We Made Our AI Dumb on Purpose
Everyone's trying to stop AI hallucination with better prompts. We stopped it by keeping the AI ignorant.
Your AI is making up deal numbers. Right now. You just don't know it yet.
Not because it's broken. That's just what language models do when they hit a gap: they fill it. Confidently. Completely wrong.
For most use cases, that's annoying. In a capital raise, it's catastrophic.
The Problem
A real estate operator reaches out. They're running a $3M syndication and have been using AI to handle investor follow-up — routing inbound questions, sending updates, keeping the funnel moving.
Here's what was happening behind the scenes:
An investor asks: "What's the projected IRR on this deal?"
The AI doesn't have the number in context. So it does what it's trained to do — generates a plausible response. Maybe it pulls from something mentioned earlier in the thread. Maybe it just sounds confident and estimates.
Either way, the wrong number goes out to a potential investor. In writing.
This isn't hypothetical. It's the default behavior of every AI follow-up tool we've seen — GHL-native AI, off-the-shelf chatbots, custom GPT wrappers. Same failure mode across all of them: when the answer isn't in context, they invent one.
We don't do that.
What We Did
We built an investor concierge for a capital raise. It handles the full pipeline — Zoom invitations, deal Q&A, 1:1 call booking — with one hard constraint: it cannot share information it hasn't been explicitly given.
Sounds obvious. It's not. Getting there took two layers.
Layer 1: Data Withholding
The most important thing we did wasn't prompt engineering. It was not putting data in the prompt.
If the AI doesn't have the deal numbers, it can't make them up.
We split investors into two categories based on funnel stage. Pre-session investors — people who registered but haven't attended the deal presentation — get one thing: the Zoom link and a warm reminder. Deal metrics, projected returns, fee structures — none of that is in their context window.
Post-session investors get the full picture. Deal facts pulled from a structured dashboard. Specific numbers, specific projections, actual terms.
The AI doesn't decide what to reveal based on good behavior. The data is structurally absent until the stage unlocks it. You can't hallucinate a number you were never given.
Layer 2: Prompt Rules as Backup
Layer 1 handles most cases. Layer 2 handles the edge cases where a post-session investor asks something that isn't in the structured facts — a niche question about the submarket, a specific clause in the PPM, something from a call two months ago.
Here, the AI has two options:
Pull from the deal's knowledge vault — a searchable database of deal documents, broken into chunks, embedded locally, retrieved by hybrid vector + keyword search. If the answer is in a document, it finds it.
Say: "Great question — let me confirm that with Ben directly."
Option 2 is not a failure state. It's a feature.
We trained the team to read "I don't know, let me check" as a sign of trustworthiness, not a limitation. An AI that confidently answers everything is lying some percentage of the time. An AI that defers when it's uncertain is protecting the deal.
The Results
Across the full investor pipeline — dozens of contacts, hundreds of touchpoints — zero hallucinated financial figures went out to investors.
Not because the prompts were perfect. Because the architecture made it impossible.
Every outbound message went through operator review via Telegram before sending. Of the messages that came back, fewer than 8% were rejected. The rest were approved as-is or with light edits.
The operator's time on investor communications dropped from roughly 3 hours a week to under 30 minutes. Follow-up timing got consistent — no more forgetting to respond to an investor for 11 days. The AI runs on a cadence. Nobody falls through the cracks.
And when someone asked a number the AI wasn't cleared to share? It escalated cleanly. The operator got a Telegram notification. The investor got a human response. The deal stayed clean.
The Takeaway
Everyone's fighting hallucination with better prompts. More instructions. More "do not make up information" in the system prompt.
That works about as well as telling someone not to think about a pink elephant.
The fix is architectural. Don't give the AI data it shouldn't have. If it doesn't have it, it can't invent it.
Layer 1: Withhold data by stage. The AI is structurally ignorant of what it isn't allowed to share.
Layer 2: Give it a retrieval system for what it is allowed to know, and an honest fallback for everything else.
The result is an AI that's reliably accurate — not because it's trying hard, but because it can't fail in the ways that matter.
We use this architecture any time high-stakes information is involved. Capital raises. Deal documents. Client onboarding. Anywhere a wrong answer has real consequences.
If you're using AI somewhere a hallucinated number could cost you a deal, a client, or a lawsuit — stop asking "how do I make it more accurate?" Ask instead: "what does it actually need to know?"
Usually, the answer is a lot less than you're giving it.
If you're running a capital raise, a sales process, or any workflow where AI is touching high-stakes communications — we'd love to show you how we'd approach it.