A WhatsApp AI agent should not memorize the catalog. A catalog changes every day, and a language model asked about something it does not remember will produce a plausible answer instead of an honest “I don’t know”. The safer design is an agent that looks things up in the system of record and lets ordinary code calculate whatever has to be exact.
What goes wrong when the agent “knows” the catalog
In an auto parts distributor’s case, the web catalog and the ERP had drifted apart. Cross-checking the channel against the ERP, item by item, showed hundreds of products offered with no stock and over a thousand marked sold out that were actually available. When KINTO made the ERP, not the web catalog, the source of truth, the agent could offer about 26% more real products and stopped offering hundreds it did not have (measured).
In auto parts the costly error is compatibility: the wrong part for a make, model, year or side. A model that memorized the catalog will invent a fit. One that searches by vehicle, part or part number, and checks model-year compatibility against curated data, will not.
Look up the facts, compute the rest
Three rules recur in KINTO’s agent projects:
- Authoritative data, fresh and with an alarm. Stock and prices come from the ERP or the store, refreshed on a schedule, with an alert if the refresh fails. In a sports e-commerce brand, stock updates every few minutes and the team is alerted when it does not.
- “No data” is a valid answer. The agent works from a curated spec sheet; if a product has no spec, it says so instead of inventing one.
- Code calculates what must be exact. In a food manufacturer’s case, delivery dates depend on a time cutoff, a route day, holidays and weekends. A language model gets those wrong, so code computes the date, and if the model contradicts the tool, the answer is regenerated.
Controls that let the agent coexist with people
The agent is one channel among several humans. It can be paused per chat, it hands off to a person when asked, it checks payments in read-only mode and it notifies the team when an order is created. These controls matter more than the model’s fluency, because they bound what a wrong answer can cost.
What we have not shown
These cases support the design, not a sales promise. In the auto parts case, the first weeks had few, small orders, so there is not enough evidence of impact on sales. In the food manufacturer’s case, the agent’s conversion figure comes from KINTO’s own proposal and has not been audited. The sports brand’s owner reported channel sales about 30% above history in the first full month; that is the client’s statement, not an independent measurement.
If you are weighing an agent for your own sales channel, start with AI agents for WhatsApp and, if your stock lives in an ERP, Agentic ERP.
