AI for Custom Chatbots.
Most chatbots are bad because they're trained on nothing in particular. A custom one is trained on your real docs, your real policies, and the questions your team actually gets asked.
- $$
Site assistant trained on your real docs
Lives on your marketing site. Answers questions from your actual product pages, FAQ, pricing, and case studies. Cites sources. Hands off to a human or books a call when it can't answer.
- $$
Internal staff tool over your knowledge base
For teams where the answer is somewhere in the SOP doc, the email archive, or the senior person's head. Returns cited answers. Flags things it isn't confident about for human review.
- $$
Support deflection without hiding the human
Handles the 30-40% of tickets that are policy questions or order status, hands the rest to a real agent with context attached. Visible 'talk to a human' button at all times — that's the difference between adoption and resentment.
- $$
Lead qualifier in conversation form
Replaces the long contact form with a short conversation. Asks the questions sales needs to know, books qualified leads into a calendar, sends the rest a useful resource.
- $$$
Domain-specific assistant in your voice
For situations where tone matters — a financial advisor's site, a clinic, a luxury brand. Trained on your real writing samples so it doesn't sound like generic ChatGPT. The voice work is most of the build.
- $
Admit-when-you-don't-know prompting
The unsexy part. The system prompt has to make the model say 'I don't know, here's a human' instead of inventing an answer. Most off-the-shelf chatbots fail at this. It's the first thing I build.
How I think about this.
- BRIEF
What they asked for: a chatbot. What they actually needed: a better intake form.
Read note - NOTE
Watched a client click the "regenerate" button six times in thirty seconds today, getting six slightly different…
Read note - BRIEF
An SMB owner who almost spent $90k. The right answer was $0 for now.
Read note
Things people ask before getting started.
Why not just use Intercom Fin or a similar off-the-shelf bot?
For some businesses, those are the right answer — especially if you have a clean help center and a high volume of repetitive support tickets. The custom build is for when off-the-shelf hits its ceiling: when the answers live in your SOPs and email archive instead of a public help center, when tone matters, or when the bot needs to do real work like booking, qualifying, or pulling from systems Intercom doesn't know about. Build vs buy is a real conversation, and I'll tell you when buy is the right call.
What stops it from making things up?
Two things. First, the chatbot only answers from your real documents, with citations — if the answer isn't in source, the model is instructed to say so and offer a human handoff. Second, about 40% of the system prompt is rules against confident guessing. The first thing I build, before anything else, is the 'admit when you don't know' behaviour. Most off-the-shelf chatbots fail at this and that's why people stop trusting them after the second wrong answer.
How much content does the AI need to be useful?
Depends on the use case. A site assistant covering pricing, features, and FAQ usually only needs a few dozen pages of real content to be helpful. An internal staff tool over a 200-page SOP is fine too. Where I'd push back is if you have almost no documented content and you're hoping the chatbot will paper over it. The bot is only as good as the source — it won't invent good answers from thin material, and you don't want it to.
Can a customer always reach a real person?
Yes, and the 'talk to a human' button is visible at all times by default. I won't ship support deflection that hides the human path — that's the design choice that turns chatbots from useful into hated. The bot's job is to handle the questions it can answer well and hand off cleanly with full context. Trying to trap people in the bot to inflate deflection numbers is a short-term win and a long-term reputation cost.
What does ongoing maintenance look like?
Mostly content drift. When your pricing changes, your features ship, or your policies update, the bot's answers need to keep up. For most clients that's a re-index when content changes, plus a quarterly read of the worst conversations to spot where the bot is struggling. A few hours of work per quarter from your team or me. Model costs run on the order of pennies per conversation depending on volume.
Don't see your situation here?
Most engagements start with a use-case nobody's built before. Tell Echo what you're actually trying to make happen.
Talk To Echo About Your Situation