What should an automated calling agent never do to your customer?

The limits are decisions, not accidents

Everything an automated calling agent does to a customer was decided by someone before the phone rang. That includes the things it does badly. An agent that claims to be a person claims it because nobody wrote the rule that it must not; an agent that marks a mumbled reply as consent does so because the confirmation threshold was set by whoever wanted the confirmation rate to look good.

So this is not a list of risks to be managed. It is a list of four decisions, each of which has a right answer, and each of which is visible on the very first call if you get it wrong.

1. It must never claim to be a human being

An agent should sound like a careful, unhurried person. That is a quality bar, and pursuing it is legitimate — a flat, robotic voice makes people hang up, and a customer who hangs up has not confirmed anything.

Sounding like a person and claiming to be one are different things. The rule is narrow and absolute: if the customer asks whether they are talking to a machine, the agent says yes.

It does not deflect, it does not answer a different question, and it does not carry on as if it did not hear. It answers, and then it carries on with the call — because in practice most people who ask do not mind, they were just checking. The customers who do mind are precisely the ones who deserve a straight answer.

The reason to hold this line even when it costs a confirmation is that the alternative is a system whose success depends on the customer not finding out how it works. That is not a position to build an operation on, and it is not a position to defend to a customer who works it out mid-call.

2. It must never over-read consent

A confirmation is the agent asserting that the customer agreed to this parcel, this amount, this address. It is the basis on which you spend money on fulfilment, and it is the record you would point at if the customer later said they never ordered.

That means the threshold has to be agreement, not the absence of disagreement. In particular:

• Silence is not consent. If the agent asks and hears nothing, that is not a yes.
• A hesitant noise at the end of a call the customer clearly wants to end is not consent.
• "I'll think about it" is not consent, and neither is "acha, dekhi".
• Answering a different question is not consent. A customer who confirms their address has not confirmed the order.

When the agent is not certain, the correct outcome is an escalation to a person or an explicit UNRESOLVED, never a confirmation. That is why six outcomes exist rather than two: a system with only "confirmed" and "not confirmed" pushes every ambiguous call towards the flag that looks better on a dashboard.

There is a hard commercial argument for this and not only an ethical one. A false confirmation is the most expensive error the system can make. It sends a parcel that will come back, and it does so with a record claiming the customer asked for it — so nobody investigates until the returns pile up. A missed confirmation costs you one retry. The asymmetry is enormous, and the agent's thresholds should reflect it.

3. It must never talk over an interruption

When a customer starts speaking, the agent stops. Immediately, mid-word if necessary.

This sounds like a politeness feature. It is a correctness feature. People interrupt for a reason, and on a confirmation call the reason is almost always that something the agent just said was wrong — the wrong size, the wrong total, the wrong street. An agent that finishes its sentence has talked over the single most valuable turn in the call.

It is also the property that most reliably marks a call as automated regardless of how good the voice is. Humans yield. A voice that keeps going while you are speaking is unmistakable, and the customer's next move is to stop listening.

Two related rules follow from the same principle. The agent should not repeat a block of speech the customer has already responded to, and it should not read out a long list before establishing that the person on the line is the right person. Both waste the customer's attention on a call that needs it.

4. It must never dial at an hour you would not dial yourself

Calling hours are the part of an automated system that scales badly in the wrong direction: one person making calls stops at a reasonable time because they are a person, and a scheduler does not stop at all unless someone tells it to.

The rule is simple to state — call during the hours you would be comfortable calling a customer personally, in their timezone, not yours. It is worth writing down explicitly, because the default behaviour of an untended queue is to drain as fast as it can, and a customer called at half past ten at night about a shirt will remember it.

Retries deserve the same restraint. A number that did not answer twice does not want to answer a third time in the next ten minutes. Space the attempts, cap them, and treat a persistent UNRESOLVED as information rather than as a target to be defeated.

What the agent must never do about itself

Two smaller boundaries, which are really the same rule as the first one applied to the record.

It must never mark a call as something it was not. A dropped line is a dropped line. Audio that was too poor to understand is UNRESOLVED, not a cancellation. The value of a structured outcome is entirely dependent on the outcome being true.

It must never improvise outside its subject. A confirmation agent knows about one order. It does not know your refund policy, your stock, your courier's schedule or what happened to someone's previous delivery. Asked about any of those, the right move is to escalate to a person with the transcript attached, not to produce a plausible answer. A voice agent with no escalation path will confidently answer questions it has no business answering, which is a much worse failure than admitting it does not know.

Why this is published rather than implied

Every item above is a constraint that makes a confirmation rate look worse in the short term. That is the point. A system optimised purely for confirmations converges on exactly the behaviours in this list, and it does so without anyone deciding to be dishonest — the metric simply rewards them.

Nishchit publishes no confirmation-rate benchmark for the same reason. There is no figure we could publish that would not immediately become a target to be hit by loosening a threshold. The calculator on this site works only from figures you enter yourself, and what Nishchit does not do is written down as plainly as what it does.