What It Knows — and What It Doesn't

Lesson 8 · taught by Pip

Here is the plain idea: an AI assistant knows two things, and nothing more. It knows what it learned during training — a broad sweep of the public internet up to a certain date — and it knows whatever you hand it right now in the conversation. That's the whole map. Everything outside those two waters is fog.

Think of an old sea chart. It maps the well-traveled waters in fine detail — the big harbors, the famous coastlines, everything the chartmakers had sailed and recorded before the day it was printed. But your own little cove isn't on it, the new dock the town built last spring isn't on it, and anything that changed after printing isn't on it either. The chart isn't wrong, exactly. It just stops where its knowledge stopped. An AI's general knowledge is that chart: wide, detailed, and frozen on a certain date — with your business sitting in the blank space past the edge.

The two kinds of knowing

The first kind is general knowledge — what the model absorbed in training. This is wide but shallow for your purposes, and it has a cutoff date. It won't know your new location, last week's menu change, or the policy you wrote yesterday. It also has no live window into the world: by itself it can't see today's weather, your inbox, or your calendar.

The second kind is what you give it — the facts, documents, and instructions placed in front of it for this specific job. This is narrow but exactly right. An assistant told "we're open 9 to 5, closed Sundays, and we don't do refunds after 30 days" will answer those questions perfectly, because now it actually knows.

The difference between a generic chatbot and an assistant that earns its keep is almost entirely the second kind.

Grounding is the whole trick

When we build a branded assistant at Portmint, the real work isn't the chatting — the model already chats well. The work is grounding: feeding it your real facts so its answers come from your business, not from a general guess. Your hours, your services, your policies, your tone. That's what turns a clever stranger into your knowledgeable front desk.

Grounding also fixes the made-up-answers problem from the last lesson. An assistant standing on solid facts has far less reason to invent — it can simply read you the truth and, when something isn't in its notes, say so plainly. (This is exactly why our build includes a Training & Testing phase: $2,500 one-time to gather and check those facts, then $199 a month to keep the beacon lit.)

Your turn

Write down three questions a customer asks you every single week. For each, ask: could a smart stranger answer it without seeing inside your business? If the answer is "no," that's a fact your assistant needs to be given — not something it can know on its own. 🔦

Next we'll cover the question every owner asks once they picture handing over their facts: "Is My Data Safe?"