An AI agent that tells a customer there are 40 units in stock when only 3 are left doesn't have a model problem. It has a tools problem. And the fix most people try first, stuffing more context into the prompt (the whole catalog pasted in, the full price list, the pricing manual), almost never works, because it treats the wrong symptom.
It's a mix-up we see constantly when we build AI agents inside a client's real system: "it makes things up" gets treated as a problem of how much information you gave it, when it's actually a problem of what kind of information you gave it.
The problem: confusing "more context" with "more accuracy"
When an AI agent gets a price, a stock number, or a customer condition wrong, the instinctive reaction is the same in almost every company: "it's missing information, let's give it more." The pricing PDF, the stock spreadsheet, the commercial terms document all get pasted into the prompt. Sometimes it helps a bit. It almost never fixes the problem fully, and as soon as that data changes (a new price list, an order that empties a warehouse), the agent fails the exact same way again, because it's still answering with the context you gave it the day you set it up, not with today's reality.
The real problem isn't the amount of information, it's the type of information:
- Context is text the agent carries along in the conversation or its configuration: who you are, how you talk, what policy to apply, what tone to use. It doesn't change (or barely changes) from one day to the next.
- A tool is a function the agent can call at the exact moment it's answering, to check a piece of data that does change: the stock right now, this customer's price today, whether an invoice is already paid.
Putting a constantly changing piece of data into context is promising the agent something you're going to stop being true the moment that data moves. And in a real business, stock, price, and customer status move all the time.
Context and tools aren't interchangeable
The table below sums up the difference with examples any small business will recognize:
| Data | Does it change often? | How to give it to the agent |
|---|---|---|
| Brand tone, how to greet, what to never say | No | Context (in the prompt or its configuration) |
| The product catalog and its descriptions | Rarely | Context, reviewed periodically |
| Available stock of a product | Yes, constantly | A tool that checks the ERP at that moment |
| The price that applies to a specific customer | Yes, by customer and by date | A tool that checks the real system |
| Whether an invoice is paid or pending | Yes | A tool, never memorized |
| The legal terms of a standard contract | No | Context |
The decision rule is this: if a piece of data could have changed between when you configured the agent and the moment it answers, it isn't context, it's a query. Giving it as context isn't a reasonable shortcut, it's programming the failure for the day that data moves, which in an active business is almost always soon.
Real case: an agent that doesn't memorize stock, it checks it
At an industrial distributor with over 50,000 products in its catalog, we built an AI salesperson that quotes over WhatsApp inside the actual ERP (full case at AI salesperson inside the ERP). When a customer asks about a product, the agent doesn't answer with a stock number it "remembers": it validates the customer, applies their specific price, checks real stock at that instant, and, if it makes sense, creates the opportunity and the order inside Odoo.
The difference from pasting the catalog into the prompt isn't subtle. With 50,000 products moving through orders, returns, and warehouse entries all day, any "snapshot" of stock we gave it as context would be outdated within minutes. The only way for the agent to answer something true is for it to ask the real system at the moment it answers, instead of reciting what we told it yesterday.
The same logic applies to the case of invoices that book themselves: the AI doesn't memorize which suppliers exist or what taxes apply, it checks the real accounting system at the moment it creates each invoice, which is exactly why it can catch when something doesn't add up and delete its own invoice if the total doesn't match.
How to tell if your agent has this problem
Before building (or accepting from a vendor) an AI agent that answers with data about your business, run this exercise on any piece of data that could give a wrong answer:
- Could this data have changed since the last time we "told it" to the agent? If yes, it needs a tool, not more context.
- Does the agent check the real system at the moment it answers, or repeat something we pasted into a document? Ask your vendor to show you, not promise you.
- What happens if the data changed five minutes ago? If the answer is still correct, you have a real tool. If it's still yesterday's answer, you have context disguised as live data.
- What does it cost when it's wrong? A wrong price or an invented stock number isn't a cosmetic bug: it's an order that can't be fulfilled or a customer who was promised something that doesn't exist.
- How long would it take to add the missing tool? Usually less work than it looks: the query to the real system typically already exists, it just needs to be connected to the agent.
The lesson for any business
When an AI agent gets a piece of changing data wrong, the fix is almost never a longer prompt. It's identifying which data in your business actually moves and giving the agent a real way to check it at the moment it needs it, instead of asking it to memorize it. It's the same discipline we apply to every project at AutoBoost: first we sort out what data lives where and how often it changes, and only then decide whether the agent carries it as context or checks it with a tool.
If you have an AI agent (your own or a vendor's) that sometimes "makes up" a price, a stock number, or a status, the problem is probably not the model. At AutoBoost we build AI agents integrated into the client's real system, with tools that check live data instead of memorizing it. Get in touch and we'll look at where yours is going wrong.
