Answering a commercial question, end to end

The stages#

1. Trigger — a customer asks what something costs, how fast it can be done, or what is guaranteed.
2. User or external event — the question arrives as mail, a WhatsApp message or speech on a live call.
3. Authentication and workspace resolution — the workspace is entered before any source is read, so the answer is grounded in your Knowledge and nobody else's.
4. Ingest or request — the message is canonicalised, or the call is in session with its instructions already loaded.
5. Canonical record — the question is part of the thread or the call record, attached to the Person it came from.
6. Reasoning — the answerable and the commercial parts of the question are separated. Most questions are a mixture, and the mixture is why a blanket refusal would be the wrong design.
7. Knowledge, memory and rules — grounded facts are selected within a budget of roughly 2,000 characters and four facts; a figure no source carries is refused rather than approximated.
8. Autonomy and approval — the commercial part is escalated whatever the channel's autonomy says. Autonomy governs sending, not authority over terms.
9. Action through a provider — the answerable reply goes out, or waits for approval, under the ordinary send boundary.
10. Result — the customer has an answer to what could be answered and a clear statement that a person is handling the rest.
11. Relationship, timeline and memory — the exchange lands on the person's timeline, and the deal, where one exists, is marked as waiting.
12. Audit, usage and Needs You — the refusal is recorded with the rule that produced it, and the question appears in Needs You for a person to decide.

Where the answer comes from#

Arithmetic across sources is not grounding. Deriving a bulk figure from a unit figure is invention with a calculation in front of it, and it is refused in the same way an outright guess is.

On a live call#

The turn-based carrier path produces each sentence and can check it before it is spoken. The realtime speech-to-speech path has no such seam: the model speaks to the caller directly, so the rule lives in the instructions the session opened with, and enforcement is honest about being detection rather than prevention.

• Lines that priced something are counted, using the same expression the carrier path applies to each sentence, so the two paths agree on what a price is.
• A call with a non-zero count is visible afterwards, with its transcript, and is the right thing to listen to before deciding whether that workspace belongs on the realtime path.
• The instructions can be corrected and the correction takes effect on the next call, not the current one.
• This asymmetry is documented rather than smoothed over, because a page claiming identical guarantees on both paths would be false.

The human decision, and stopping the repeat#

1. Open the item in Needs You. The customer's words, the conversation and the deal are attached to it.

   Result You are deciding a commercial matter with its context in front of you rather than reconstructing it.

2. Decide, edit the reply so it says what you decided, and approve.

   Result Your words go out through the same boundary a person's own send uses, and the state is uncertain until the provider acknowledges it.

3. Clear the review on the deal.

   Result It returns to the ordinary board; the audit keeps the fact that it was raised and who answered.

4. If the decision should hold for everyone, add it to Knowledge; if only for this customer, write it as memory against them.

   Result The next identical question is answered from a source instead of escalating — which is the only step that changes tomorrow.

Questions#

Related#

• Price, SLA and warranty: what Connect may say
• Escalating a commercial decision
• Knowledge in Connect
• Connect would not give a price
• Why a sales agent should refuse to answer
• The realtime voice engine