Your AI Support Bot Shouldn't Cost More Than Your Support Team

Here's a conversation that happens at every growing startup:

"We should add an AI chatbot to our site."

"Great, what's the budget?"

"Well... it depends on how many customers use it."

That last line is the problem. Every AI support solution today ties your costs to your success. More customers, more queries, more spend. The tool that was supposed to reduce your support burden becomes a line item that grows faster than your revenue.

We built Kanha to fix this.

---

The math that breaks

We had an independent cost analysis done across three real scenarios. Here's what it looks like.

Small site: 50 pages, 1,000 queries/month

At small scale, everyone's roughly comparable, $29 to $99. Except the per-resolution players who are already bleeding money. If you're building it yourself with an API, the tokens are cheap but you're still paying for hosting, a RAG pipeline, monitoring, and error handling. Realistic floor: $150/month even for a light-touch setup.

Even here, Kanha offers something the others don't: real privacy, zero vendor lock-in, and a model trained on your data; not a generic LLM with your FAQ pasted into a system prompt.

Moderate site: 500 pages, 10,000 queries/month

Now it gets interesting. The per-resolution tools are catastrophic at this volume, one popular product charges $0.99 per resolved conversation, and at 10K queries that's $9,900/month. The credit-based builders look fine until you realize "credits" aren't queries, premium model usage can 20x the effective per-message cost.

DIY with an API looks cheaper on paper, but you're now maintaining real infrastructure and every customer question is flowing through OpenAI or Anthropic's servers. More on why that's a problem below.

Large site: 2,000 pages, 50,000 queries/month

At scale, the story is stark. The per-resolution model is now costing you more than a junior support engineer. The credit-based builders start hitting overages. And DIY? The token cost looks manageable, but at 50K queries/month you need autoscaling, caching, vector DB, monitoring; a permanent engineering commitment that pulls your team away from building your actual product.

The line that never goes up

Here's the real chart. This is monthly cost as query volume scales:

Kanha's line is flat. Every other approach scales linearly, or worse, with usage.

Why DIY isn't actually cheaper

The DIY column looks competitive on raw cost. But those numbers only tell part of the story. "Build it yourself" means:

• 2-4 weeks to set up a RAG pipeline (chunking strategy, vector DB, embedding model, retrieval tuning)
• Ongoing hosting: A server that's up 24/7, autoscaling if traffic spikes, monitoring for when it goes down at 2 AM
• Maintenance: Every time your docs change, someone needs to re-index. Every time OpenAI or Anthropic changes their API, someone needs to update your integration
• Guardrails: Prompt injection protection, hallucination detection, fallback handling. All on you
• Privacy exposure: Every customer query goes through a third-party API. You're trusting their data retention policies, not yours

Factor in the engineering salary, the maintenance burden, and the opportunity cost of not shipping your actual product and DIY is the most expensive option on the list. It just doesn't show up on the infrastructure bill.

Kanha gives you the same result, a trained bot on your site, in 20 minutes. No infra, no maintenance, no ongoing engineering cost. And if you're operating at serious scale, our Enterprise plan gives you custom page limits, dedicated support, and SLAs, without hiring a single engineer to maintain it.

At 100K+ queries/month, a business saves $1,000+/month vs a typical chatbot builder, $14K+/month vs premium platforms, and $98K+/month vs per-resolution tools. By choosing Kanha.

---

How this actually works

No magic. Just a different architecture.

Step 1: You point Kanha at your site. We crawl your pages, product descriptions, FAQs, return policies, size guides, whatever you've got. We generate question-answer training pairs from your actual content.

Step 2: We train a small language model specifically on your data. This isn't a generic chatbot with a system prompt pasted in. It's a model that has learned your business.

Step 3: You add our SDK to your site; a script tag, a React component, or a Web Component. Three lines of code.

Step 4: When a customer opens the chat, the model loads into their browser and runs on their device via WebGPU. By default, responses are generated locally. No round-trip to a server. No API meter running.

The result: your bot knows your products, speaks in your voice, and costs you nothing per conversation.

---

Why "fine-tuned" matters more than "RAG"

Most AI support tools use retrieval-augmented generation (RAG); they search your docs at query time and paste relevant chunks into a prompt for GPT-4. This works, but it has problems:

1. Every query hits an LLM API. That's the per-query cost you can't escape.
2. Retrieval can miss. If the chunking is wrong or the question is phrased unusually, the bot gets the wrong context and hallucinates confidently.
3. You don't control the model. You're renting someone else's intelligence and hoping the prompt engineering holds.

Kanha fine-tunes a model on your data. The knowledge is in the weights, not in a prompt. The model doesn't need to "find" the answer at query time, it already knows it. And because the model is small and specialized, it runs fast enough to work on a laptop GPU.

---

The privacy thing (it's not just marketing)

When your chatbot runs through an API, every customer question hits a third-party server. "What's the return policy for the dress I bought last Tuesday?", that's going through someone else's infrastructure.

With Kanha, the model runs in the browser by default. The question stays on the customer's device, no third-party API in the loop. For most use cases, there's nothing to log, nothing to leak, nothing to worry about.

Down the road, we plan to offer an optional server-side mode for businesses that want query analytics, understanding what customers are asking, spotting trends, identifying gaps in their knowledge base. But that'll be opt-in. The default is on-device, private, and zero-cost per query.

For industries like healthcare, finance, or legal, where data handling is regulated, on-device inference is a strong default. You control the data story, not your chatbot vendor.

---

What it costs, no surprises

You saw the comparison above. Here's our pricing:

Extra pages are $0.10 each. Extra retrains are $15. That's it. No per-query fees, no credit system, no "tokens" that cost more when you use a better model.

Need more? Our Enterprise plan gives you custom page and training limits, priority support, and SLAs tailored to your scale. Get in touch, we'll build a plan around your needs.

The SDK auto-selects model size based on the visitor's device — a customer on a MacBook gets the largest model, someone on a budget Android gets the smallest. You don't configure this. It just works.

The free tier gives you 50 pages and 1 training run, no credit card, no trial timer, no "sandbox mode." This isn't a playground. It's a production bot that serves every visitor on your site. Train it on your FAQ, shipping policy, return policy, and a few dozen product pages. Embed it. Your customers start getting instant answers.

Most businesses start here and upgrade when they want to retrain regularly or index more pages as their site grows. But the free bot is the real thing, same fine-tuned model, same on-device inference, same zero per-query cost. It's not a trial. It's the product.

---

A real example

We trained a bot on a fashion brand's site — 47 product pages, a size guide, shipping policy, and FAQ. Total: 52 pages. Took about 8 minutes to crawl, 3 minutes to generate training data, and 12 minutes to train.

The bot correctly answers:

• "Do you have the silk wrap dress in green?" → Yes, checks the product page, knows available colors
• "What's your return window for sale items?" → Different from regular items, the bot learned the policy distinction
• "I'm 5'6 and usually wear a US 6, what size should I get in the oversized blazer?" → Uses the size guide context to recommend

These aren't retrieval lookups. The model knows the catalog. It handles follow-up questions, remembers context within the conversation, and doesn't hallucinate product features that don't exist.

Cost to the brand: $39/month (Starter plan). Cost per customer conversation: $0.

---

Who this is for

Growing DTC brands with a product catalog, FAQ, and size guides. You're getting 100+ support tickets a week about things that are already on your site. A Kanha bot handles 60-80% of those instantly.

SaaS companies with help docs and knowledge bases. Your users ask the same 50 questions. Train a bot on your docs, embed it in your app, and let it handle L1 support while your team focuses on real issues.

Agencies managing multiple client sites. Each client gets their own bot, trained on their content. The Business tier at $299 gives you 10,000 pages and 20 trainings a month across all your bots. Need more? Enterprise scales with you.

Anyone tired of per-query pricing. If you've built a chatbot on OpenAI's API and watched the bill climb every month, you know the problem. Kanha makes that bill flat.

---

Try it now

Go to kanha.ai, sign up, point it at your site. You get a real, production-ready bot trained on up to 50 of your pages; free, no credit card. Your customers can use it today.

If you're a fashion brand or DTC startup reading this: we've probably already trained a demo bot on your site. Check our showcase or reach out at contact@kanha.ai and we'll send you a working demo.

The support bot that gets cheaper as you grow. That's Kanha.