The short version: Training an AI assistant on your own content, blog posts, emails, transcripts, proposals, can answer 70 to 80 percent of the repetitive questions your business gets, for less than the cost of one hour of your time a month. It will not close a sale for you and it will expose every gap in your content faster than any consultant will. Build it in a weekend, watch it for a month before you trust it.
Why I even tried this
Last year I was answering the same six questions on repeat. What do you charge. Do you do one off sessions or ongoing. Can you speak at our event in Tel Aviv or does it have to be UK only. Do you work with startups or only bigger businesses. I was typing near enough the same reply into WhatsApp, email and LinkedIn DMs, sometimes three times in one day.
So I did something that felt slightly mad at the time. I pulled every blog post, every speaker one sheet, every proposal I’d sent in the last two years, every FAQ page, and fed the lot into a custom GPT inside ChatGPT. Took me a Sunday afternoon, maybe five hours including tea breaks. Cost nothing extra because it sits inside the ChatGPT Plus subscription I already pay 20 dollars a month for.
I called it, unimaginatively, “Ask Lilach Anything” and put a link to it on my speaker page instead of a contact form for the basic questions.
What happened in the first month
The questions it handled well were the boring, factual ones. Fee ranges. Whether I travel for keynotes. What topics I cover. Whether I do half day or full day workshops. Those answers were sitting word for word in old proposals, so the assistant just found them and served them back, no drama.
What it got wrong was more interesting. Someone asked if I’d done work with fintech companies. I have, quite a bit, but the case study I’d written about it was three years old and the assistant pulled details that were out of date, an old day rate, a client name I’m no longer allowed to use publicly. That’s not an AI problem. That’s a content problem. The bot only ever knows what you’ve told it, and most of us have never gone back and cleaned up what we told it years ago.
That’s the uncomfortable bit nobody selling you an AI chatbot wants to say out loud: the tool doesn’t fix messy, outdated content, it broadcasts it faster and with more confidence. A human reading an old case study might think “this feels dated.” An AI reads it with the same certainty as your newest, sharpest page. Same tone, same authority, no flag saying “this bit’s from 2022.”
How to build one (steps, not theory)
- Step 1, gather your content. Everything you’ve written that answers a question a prospect or customer has asked you. Blog posts, FAQ pages, old proposals with client names stripped out, transcripts of talks, even your best email replies. Aim for at least 15 to 20 solid documents.
- Step 2, pick your tool based on budget. If you already pay for ChatGPT Plus at 20 dollars a month, use the custom GPT builder, it’s included. If you want it embedded on your actual website with a chat widget, tools like Chatbase run from about 40 dollars a month for the basic tier. Either way you’re uploading documents, not writing code.
- Step 3, write the instructions carefully. This is the part people rush. Tell it explicitly what it should never do: quote a specific price without a disclaimer, promise availability, name a client who hasn’t given permission. Tell it what to do when it doesn’t know: say so and point to a real contact method, not guess.
- Step 4, test it with the ten worst questions you can think of. Not the easy ones. The awkward, ambiguous, slightly rude ones real people send. If a stranger asked “why should I pick you over anyone else,” what does it say. If it goes vague or starts inventing detail, fix the source material, not the prompt.
- Step 5, put it somewhere low stakes first. I linked mine from a speaker page, not my main contact form, for the first six weeks. Watch what people ask it before you hand it anything customer facing.
What it’s good for
Pre-qualifying. Someone thinking about hiring me for a keynote can ask the assistant three or four questions before they ever fill in a form, and by the time they do reach me they’ve already ruled themselves in or out on budget and availability. That’s a proper win, it means the calls I take are with people who were already fairly sure.
It’s also good at consistency. My tone doesn’t drift depending on whether I’m tired or rushed. If you’ve ever read a brand’s customer service replies and noticed how the wording changes person to person, day to day, an assistant trained on your own material keeps that voice steady. Look at how Revolut built consistency into every touchpoint of their brand, that’s the same principle at a much smaller scale, one person’s business instead of a fintech giant.
And it’s a decent early filter for interactive tools too. A few years back I wrote about using interactive calculators to engage visitors, and this works on the same logic, give people something to poke at before they commit to a form or a call, and more of them stick around long enough to convert.
What it will never do, and don’t let anyone tell you otherwise
It doesn’t build trust. Dale Carnegie’s whole point, the one that still holds up nearly a century later, was that people buy from people they feel understood by, not from the most technically correct answer in the room. I wrote about this in my piece on business lessons from Dale Carnegie, and the short version is: an AI assistant can answer “what do you charge,” it cannot make someone feel like the only client in the room. That’s still a human conversation, always will be.
It also can’t read a room it’s never seen. If a prospect is nervous, or annoyed because a previous consultant burned them, or testing you to see if you’ll be honest about your limits, a scripted assistant trained on your polished content will miss all of that. It’ll answer the words, not the person.
Where I’d tell a small business owner to stop and think first
Before you build one of these, look at what people get stuck on. I only knew my top six repeated questions because I’d been keeping half an eye on my inbox for two years, not because I had proper data. If you want to do this better than I did, use something like Hotjar first, the heatmap and session recording tool, to watch where visitors hover, scroll back, or abandon a form. I covered how that kind of behavioural data changes decisions in my piece on Hotjar’s own marketing strategy, and the lesson transfers directly, watch behaviour before you build the answer.
Also think about personalisation rather than as a buzzword. Duolingo does this brilliantly, the app adjusts what it shows you based on what you’ve done, not what a generic user might do. I looked at how that’s built into their whole approach in the piece on Duolingo’s marketing strategy. A content trained assistant can do a rough version of this too, if someone’s asked about keynote fees, don’t have it repeat generic pricing on the next question, have it reference what they’ve already told you.
The real cost, in full
Here’s the honest breakdown, because most posts on this topic are strangely vague about money. My setup costs me nothing on top of the 20 dollars a month I was already paying for ChatGPT Plus. If you want it embedded as a chat widget on your website rather than a separate link, budget 40 to 150 dollars a month depending on how many conversations it handles and whether you want it connected to your CRM. If you pay someone to build and train it rather than doing the Sunday afternoon version I did, expect a one off fee somewhere between 300 and 1500 pounds for a small business setup, depending on how much content needs organising first. That’s roughly what I’ve seen quoted by AI consultants doing this kind of work for small firms, and it’s covered in more detail on the page about what an AI consultant costs.
If you’d rather not do the Sunday afternoon version yourself and want someone to set this up, sorting the content, writing the guardrails, testing it against real questions, that’s exactly the kind of project an AI consultant for small business handles day to day, and it’s usually a two to three week job rather than months.
What I’d tell you honestly if you were sat across from me
Do it if your content is already decent. If you’ve got two or three years of blog posts, proposals, FAQs, case studies that are current, this is one of the highest return, lowest cost things you can build this year. Don’t do it if your website copy is thin or three years stale, because you’ll spend the weekend building an assistant that confidently repeats mistakes you haven’t noticed yet. Fix the content first. The AI part is the easy bit.
Frequently asked questions
Do I need to know how to code to train an AI on my own content?
No. Tools like the custom GPT builder inside ChatGPT and platforms such as Chatbase work by uploading documents, no coding involved. You need organised content and clear instructions, not technical skill.
How much content do I need before this is worth doing?
Aim for at least 15 to 20 solid pieces, blog posts, FAQs, old proposals, case studies. Below that the assistant has too little to work with and will start guessing, which is worse than having no assistant at all.
Can an AI assistant replace my customer service or sales calls?
No, and treat anyone who tells you otherwise with suspicion. It handles repetitive, factual questions well, pricing, availability, what you offer. It cannot build trust, read tone, or handle an upset or hesitant customer the way a person can.
What’s the biggest mistake small businesses make with this?
Uploading outdated content without checking it first. The assistant repeats old prices, old client names, and old claims with total confidence, because it has no way of knowing your content is stale unless you tell it.
Related reading: Write for Us, Startup Blog | Guest Post on Lilach Bullock and Write for Us, Travel Blog | Guest Post on Lilach Bullock.
