How it works under the hood

This page explains what happens between a user writing to your organization’s WhatsApp and getting a reply. You do not need it to use the agent, but it helps understand why it is safe — and which guarantees are structural, not promises.

A message’s journey

1. Verified reception. The message arrives from Meta with a cryptographic signature that Inspow validates before processing it. Each message is processed exactly once, even if Meta retries it (retries are discarded automatically). If it is a voice note it is transcribed; if it is a photo or a PDF (for example a prescription), the product is extracted so it can be looked up in your catalog — from a medical prescription only the product names are taken, never doses or directions.

2. Identity first. Before any data, the system resolves who is writing: verified contact, pending code, or unknown. Identity is determined by the system with fixed rules — AI plays no part in verification and never decides “who” someone is.

3. Pre-checks. A battery of checks decides whether the agent may reason: valid consent (did they write BAJA?), business hours, emergency stop, AI usage cap and available conversation quota, per-contact abuse limit (if you enable it, a single number cannot consume AI without a brake), conversation state (has a person taken it over?) and, if clinical hand-off is enabled, questions about doses, symptoms or diagnoses go straight to your team. Any failing check cuts the process short with a fixed reply or an escalation, without the agent ever reasoning.

4. Reasoning with tools. Only here does the agent reason. It can use only the enabled tools: lookup tools (account status, prices, menu, attendance, available times) read real data and return results the AI merely words; tools that register something (an order, an appointment, a payment receipt) do so with system data and are visible to your team — the receipt, in addition, stays pending until a person approves it. The process has hard limits on time, steps and reply length; the history it sees is bounded (the last 20 messages, and it resets after the inactivity period you configure).

5. Output check. Before approving the text, the system verifies that every monetary figure in the reply is literally present in the query results. If the reply does not pass, the system retries a single time (with a more capable AI model when one is available); if that fails too, it blocks the send and escalates the conversation to a person. That is why the agent cannot “make up” a balance or round it.

6. Shadow or send. In shadow mode, the reply waits as a proposal for human review. In direct mode, it is sent — always within WhatsApp’s 24-hour window; outside it, free text does not go out.

Everything is traced

Every agent reply generates an activity record you can inspect in the console (Agent activity): which checks ran, which tools it queried, with what arguments and results, how long it took and what it decided. Your team can audit any reply, from any day.

The 24-hour window, explained

WhatsApp allows free-text replies during the 24 hours after the user’s last message (each message of theirs resets the clock). Those conversations are free. Outside the window only Meta-approved templates can be sent, at a low per-message cost that Meta bills directly to your organization.

In practice: if your team reviews the shadow queue at a reasonable pace, the whole operation happens inside the window and costs $0 in messaging.

The economic brake

AI spending accrues in real time against the monthly budget that Inspow sets for your organization:

• 80% → a single alert to your team.
• 100% (with hard stop) → the AI halts; the agent switches to “escalate only” and the non-AI features (verification, BAJA, confirmations) keep operating.

The same happens if your plan has a conversation quota and it runs out: the agent hands over to your team. In the statistics those conversations show up as “The assistant couldn’t reply”.

Honest limits

• The agent does not know what is not in Inspow: if the menu is not published or an FAQ does not exist, it says so and offers to escalate — it does not fill gaps with guesses.
• The attendance it reports is the cafeteria record (whether they claimed their service or not), not academic attendance, and it always says so.
• Payment receipts stay pending: someone on your team approves or rejects them, same as always.
• Orders stay awaiting confirmation until your team reviews them in Orders.