The short answer
Use a general assistant (Copilot inside Microsoft 365, ChatGPT on a business plan, or the equivalent inside the tools you already run) for the work a person is doing anyway: drafting, summarising, finding, rewriting, first passes at analysis. Set it up properly, on business accounts, with permissions checked. That covers most of what most people will use AI for, and you probably already licence it.
Build something for you when a piece of work needs to happen without a person driving it each time, when it has to read from and write to your systems, when it must follow rules a general assistant cannot be told, or when the answer has to come from your knowledge with a source. That is an agent, an automation, a document pipeline or a knowledge assistant, and it is designed rather than bought.
What a general assistant is good at
A general assistant is at its best with a person in the chair. It drafts the email, and the person sends it. It summarises the forty-page report, and the person checks the summary against the two pages that matter. It writes the first version of the proposal from the meeting notes, and the person makes it true. Inside a platform like Microsoft 365, it can also see your documents, which makes it far more useful than the public version and raises the permissions question you must answer before switching it on.
It is cheap per person, needs no build, and improves every month without you doing anything. For the majority of staff, it is all the AI they need, and a business that has never configured the assistant it already pays for should do that before building anything.
Where a general assistant stops
Every one of those is fine when a person is in the loop, because the person is the rules, the action and the record. They are not fine when the work is meant to run on its own.
- It waits to be asked. Nobody asks it to process the two hundred invoices that arrived overnight, so they wait for a person.
- It does not act in your systems. It can draft a reply; it cannot look up the order, update the CRM and schedule the follow-up.
- It cannot be given your rules in a way that holds. You can ask it nicely to follow policy; you cannot make it wait for approval before a refund.
- It answers from what it was trained on unless it is connected to your knowledge, and even connected, it does not reliably cite the source or say when it does not know.
- It leaves no record you control. When something it produced reached a customer with an error, you cannot show how.
What “built for you” actually means
It rarely means training your own model. It means designing a system around a model: the knowledge it may draw on (your documents, your systems, with permissions respected), the tools it may call (read the order, draft the reply, create the record), the rules about what it may do without a person, and the record of every step. The model inside is usually one of the same vendors’ models the assistant uses. The design around it is what makes it yours.
That design is also what makes it cost more than a licence. It is worth it for a process that runs many times a day, touches several systems, and currently consumes a person’s time. It is not worth it for a task a person does twice a week with the assistant’s help.
Six questions that settle it
- Does the work need to happen without a person starting it? If yes, built.
- Does it need to read from one system and act in another? If yes, built.
- Must it follow rules that cannot be broken by a clever request? If yes, built.
- Must the answer come from your knowledge, with a source, and say when it does not know? If yes, built, or at least a knowledge assistant configured over your documents.
- Is a person in the chair every time, checking the output before it goes anywhere? If yes, assistant.
- Is the volume low and the variety high? If yes, assistant; the design cost is not recovered.
They are not rivals; they share an architecture
The businesses that get the most from AI run both, on one design. The assistant for everyone, configured on the platform they already own, with permissions checked. Two or three built systems for the processes that justify them, drawing on the same knowledge base so the assistant and the agent give the same answer. One policy, one record, one owner. When the assistant and the systems are bought separately from different suppliers, the knowledge diverges, the permissions are checked twice or not at all, and nobody can say what the AI in the business knows.
The design that holds them together is the intelligence layer: people, applications, agents, models, knowledge, integration, with security around all of it. It is the same drawing whether the first thing you switch on is Copilot or a document pipeline.
Where to go from here
The AI Architecture service on this site produces that drawing for your business: which work stays with an assistant, which becomes a system, what they share, and what it will cost to run. If the honest answer for you is “configure the assistant we already pay for”, that is what the architecture will say. The AI Opportunity Score below is the five-minute version of the same question.
Published 12 September 2026 · Be AI