How do you ground an AI assistant in a real product catalogue?
Stop the model inventing products by giving it only the catalogue as its source of truth: it sizes the customer's requirement, then selects from stock the business actually holds, at that market's real prices. The recommendation becomes an order the company can fulfil, rather than a plausible-sounding answer nobody can sell.
Key facts
| PARAMETER | VALUE |
|---|---|
| Product | Solar AI (SA-200) |
| Client | Joshville Global Auto & Energy Ltd |
| Grounding source | Live product catalogue |
| Pricing basis | NGN, local market |
| Surfaces | Android, iOS, Web |
A good conversation and no order
The default way to add AI to a commerce product is to put a chat window on the site and connect it to a general model. The customer describes their problem, the assistant answers fluently, and everyone agrees the demo went well.
Then you look at what it actually recommended.
For a solar business in Nigeria, a general model will confidently propose a 5kW inverter and a panel array, quote figures that look reasonable, and name brands from its training data. The problem is not that the answer is wrong in the abstract. It is that the company does not stock that inverter, the panels are not available at that specification locally, and the prices belong to a different market entirely. The customer has been given a shopping list the business cannot fill.
That is worse than no assistant, because the customer now has an expectation anchored to numbers the sales team has to walk back.
The assistant should not know anything
The fix is to stop treating the model as a source of knowledge about products.
In Solar AI, the assistant's job is narrow: understand what the customer needs, then select from what the company holds. Product identity, specification, availability and price all come from the live catalogue at request time. The model handles the conversation and the sizing logic. It does not supply a single fact about a product.
This sounds like a limitation. In commerce it is the entire value. A recommendation is only useful if it terminates in something the business can quote, deliver and install. Anything else is a conversation.
The practical test is simple: can the sales team act on the output without correcting it? If the assistant recommends a configuration, the team should be able to turn it into a quote unchanged. The moment a human has to substitute products because the AI named something unavailable, the feature has created work rather than removed it.
Pricing is the part that breaks first
Of everything a model can get wrong, price is the most damaging and the most certain to be wrong.
Solar equipment pricing in Nigeria does not resemble the figures a general model has absorbed. Import duties, exchange rate movement, local distribution and installation labour all move the number substantially. A model asked to price a system will produce something plausible and wrong, and plausible-and-wrong is harder to catch than obviously-wrong.
So prices are never generated. They are read from the catalogue in local currency at the moment of the response. If a price is missing, the correct behaviour is to say so and route to a human quote — not to estimate.
The same applies to availability. "We have this in stock" is a claim about the present, and only the inventory system can make it.
Sizing is where the model earns its place
None of this means the AI is doing trivial work.
The genuinely hard part of solar retail is that customers do not know what they need. They know their electricity is unreliable, roughly which appliances they want to keep running, and roughly what they can spend. Turning that into a load profile, and a load profile into a system specification, is the expertise that normally requires a sales engineer on a call.
That translation is what the assistant does well: asking the right follow-up questions, handling vague answers, converting "my fridge, some lights, and the TV" into an energy requirement. It is a conversation problem, which is what the model is genuinely good at.
The division holds up: the model owns the conversation and the reasoning, the catalogue owns the facts. Neither does the other's job.
Why this matters beyond the demo
An assistant grounded this way stops being a feature and becomes part of the sales pipeline. Its output is a specification tied to real stock, which flows into a quote, which flows into an order and an installation schedule.
That connection is the reason to build it at all. The alternative — a chat widget that answers questions and hands off nothing — is an expensive way to make a website feel modern.
If you are adding AI to a commerce product, the first question is not which model to use. It is: what is the source of truth this thing is allowed to speak from, and what happens when the answer is not in it?
Frequently asked
- What does it mean to ground an AI assistant?
- Grounding means the assistant answers only from a specified source of data rather than from its training. For a commerce assistant, that source is the product catalogue and its current prices, so every recommendation maps to something the business genuinely stocks and can install.
- Why not just use a general chatbot for product recommendations?
- A general chatbot will confidently recommend products the business does not sell, at prices from another market. That produces a good conversation and no order. Grounding the assistant in the live catalogue means the end of the conversation is something the sales team can actually quote.
- How do you handle local pricing in an AI recommendation?
- Prices have to come from the catalogue at request time, in local currency, rather than from the model. Solar equipment pricing in Nigeria does not resemble the figures a general model has absorbed from other markets, so any price the model produces on its own will be wrong.