An AI agent for WhatsApp is software that answers customers there, looks up prices, stock and dates in the company's own systems, and hands off to a person when needed.
What it does, in order, and what it never does.
Connections
What the system never does
KINTO builds these agents for companies that already sell or support over WhatsApp and are losing sales to slow replies, wrong answers or unattended hours.
Many companies close most of their sales inside WhatsApp, and the weak point is the chat itself. In a sports e-commerce case, order automation worked well, but replies were slow, sizes were wrong, sold-out items were offered and some orders never reached the store. In a real estate case, WhatsApp support depended on the office.
A generic chatbot makes this worse. If it memorizes the catalog, it invents products, prices and dates. A cross-check in the auto parts case found hundreds of products offered with no stock and over a thousand marked as sold out that were in stock.
KINTO starts from the conversations the company already has, not from a script. In the real estate case that meant analyzing more than a hundred thousand messages from more than four years. Each error, including each lost sale, becomes a rule that can be tested.
The agent does not hold the catalog in memory. It looks up price, stock and availability in the ERP or the store at the moment of the question. What must be exact, such as delivery dates, is computed by code, not by the model. Before each release, KINTO runs simulated customer scenarios, and the result decides whether the version ships.
The same approach extends to voice: in the food manufacturer case, KINTO also built an outbound AI calling pilot on the same lookups.
Companies with a catalog or a service sold over WhatsApp, an ERP or store that holds the real data, and a team that answers by hand. KINTO reports what each case measured and what it does not yet know: the sports e-commerce owner reported channel sales about 30% above their history in the first full month, and the agent converted about 9 to 10% of conversations in its first weeks, but the effect of the agent separate from seasonality is not isolated.
Headless e-commerce
By making the ERP, not the web catalog, the source of truth, KINTO got an auto parts distributor’s WhatsApp agent to offer ≈26% more real products and stop offering hundreds it did not have.
WhatsApp & voice agents
By analyzing a year of invoices, KINTO found that fewer than half of a food manufacturer’s customers drove almost all of its volume; we built a production WhatsApp agent, an AI calling pilot and a headless store that kept every URL.
Agentic ERP
KINTO gave a real estate firm locked inside a sector ERP an agent layer that reads the ERP, shows lease by lease the status of each document, and writes only when a person approves.
WhatsApp & voice agents
KINTO fixed the weak link of an e-commerce brand that sells over WhatsApp with a multi-model audited agent and order-by-order reconciliation; the owner reported a month with channel sales ≈30% above their history.
To turn years of WhatsApp conversations into knowledge an AI agent can trust, classify them with a closed taxonomy, transcribe voice notes, and admit a fact only if it appears in at least two distinct, recent conversations.
A WhatsApp AI agent should not memorize the catalog because memory goes stale and models fill gaps with plausible inventions. It should look up products, stock and prices in the system of record, and let code compute dates and totals.
It can if it memorizes the catalog, which is why KINTO does not build it that way. The agent queries the company's ERP or store for each answer. In an auto parts case, making the ERP the source of truth let the agent offer about 26% more real products and stop offering hundreds that did not exist.
Code. Dates depend on cutoff times, routes and holidays, and a language model gets them wrong. In a food manufacturer's agent, code computes the delivery date, and if the model contradicts the tool, the answer is regenerated. Prices come from the official source, never from the model's memory.
KINTO turns real conversations, including lost sales, into testable rules and runs simulated customer scenarios before every release. In the food manufacturer case that meant dozens of scenarios per release; in the auto parts case, hundreds of automated agent tests. Failures block the release instead of reaching a customer.
When the request falls outside what it can verify: a payment it cannot check, a complaint, an exception to policy. The agent alerts the team and passes the conversation with its context. In the auto parts case it checks payments but cannot change them, so a person always decides money matters.
A health monitor alerts the team. In a real estate case KINTO added a monitor that warns when the channel cannot operate even if every service looks healthy. In the auto parts case, an email alarm fires if the team stops receiving alerts.
KINTO designs agents so they do not need to. Price and stock are looked up in the system of record, delivery dates are calculated by code, and when a product has no data the agent says so. This reduces the risk, but it does not remove it, which is why agents are tested with simulated conversations before each release.
Tell KINTO what is going on. The first conversation is to understand it, not to sell.
