The health copilot page shows the whole exchange rather than describing it.

Claude is unusually good at declining to guess, and that is a genuine virtue. It is also not free.

What happens after a well-calibrated "I cannot advise on this specifically" is that the question gets postponed. Someone asks about a clause, gets an accurate general explanation, decides an attorney is not worth it for one paragraph, and the document goes back in the drawer. Three months later the auto-renewal fires or the notice period turns out to have been ninety days.

That chain is not a failure of calibration - the model was right to be careful. It is a failure of the exchange to produce a decision, and a decision is what the person came for.

The useful middle is narrower than either extreme. Not advice, and not a refusal: the specific rule that applies given your facts, the deadline if there is one, and a draft you can send - with an explicit statement of which part still needs a professional. That last part is what makes the rest safe to act on.

It is a smaller claim than replacing anyone. It is also the one that turns a postponed question into a resolved one, which is the whole difference in outcome.