Imagine a customer calls your shop on a Tuesday evening and asks about a coffee machine she saw on your website. The AI agent answers warmly, tells her it is in stock, and adds that there is a 15% discount until Friday. She orders. The problem: there is no discount, and the machine has been sold out for a week. Nobody lied on purpose. The system simply produced what sounded right. Now you have a customer who feels cheated, a colleague who has to apologize, and a call recording that proves exactly what was promised.
A confident guess is the most expensive mistake
Language models are built to produce likely text. Ask for a price and you can get a likely-sounding price. It will be fluent, friendly and sometimes wrong. On the phone this is worse than in a chat window. There is no screenshot to compare, the voice sounds sure of itself, and the customer decides within seconds.
A price said out loud is an offer in the customer's mind. If you honor a wrong discount, you eat the margin. If you refuse it, you create a complaint and probably lose the customer. Either way you pay. Imagine a wrong promotion quoted on just ten calls in a single day. That is ten small conflicts your team has to untangle, each one starting with "but your assistant told me".
The damage also spreads. Once a customer catches the agent on one wrong fact, she will doubt everything else it says, including the correct things. Trust in a sales conversation is built slowly and lost on one sentence.
Split the knowledge: what the AI knows, what it looks up
The fix is not a smarter guess. It is a clear division of labor. Some knowledge is stable and can live in the agent's training and instructions. Other knowledge changes daily and must never be remembered, only looked up.
Stable knowledge includes how to open a call, how to uncover what the customer actually needs, how to answer "it's too expensive" or "I need to think about it", and how to describe products in general terms. This is where an agent trained on real sales conversations earns its keep, because the craft of selling does not change with your price list.
Volatile knowledge is everything that can be different tomorrow morning: prices, active promotions, stock levels, delivery dates, order status, return eligibility. A simple test works well. Ask, "Could this answer change without anyone retraining the agent?" If yes, it must come from a live source.
How to connect the AI to your systems
You do not need a giant integration project to start. You need a few clear decisions.
Choose one source of truth
Decide where the real price, stock and promotion data lives: your shop platform, your ERP, or a product feed. If your website and your ERP disagree, the agent will inherit the mess and repeat it with great confidence. Fix the disagreement first.
Start read-only
Let the agent read prices, stock, promotion rules and order status through an API. Only later add actions that write data, such as creating an order or changing a delivery address, and even then have the agent read the details back to the customer before confirming.
Look up in real time, not from last night's export
A nightly export is fine for product descriptions. It is not fine for stock. If you sell items that sell out quickly, the agent should check at the moment the customer asks. Alphacall, for example, offers API and webhooks, so the agent can query your systems during the call and push results back afterwards, such as a new order, a callback request or the call outcome in your CRM.
Turn promotions into structured rules
"Summer sale" written as free text invites improvisation. A promotion the agent can safely use has a start date, an end date, a list of eligible products, a minimum basket if there is one, and a statement on whether it stacks with other offers. If a rule is missing, the agent should treat the promotion as unknown, not as probably true.
Treat a failed lookup as "unknown"
Systems time out. Products get entered with typos. When the lookup fails or returns something unclear, the agent needs a defined behavior, which brings us to the most underrated skill in a sales AI.
Teach it to say "I can't confirm that yet"
Good salespeople do not know every number by heart. They say, "Let me check that for you," and they do. An AI agent needs the same habit, built in on purpose.
For example, if the stock lookup fails, a good answer sounds like this: "I can't confirm availability at this moment. I can have a colleague call you back within the hour, or send you an SMS as soon as it's confirmed. Which would you prefer?" The customer stays in the process, nothing false has been said, and the sale is still alive.
Watch out for soft guesses too. Phrases like "usually", "I think" or "around Friday" are invention in disguise. If the delivery date is not in the data, the agent should not estimate it.
The same applies to concessions. Decide outside the conversation what the agent may offer, for example free shipping above a certain basket value, and keep it as a rule. Negotiating creatively on the fly is how invented discounts are born. For everything beyond the rules, such as custom quotes, exceptions and difficult complaints, the call should go to a person. With Alphacall the agent hands the call to a human together with the context: what the customer wants, what is in the basket, and which point could not be confirmed. The colleague continues the conversation instead of starting from zero.
A checklist before you go live
Run through this list with your sales, IT and customer service leads. If one line gets a "not sure", fix it before real customers call.
- Every price the agent can quote comes from a live lookup or an approved list.
- Every promotion has a start date, an end date, eligible products and clear conditions.
- Stock answers come from your system, with a defined maximum age of the data.
- A failed or unclear lookup triggers a fixed fallback sentence, never a guess.
- Maximum discounts and allowed concessions are defined as rules, not left to the conversation.
- It is clear which topics always go to a human.
- Someone owns the data. When a promotion ends, a named person switches it off.
- You have run test calls with trick questions: an expired promotion, an out-of-stock item, a price changed five minutes ago.
- Supervisors listen to calls and recordings during the first weeks and flag every doubtful sentence.
The trick-question test is the most revealing one. Change a price in a test environment, call the agent, and see whether it follows the change immediately. Then ask about a promotion that ended yesterday. If the agent hesitates and checks, you are on the right track. If it answers smoothly from memory, you have found your gap before a customer did.
Let it sell, but only with facts
None of this makes the agent less persuasive. Quite the opposite. A salesperson who is sure about the facts can focus on the customer: asking good questions, handling objections, suggesting the accessory that fits the order. The facts come from your systems, the selling skill comes from the agent, and the exceptions go to your people.
If you want to hear how this sounds in practice, ask for a live call with Alphacall. Listen to how an agent quotes a real price from a connected system, and just as importantly, how it behaves when the answer is not available.