A ChatGPT trained on YOUR business, embedded where your customers already are.

What "trained on your data" actually means

When someone says they want a "ChatGPT for our business," they usually mean one of three different things — and the difference matters enormously for cost, risk, and outcome.

Approach	What it means	When it's right
Fine-tuned model	Adjusting a base model's weights using your data	Specialized writing style or jargon at high volume; rarely the right answer for business knowledge
Prompt-engineered assistant	A standard model with your content stuffed into the system prompt	Tiny knowledge bases; fragile as content grows
RAG (Retrieval-Augmented Generation)	Your content lives in a vector database; relevant chunks are retrieved and given to the model per query	The right answer for ~95% of business AI assistant use cases

Custom AI Assistant Integration is almost always the third option — RAG. It's the right architecture for nearly every "AI assistant trained on my business" use case, because:

• Your content updates anytime, without re-training
• The model cites its sources, so users can verify
• Hallucination is dramatically lower because answers are grounded in your actual content
• Your data is not used to train the LLM provider's models
• Costs scale predictably with usage, not with content volume

The architecture of a Custom AI Assistant

A production-grade Custom AI Assistant has five components:

1. Ingestion pipeline

Your content sources — documentation, knowledge base articles, FAQs, policy PDFs, product specs, support tickets — are pulled, cleaned, chunked into semantic units, and embedded as high-dimensional vectors. This pipeline runs whenever your content updates, so the assistant always reflects current truth.

2. Vector store

A specialized database that stores the embedded chunks and returns the most semantically relevant ones for any query in milliseconds. Common choices: pgvector (Postgres extension, self-hosted), Pinecone (managed), Qdrant, Weaviate. The right pick depends on your scale, your hosting preferences, and your data privacy requirements.

3. Retrieval logic

When a user asks a question, the query is embedded, the most relevant chunks are pulled from the vector store, and they're ranked. Smart retrieval is what separates a good Custom AI Assistant from a frustrating one — naive retrieval pulls plausible-but-wrong context, which leads to plausible-but-wrong answers.

4. Assistant layer

The retrieved context is passed to the LLM (OpenAI, Claude, or whichever fits your privacy and quality needs) along with a carefully engineered system prompt that defines the assistant's tone, scope, guardrails, and behavior when it doesn't know.

5. Channel integration

The assistant is exposed where your users already are: a web widget on your site, a Slack bot, a WhatsApp Business number, an in-product chat. One channel is included in the standard Custom AI Assistant Integration engagement; additional channels scope upward.

Guardrails that production assistants need

A serious Custom AI Assistant doesn't just generate text. It has:

• Scope guardrails — instructions that keep the assistant from answering off-topic or out-of-policy questions
• Uncertainty marking — when retrieval confidence is low, the assistant says "I don't have a confident answer for this" rather than fabricating
• Source citations — every answer references the page or document it pulled from, so users can verify
• PII handling — if you have user-data privacy requirements, the assistant is configured to refuse or redact accordingly
• Eval harness — a small test suite of representative questions with expected answers, so quality doesn't regress when the underlying model or retrieval changes

These aren't optional in a production Custom AI Assistant. They're the difference between an assistant your team trusts and one your team has to babysit.

Where your Custom AI Assistant lives — channel options

The standard Custom AI Assistant Integration engagement covers one channel. The most common:

• Website widget — a floating chat bubble or full-page chat embedded on your marketing site, documentation, or app
• Slack integration — a bot answering inside your team's Slack workspace, often used for internal employee assistants
• WhatsApp Business — a number your customers can message directly; common for support and sales in MENA and emerging markets
• In-product chat — embedded inside your existing SaaS or app, often for in-product help and onboarding

Additional channels can be added as scope grows. The retrieval and assistant layers are shared across channels; only the channel-specific integration code differs.

Data privacy: where things actually live

This question comes up on every discovery call. Plain-English answers:

Will my data be used to train OpenAI's or Anthropic's models? No, not when used through their business API tiers under standard terms.

Where is my content stored? Your choice. Two common paths:

• Self-hosted — your content lives in a vector store on your infrastructure. Nothing leaves your environment except per-query API calls to the LLM provider.
• Managed — your content lives in a managed vector service like Pinecone. Faster to set up, simpler to operate, but your content sits with a third-party provider under their terms.

We pick the right path together during discovery, based on your privacy posture and what your data actually contains.

What it costs to run a Custom AI Assistant after launch

Two cost components, both predictable:

1. LLM API costs — these scale with usage (number of conversations and tokens per conversation). Typical SMB traffic ranges from $50 to $300 per month. High-volume use cases scale linearly; we estimate this on the discovery call from your expected traffic.
2. Infrastructure costs — vector store, small backend, hosting. Typically $20 to $100 per month for SMB scale. Lower if self-hosted on infrastructure you already pay for.

The total monthly cost for a typical SMB Custom AI Assistant is in the $80 to $400 range. There are no per-conversation hidden fees from the integration itself.

When Custom AI Assistant Integration is the wrong product

• You expect a flawless human-level support agent on day one. Even the best RAG assistants need a few weeks of tuning before they handle the long tail well. Plan for the tuning window.
• You can't or won't share source content. The R in RAG isn't optional.
• You're in a heavily regulated space and any hallucination is catastrophic. RAG dramatically reduces hallucination but doesn't eliminate it. Plan for human-in-the-loop review.
• Your "knowledge base" is people's heads, not documents. That's an Audit problem first.

The standard engagement, in one paragraph

Two weeks. One channel. One knowledge domain. Full source code, full ownership, two weeks of post-launch tuning, and a 60-day bug-fix window. Quoted from $2,000 on the discovery call, with additional channels and domains scoped on top. The Custom AI Assistant you build with me is yours forever — including the right to swap LLM providers, host on different infrastructure, or extend it without me.

Quick answers to the questions every buyer asks

Is RAG the same as fine-tuning? No. RAG keeps the model generic and feeds it your content per query. Fine-tuning permanently changes the model's weights. RAG is almost always the better fit for business knowledge assistants.

Will the assistant work in Arabic? Yes — modern LLMs handle Arabic well. We can configure the assistant for Arabic-only, English-only, or bilingual responses.

Can I see one of your assistants live? Yes — the chat widget on this site is itself a Custom AI Assistant Integration trained on this site's content. Try asking it about a service.

Ready to give your business its own AI assistant?

Book a free 20-minute Discovery Call to scope the right Custom AI Assistant Integration for your situation, or send a brief about your project and I'll respond within 24 hours.