Essay · · 4 min read · By Amalgament
ChatGPT can handle much of your writing and thinking. It can't run the parts of your business that live in other systems, repeat every week or need someone accountable.
What it does well
A chat assistant is good at work that starts and ends in the conversation. It drafts an email, tidies a proposal, summarizes a long document, explains a long email thread, turns notes into a checklist, or helps you think through a decision. For a busy owner that is real time back, and it costs a monthly subscription.
If that is most of what you need, you may not need anything else. Use it, and use it often.
Where it stops
The limits show up when the work leaves the chat window. There are five worth knowing.
It sees what you give it. Unless you connect it to something, the assistant knows only what you paste or upload in that conversation. It does not see your accounting, your schedule or your CRM. If the question needs all three, someone has to collect the pieces first, and the answer is only as complete as what they collected.
It does not run on its own. A chat answers when someone asks. The weekly follow-up, the reminder to a client who has not sent a document, the copy of a new customer into accounting: those need to happen every time. They cannot wait for someone to remember to open a chat.
What it learns lives in chats. Over time, people write good instructions: how to describe your services, which tone to use, what never to promise. Those instructions end up in saved conversations and in one person's habits. When that person leaves, much of it leaves with them.
The model underneath changes. The companies that make these models replace them regularly. Anthropic's developer documentation says: "As safer and more capable models launch, Anthropic regularly retires older ones." It adds: "Requests to retired models fail." OpenAI's says: "As we launch safer and more capable models, we regularly retire older models. Software relying on OpenAI models may need occasional updates to keep working." Those pages are written for developers, but the same thing happens behind a chat app. The model answering you next year will not be the one answering today, and it may read your saved instructions differently.
It can be wrong, smoothly. Every one of these assistants sometimes gets a fact or a figure wrong, and the wrong answer reads exactly like a right one. In a draft, that costs a minute of checking. In a message already sent to a client, it costs more.
Drafting is cheap. Sending is not.
The most useful rule for AI at work is simple. Let it prepare as much as you like, and keep a person on the step where the work leaves the business. An email to a client, a change to an invoice, a note in a customer's record: once it goes, someone outside the room sees it.
A person who reads a draft before it goes out catches the obvious mistakes. That review is quick when the draft is good, which is most of the time.
How to get more from it without more software
- Keep your instructions in a document you own. How you describe your services, your tone, what you never promise. Paste it in when you need it. When the model changes, you test one document instead of rediscovering everything.
- Decide what may go in. Read your plan's terms on how your conversations are used and stored, and set a simple office rule about client details. If you would not put it in an email to a stranger, think before pasting it.
- Give it the whole picture for the question. If you want an answer about a client, paste the relevant invoice and the latest email together, and say which one is right if they disagree.
- Check numbers against their source. A figure from the assistant is a lead to verify, not an answer to forward.
- Keep a person on anything that goes out. Draft with it freely. Send yourself.
When you need more than a chat
There are three signs that a chat assistant alone is no longer enough.
- Someone spends time each week copying information from your systems into a chat, or from a chat back into your systems.
- A task needs to happen every time, on a schedule or when something changes, and today it depends on someone remembering.
- The answer you need crosses two or more systems, such as which jobs were invoiced but not paid, and assembling it by hand is a weekly chore.
None of those is really an AI problem. They are connection and automation problems. Sometimes AI is part of the fix: reading an incoming form, drafting a reply for a person to approve. Often the fix is a plain connection between two tools that already exist, with no AI at all.
The useful question is not whether to use ChatGPT. Most firms should, for the work it is good at. The question is which parts of your week need to work reliably without anyone opening a chat, and who will set those up and keep them working.
Amalgament is a fractional hire who looks after growing, owner-led firms. We learn how your business works, find what could be better and build it, adding AI only where it helps a person do the work. If you are not sure which of your problems need more than a chat, we are happy to help you sort them.
Note, October 28, 2026: some wording, including this closing paragraph, was updated to match how we now describe our work. The advice is unchanged.