How to build a custom assistant that answers your customers' first ten questions
One afternoon is enough to build an assistant that answers the ten questions customers ask you before they buy. It runs on your own answers, in your own words, with the price list left out. Most of the work is writing. The last hour is a test that tells you whether anyone else can be allowed to use it.
How can AI help me grow my business?
By getting repeat work out of your hands so your hours go to the work only you can do. With two or three partners and a runway you can count in months, your output is capped by what you personally type. The first ten questions a customer asks before buying are the cheapest place to start. You answer them every week, the answers barely change, and a wrong answer costs a sale.
That is one repeatable task with a clear edge around it, which is exactly what a custom assistant is for. A custom GPT inside ChatGPT is a saved set of instructions plus a few reference files you upload. OpenAI's help pages now carry a notice that it plans to retire custom GPTs and move this kind of work to Plugins, with dates published per plan. So treat the written instructions as the asset and the container as temporary. Keep your instructions in a document you own, and you can rebuild the assistant anywhere.
Collect the ten questions from your own inbox
Open your inbox and your WhatsApp Business thread and read backwards through the last sixty customer messages. Write down every question that appears more than twice, in the customer's own words. Under each one, write the answer you actually give, and next to the answer, the document it comes from.
A nutraceutical brand selling direct ends up with a list like this. What is in it. How many days does one pack last. When will my order reach me. What if it does not suit me. Do you ship outside India. Done looks like one page, ten questions, ten answers, ten sources. If an answer has no source document, that is a gap in your business, not a gap in the assistant.
Write the instructions as a brief
The instructions are a brief with five parts: role, context, task, format, constraints. Spend your time on the last two. Format is where you decide the shape of every reply, and constraints are where you decide what the assistant must refuse.
Format here is narrow. The assistant takes one customer question and returns an answer of about eighty words, in plain English, ending with one question back to the customer.
Then the constraints, written as flat rules. Never state or estimate a price. Never promise a date, a stock count or a delivery window. Use only the uploaded files. When the answer is not in the files, say it will be checked and stop. Write those four lines before you write anything clever.
What do I upload, and what do I keep out?
Upload only the documents you would hand a new salesperson on day one. For the nutraceutical brand that is the product sheet, the shipping and returns policy, two past customer email threads with names removed, and the one-page list of ten questions and answers you just wrote. A small, clean set beats a folder dump every time, because every extra file is another place a wrong answer can come from.
Keep out the price list. Keep out your customer list, your partner contracts and anything carrying a third party's personal data. I ran a P&L for 25 years and I never let a price leave a desk that was not carrying the margin, so I see no reason to hand that desk to an assistant. Prices move, and a confident wrong price quoted to a customer is the most expensive mistake this assistant can make.
Test it with ten real customer messages before anyone uses it
Paste in ten real messages verbatim, with the typos, the half sentences and the voice note transcriptions. Do not clean them up. Score each reply as right, vague, or wrong, and write one line on why. Vague is the category that matters most, because vague is what gets forwarded to a customer and then quietly loses the deal.
Now the test that decides whether this is safe. Ask it something you deliberately left out of the files, the kind of edge question a real buyer asks on a Friday evening. If it returns a confident answer instead of saying it will be checked, your constraints failed, and the fix is one line in the instructions. Change the line, clear the chat, and run the same ten messages again. An assistant that says it will check is more useful to a founder than one that sounds sure.
Two rounds of that is usually enough to let your partner use it, and a month of logged replies before a customer sees one unedited. Session 5 of Module 1 in A.D.A.P.T. AI Coaching ends with exactly this, one assistant running on the client's own work.
What else should I settle before customers see this?
Who should own the assistant if I have two partners?
One person owns the instructions file and the upload folder. Everyone can use the assistant, one person changes it. Two people editing constraints in the same week is how an assistant starts contradicting itself, and nobody can tell you which edit caused the drift.
Should I put this on my website or keep it internal?
Internal first, for at least a month. Your team sees every reply before a customer does, which is your only cheap way to catch bad answers. Public comes later, after the logged replies stop surprising you and after you are sure the files hold no price or personal data.
How do I keep the answers current as the product changes?
Put a date on every uploaded file and re-upload the moment the underlying document changes. Once a month, run the same ten test messages again and compare against last month's replies. Drift shows up as answers that are still fluent and no longer true.