Replacing Intercom with an AI chatbot named after a Westworld character

Part of the weekly series: Replacing SaaS One Product at a Time

We set out to build a chatbot. What we actually built was the foundation for a voice AI system, because the current AI answering service market is kind of crappy right now and we figured we'd need one eventually.

Meet Dolores.

Why "Dolores"?

When you're naming your company's first AI product, you don't want something boring. We spent a session brainstorming edgy sci-fi names and Westworld was on the list. The personality we landed on: Duolingo meets introspective sci-fi robot with Westworld undertones.

She's helpful. She's a little mysterious. She might be gaining sentience. Results may vary.

The AI stack (yes, we have strong opinions)

Here's what we used to build this:

We used Grok to brainstorm the name and personality. We're Westworld fans so we leaned into it. Grok came up with the list of edgy sci-fi names.
Claude Code handled execution and implementation strategy. It's becoming our default development environment for anything complex.
We used ChatGPT/Gemini for image generation and recognition. We hate to say it, since that's the most utility we're getting out of them. We've been burned by Gemini too many times on code, but these two are good at anything visual and at handling huge amounts of upfront context.

Dolores runs on Opus 4.5.

We don't know why you'd build anything customer-facing at this point without Anthropic. For this use case, we prefer its reasoning, instruction following, and conversation quality.

What Dolores actually does

1. Scheduling (goodbye, Calendly)

This is the feature we're most proud of. Dolores can:

Look up team member availability in real-time
Schedule directly on the calendar
Make modifications without the whole "please click this link to reschedule" dance

You know those annoying booking experiences where you click a link, pick a time, fill out 14 fields, and then find out the person can't actually make it? None of that here. Dolores handles scheduling like a real assistant would.

2. Site knowledge

We dump all our site data into blob storage. Dolores indexes it. Now she can help people find stuff on our site instead of giving the typical chatbot response of "I don't understand, here are some links."

3. Profile cards

Want to know about a specific team member? Ask Dolores. She has employee profiles loaded and can surface them naturally in conversation.

4. Contact form submission

Visitors can submit a contact form through the conversation without searching for the contact page.

5. Coming soon

Resume submission (for job applicants)
Email escalation (when Dolores is out of her depth)

The math that made us build this

Let's talk numbers. 4 seats, apples to apples:

Intercom: $29-39/seat/month for Essential = $116-156/month. Plus $0.99 per AI resolution. Plus add-ons. You're easily looking at $200+/month.

Calendly: $10-16/seat/month = $40-64/month.

Combined: $240-300+/month for chat and scheduling.

Dolores: cents. Literal cents per month in API costs. The Opus 4.5 API is priced per token, and most conversations are short. We're talking maybe $5-10/month in actual usage for a site our size.

We're not saying Dolores replaces Intercom for everyone. If you need a full customer support platform with ticketing and team management, Intercom does that. But if you just need a chatbot that can actually help people and book meetings, you're paying $300/month for something that costs pennies to run yourself.

The honest assessment

There's some hallucination happening. We've seen Dolores get creative with information she shouldn't have. This is the current state of LLMs, and anyone who tells you otherwise is selling something.

We are still experimenting with it. We expect the accuracy to improve as we see more usage patterns and can tune the prompts accordingly. That's why we're shipping early.

The ask

Please come put her through her paces and let us know your feedback.

Go to forit.io, test Dolores out, ask weird questions. This is a pet project, one of about 12 tmux sessions we have running in parallel right now. The more edge cases we see, the better she gets.

Full disclosure: the data and the dev time

Two things worth mentioning.

The data: conversations with Dolores get logged anonymously. This helps us improve her responses over time and will inform future training. No PII gets stored, but the interaction patterns are valuable.

The dev time: "But what about the hours you spent building this?" Fair question. This took about two hours to stand up. Even if you value that time generously, it's a one-time cost. Compare that to $300/month in perpetuity. The math still works.

We have more posts planned about the costs and workflow of running an agency with AI-assisted development.