Most small business owners have now tried ChatGPT for something work-related. A fair number gave up.
The usual story goes like this. You ask it to help with a supplier email and it does a decent job. You ask it something about VAT treatment and get an answer that sounds authoritative and turns out to be wrong, or right for the wrong country. You ask it to look at last quarter's numbers and remember that it has no idea what last quarter's numbers were. After a few rounds of this you go back to doing things by hand.
The tool is not the problem. The shape of the tool is the problem.
The two things a general assistant does not have
Context. A general chatbot starts every conversation knowing nothing about you. It does not know your VAT scheme, your headcount, your payment terms, which clients are slow payers, or that you operate under French labour law rather than American employment-at-will. You can tell it, every time, in a long preamble. Almost nobody does that consistently, which is why the output quality swings so wildly.
Boundaries. A general assistant will answer anything you ask with the same confident tone, whether it is good at that thing or not. It has no concept of being out of its depth. For casual use that is fine. For anything with a compliance consequence it is a liability.
What specialisation changes
The alternative approach is to build several narrow agents instead of one broad one, each carrying its own instructions, its own reference material, and its own boundaries.
An agent scoped to French accounting knows which VAT regime applies, what the filing deadlines are, and what a liasse fiscale is, because that is all it has been set up to handle. An agent scoped to recruitment knows the rules around trial periods and fixed-term contracts under French labour law. Neither of them will confidently improvise an answer to something outside its remit, because they are not built to.
You lose flexibility. You gain reliability, which for business work is the better trade.
The coordination problem, and how it gets solved
Split your assistant into a dozen specialists and you create a new problem. Real questions do not respect departmental boundaries.
"Can we afford to hire someone in September" is a finance question, a payroll question, a labour law question and a sales forecast question at the same time. If you have to ask four separate agents and stitch their answers together yourself, you have not saved any time. You have added admin.
This is why the better multi-agent products put a coordinator in front. You ask one question, in plain language, and the coordinator works out which specialists to consult and returns a single answer.
Mirage Cloud is built this way. An orchestrator agent named Sofia sits in front of eleven specialists covering accounting, finance, HR, marketing, legal, logistics, sales, customer support, nonprofit governance, general assistance and phone reception. Ask a question that spans three domains and the routing happens without you thinking about it.
The design goal is that you delegate the way you would to a colleague, rather than assembling a prompt.
Where this approach still has limits
Worth saying plainly, because most articles on this topic will not.
Specialised agents still run on the same underlying language models as general ones. Specialisation improves reliability. It does not eliminate errors. Anything with a legal, tax or financial consequence still needs a qualified human to review it before it goes anywhere. Any vendor telling you otherwise is overselling, and most vendors' own terms of service say exactly this in the small print.
Multi-agent systems are also only as good as their connections. An agent that cannot see your bank account or your invoicing system is guessing. The value is in the integrations, so when you evaluate one of these products, look at the integration list before you look at the agent list.
And there is a floor on complexity below which this is overkill. If you are a solo freelancer with ten invoices a month, a general assistant and a spreadsheet will serve you fine.
How to evaluate one
Ask what happens when an agent does not know something. Good systems say so. Bad ones invent.
Ask which integrations are live today versus on a roadmap. Most vendors list both in the same grid with a small label, and it is easy to miss.
Ask what the fallback is when the automated answer is wrong, and whether there is a human you can reach.
Then test it on something you already know the answer to. It is the fastest way to find out whether a product is genuinely specialised or just a general model with a name and an avatar attached.
Top comments (0)