Why Your AI Agents Keep Making Stuff Up

And how to fix it with one MCP server

January 13, 2026

Your AI agents are guessing. They don't have access to your actual docs, policies, past conversations, or customer data. So they make things up.

I've been building AI agents for e-commerce brands for over a year now. Every single time we hook up an agent to a brand's real data, the quality jump is massive. The agent goes from confidently wrong to actually useful.

But getting that data into a format AI can use? That's been the hard part.

Until now.

Introducing kb

We built kb to solve this exact problem.

It's a vector knowledge base that gives your AI agents real knowledge. You connect to it via MCP. Link your docs, messages, reviews - whatever. Your AI now knows what your team knows.

Instead of hallucinations and guessing, you’ll get accurate answers grounded in your actual business data.

Three Ways to Feed Your Knowledge Base

Looking under the hood, kb gives you three powerful ways to add knowledge:

1. Documents

Upload PDFs, Word docs, RTF files, or plain text. The system handles everything automatically.

Got a 200-page product manual? It gets chunked into searchable pieces. Got a scanned PDF that's basically images? The system detects garbled text and switches to OCR mode automatically. No manual intervention needed.

Supported formats: PDF, DOCX, DOC, RTF, TXT

Real use cases:

  • Brand guidelines and style guides

  • Product spec sheets and manuals

  • HR policies and employee handbooks

  • SOPs and operational procedures

  • Training materials

2. Web Pages

Paste any URL and the system scrapes it, extracts the content, and vectorizes it. Works with JavaScript-heavy pages too.

This is huge for competitive intel. Vectorize competitor product pages, industry articles, or your own help center. The AI can then answer questions by pulling from all of it.

Real use cases:

  • Your entire help center

  • Competitor product pages

  • Industry research and articles

  • News mentions and PR coverage

  • Documentation sites

3. Raw Content (Messages, Reviews, Feedback)

This is where it gets interesting for e-commerce.

You can feed in any text content with metadata: Slack messages, emails, customer reviews, social comments, support tickets, creator feedback. Each piece gets tagged with its source type so you can filter later.

Real use cases:

  • Customer reviews from your site and third-party retailers

  • Support ticket history

  • Slack conversations and decisions

  • Email threads with suppliers or partners

  • Social media comments and mentions

  • Influencer feedback and communications

How Vector Search Actually Works

Traditional search matches keywords. That's why searching your Google Drive for "refund process" fails when the doc says "return policy."

Vector search understands meaning, not just words.

When you add content to kb, it gets converted into numerical representations of meaning. So when someone asks "how do I get my money back?" it finds your return policy doc even though those exact words aren't in it.

The system also handles deduplication automatically. Upload the same document twice? It detects the existing content and replaces it instead of creating duplicates. Smart chunking breaks content at natural boundaries (sentences and paragraphs) so you don't get weird cut-off answers.

What This Unlocks for E-Commerce Brands

Customer Support

Your support team answers the same questions over and over. Return policies. Shipping times. Product comparisons. Troubleshooting steps.

With kb, an AI agent can instantly retrieve the correct answer from your actual policy docs. Not a guess based on what it thinks your policy might be. The real answer.

Even better - feed in your past support tickets. "What did this customer complain about last time?" Answered instantly. Your support agent (human or AI) now has full context before responding.

One brand we work with embedded their entire FAQ, help center, and two years of support tickets. Their AI chatbot went from "maybe helpful" to "actually solves problems."

Marketing & Brand

Your brand guidelines live in a PDF somewhere. Past campaigns live in random Google Docs. Customer reviews are scattered across platforms.

Put it all in kb.

Vectorize your brand guidelines document. Scrape your help center pages. Feed in customer reviews and social comments as raw content.

Now when your marketing team (or AI assistant) asks "what's our brand voice?" - it pulls from the actual guidelines. When you ask "what did customers say about our last product launch?" - it summarizes real reviews and social comments.

Product Development

All the feature requests, complaints, and ideas customers share - in reviews, support tickets, social media - it's gold. But nobody has time to read through thousands of comments.

Feed it all into kb as raw content. Tag each piece with its source (review, ticket, social). Then ask: "What improvements are customers asking for most?"

The AI retrieves the actual feedback. Not a vague summary. Specific quotes and patterns from real customers.

Internal Knowledge

Here's where most companies are bleeding time.

Your team wastes hours hunting through Slack threads, Google Docs, Notion pages, and emails. "What was decided in that meeting?" "Who owns this project?" "What's our policy on X?"

According to Gartner, 47% of employees don't even use their company's knowledge base because the search is so bad.

Vector search fixes this. It understands context. It finds what you actually need.

Vectorize your key documents. Feed in important Slack conversations as raw content. Scrape your internal wiki pages. Now anyone can ask questions and get real answers.

The MCP Connection

Here's why this matters for the agentic future.

MCP (Model Context Protocol) is becoming the standard for how AI agents connect to data sources. I've written about it a lot. Anthropic created it. OpenAI adopted it. Google adopted it. It's not going away.

kb gives you an MCP endpoint. That means it works with Claude, ChatGPT, n8n, and any other MCP-compatible tool.

You're not locked into one AI provider. Your knowledge base works everywhere.

And as you build more sophisticated agents - for CRO, for retention, for ad creation - they can all tap into the same knowledge source. One setup, everywhere.

Getting Started

Setup takes about five minutes.

  1. Sign up at kb.run

  2. Add mcp.kb.run to Claude, ChatGPT, or n8n

  3. Start adding knowledge:

    • Link your Google Drive for documents

    • Paste URLs for web pages

    • Feed in reviews, tickets, and messages

  4. Start asking questions

The Bottom Line

Every brand I talk to is trying to figure out AI agents. Most are stuck because their agents don't have access to real knowledge.

This is the foundation. Before you build sophisticated workflows, your AI needs to know what your team knows.

Ready to give your AI agents real knowledge?

Start free at kb.run - 1,000 free vectors, no credit card required.

Questions? Reply to this email or DM Bora on LinkedIn.