The short version: A generic AI chatbot doesn’t know your pricing, your tone, your client history, or the way you talk to people, and that’s why most small business owners get bland, generic answers out of it. Feed it your own documents through a Custom GPT, a Claude Project, or Google’s NotebookLM and you get something closer to a junior employee who has read everything you’ve ever written. It takes about three hours to build the first version and ten minutes to update it after that.
Why ChatGPT doesn’t already know your business
I had a client last year, a wedding photographer in Kent, who was convinced ChatGPT was “useless for her business” because every email reply it drafted sounded like it belonged to a different photographer in a different city. She’d type “write a reply to this enquiry” and get back something warm and correct and completely wrong for her.
Here’s the bit nobody tells you when they’re pushing you to “use AI for your marketing”: the model has no idea who you are. It doesn’t know she shoots documentary-style, hates posed group shots, charges £2,400 for a full day, and always mentions her insurance cover because a competitor once left a bride stranded. It’s a brilliant writer with zero memory of your business, and every time you open a new chat, it forgets everything again.
This is the single most useful thing I’ve taught clients this year, more useful than any prompt template: stop treating ChatGPT like an oracle and start treating it like a new hire on day one who needs a proper induction pack.
What a private AI assistant is (no coding needed)
You don’t need a developer, an API key, or a budget. Three tools do this job well and none of them require code:
- Custom GPTs inside ChatGPT Plus (£16 to £20 a month depending on exchange rates). You upload files, write instructions once, and it becomes a private assistant only you can access unless you choose to share it.
- Claude Projects inside Claude Pro (around $20 a month). Same idea, and in my experience Claude is better at holding a consistent tone once it’s learned yours.
- Google NotebookLM, which is free and brilliant if all you want is something that answers questions from a stack of documents without writing new content in your voice.
All three work the same way underneath: you give the tool a pile of your own documents, it reads them, and every answer it gives you afterwards is filtered through what it just learned about your business instead of the internet’s average guess.
The document pile that matters
Most people who try this fail because they upload the wrong things, usually one glossy brochure and nothing else. What moves the needle is boring, specific, internal stuff:
- Your last ten client emails, the real ones, typos and all
- Your pricing sheet with every package and every add-on
- Three past proposals that won you work, and one that lost
- Your FAQ page, even the messy internal version you never published
- A page of “how I talk”, written by you, not about you
That last one matters more than people expect. When I built mine, I wrote a single page describing how I sound when I’m annoyed, how I sound in a sales email, and three phrases I never use. Documenting your voice the way GoPro or Mailchimp would document a brand voice guide forces you to notice patterns you never wrote down before, which is honestly half the value even before the AI does anything with it.
How I built mine, step by step
I run my own consultancy, so I built a Custom GPT trained on eighteen months of my own client work. Here’s exactly what I did, start to finish, in one sitting:
- Step 1. I pulled twelve past proposals into one folder and stripped out client names and figures I didn’t want repeated.
- Step 2. I wrote a one-page brief covering my three service packages, who each one is for, and who it’s wrong for.
- Step 3. I added a “tone” document, one page, written as instructions: short sentences, no corporate padding, British spelling, no exclamation marks in client emails.
- Step 4. I uploaded all of it into a new Custom GPT, gave it a name, and wrote instructions telling it exactly what jobs it was allowed to do: draft proposal first drafts, draft follow-up emails, never invent a price it hasn’t seen in the documents.
- Step 5. I tested it on ten real scenarios from the previous month and marked what it got wrong.
The whole thing took three hours the first time, mostly spent gathering documents rather than fiddling with the tool. Updating it now, adding a new case study or a new price, takes about ten minutes.
The uncomfortable bit about privacy nobody mentions
Here’s the part most guides on this topic skip entirely because it’s inconvenient: if your documents contain client names, contract terms, or anything covered by an NDA, uploading them into a public AI tool needs a second thought, not a shrug. Free tiers and consumer accounts generally have different data retention and training settings than paid business accounts, and the default is not always the private one. Check your settings before you upload anything, turn off model training on your data where the option exists, and if you’re handling anything sensitive, redact names and figures before they go anywhere near the tool.
I’ve seen a small agency owner upload a full client contract, name and fee included, into a free chatbot account just to “summarise it faster.” That’s not a small business being efficient, that’s a small business handing a stranger’s paperwork to a company with no obligation to protect it the way she’d protect it herself. The fix isn’t complicated, it’s just a step people skip because it’s boring: strip identifying details, use a paid business-tier account when the content is sensitive, and read the actual data policy once rather than assuming.
What it cost me and what it saved
My setup costs £20 a month for the Plus subscription I was already paying for anyway, so the marginal cost of the assistant itself was zero. The time saved is where the number gets interesting.
Before I built it, a new proposal took me roughly ninety minutes: pulling old ones for reference, matching tone, checking pricing was current. After, a first draft comes back in under ten minutes and needs maybe fifteen minutes of editing. Across a typical month where I’m sending four to six proposals, that’s five to six hours back, every single month, from a tool I was already paying for and barely using.
The same logic applies to newsletter drafts, the kind of thing brands like Mailchimp have built entire businesses around helping people write faster; feeding your assistant six months of your own newsletter content means every new draft sounds like you rather than like a template.
Where this goes wrong
Two mistakes show up constantly when I watch clients try this themselves. First, they upload too much at once, three years of unsorted documents, and the assistant gets vague because it can’t tell what’s current and what’s outdated. Keep the pile tight and update it, don’t let it grow into a landfill. Second, they never test it against real past scenarios, so they don’t discover it’s confidently repeating a price that changed eight months ago until a client points it out. Ten test questions before you trust it with anything client-facing isn’t optional.
There’s also a discipline problem that has nothing to do with the AI and everything to do with habits people already had. If your team’s institutional knowledge lives in someone’s head rather than on paper, the way Joe Pulizzi talks about building content systems that outlast any one person, an AI assistant just exposes that gap faster than it would have shown up otherwise. That’s not the tool’s fault. It’s a business that never wrote things down finding out the hard way.
When to build it yourself and when to get help
If you’re a solo operator or a team of two or three, build it yourself this week. It doesn’t need outside help at that scale, and the exercise of gathering your own documents is worth doing with your own hands. Where it gets worth paying for is when you have more than one person needing access, when you’re connecting the assistant to live systems like your CRM or your booking calendar, or when the documents involved are complex enough that getting the instructions wrong could cost you a client relationship. That’s the point at which working with an AI implementation coach earns its fee, because the mistakes at that scale are more expensive than the help.
Frequently asked questions
Do I need to know how to code to build a private AI assistant?
No. Custom GPTs, Claude Projects, and Google NotebookLM all work through uploading documents and writing plain-English instructions. No code is involved at any point.
Is it safe to upload client information into a Custom GPT?
Only if you check the settings first. Turn off model training on your data where the option exists, use a paid business-tier account for sensitive material, and strip out client names and figures where you can. Don’t assume the default setting is the private one.
How long does it take to build one?
The first version takes about three hours, most of that spent gathering your own documents rather than using the tool. After that, updates take ten to fifteen minutes.
What documents should I upload?
Real past emails, your current pricing, a handful of proposals that won and lost you work, your internal FAQ, and a short page describing how you talk. Skip the polished brochure, it teaches the AI nothing about how you really operate.
Prefer to hand this over? This is a conversational AI consultant in Israel.
Want the complete version? Read where I break down AI marketing.
Related reading: Why AI Pricing Advice Is Quietly Training Small Business Owners to Undercharge and Monday.com Marketing Strategy: How They Built a Brand That Wins.
