The conversations where deflection is the expensive option

Cost per contact is the number that makes the case for automation, and it treats every contact as one unit of the same thing. In reality the value of a contact spans orders of magnitude. A question about opening hours is worth very little. A first conversation with somebody about to spend a large amount is worth a great deal. A complaint sits somewhere unusual, because its value is negative if handled badly and positive if handled well.

If the variation were random, deflecting uniformly would be fine, because you would be taking an unbiased sample off the top of the queue. It is not random. The valuable conversations are identifiable in advance by their type, which means uniform deflection is a choice to give away your best conversations at the same rate as your worst, and to do it silently.

The trap here is competence rather than incompetence. These are not cases the assistant handles badly. It will handle most of them fluently, politely and on topic, which is exactly why nobody investigates. A refusal is visible and shows up in a report. A smooth, plausible answer to a conversation that was worth a person leaves no trace at all.

The business case for putting a person on a complaint is not sentiment. It is that a complainant is a customer who is still talking to you, and that is a temporary state. The alternatives to complaining are leaving quietly and telling other people, and both of those are more expensive and neither generates a conversation you can act on.

Recovery is also cheap relative to acquisition and it is sharply time limited. The window in which a customer can be brought back is measured in hours rather than weeks, and it closes as soon as they conclude that nobody is coming. An automated reply, however well worded, reads as confirmation that nobody is coming, because that is precisely what it is.

There is a second return that gets ignored. A complaint is specific, motivated, unpaid research about your own operation, conducted by somebody with a strong incentive to be accurate about what went wrong and when. You would struggle to buy that. Routing it into an automated reply discards the information along with the customer, and the pattern across a month of complaints is usually worth more than any single recovery.

Distinguish two things that look alike. Somebody asking how cancellation works, or what the notice period is, or whether they will be charged, is asking a factual question and should get a factual answer immediately. Making that hard is user hostile and it does not retain anybody.

The other case is a person who has decided and is telling you. That conversation is worth a person for three reasons, and none of them is about persuading anyone to stay. First, the reason for leaving is the most valuable retention data your business can obtain, and it is only offered at that exact moment, unprompted, honestly, by somebody with no reason left to be polite about it. A cancellation form with a dropdown collects a category. A person collects a reason.

Second, a share of decided cancellations are reversible, and reversible specifically by somebody with the authority to do something an assistant cannot do: waive a fee, move a plan, fix the underlying problem, apologise credibly. Not most of them, and the ones that are not should be processed quickly and without obstruction. But the ones that are can only be caught by a person, and they are invisible to any process that treats a cancellation as a transaction.

Third, how somebody leaves determines what they say afterwards and whether they come back. This is the argument against the retention maze, which is the failure mode at the other extreme: an assistant instructed to deflect, delay and offer discounts before allowing anybody to leave. That converts a neutral exit into an angry one, and it is a good way to be described publicly. The person's job in this conversation is to listen and to make leaving easy, and the retention comes as a side effect of both.

The handling of this belongs elsewhere and needs care that this article is not the place for. The classification argument is simple and it is worth stating in business terms because the business terms are unusually clear.

This is the one category where the business has no upside in the conversation at all and unbounded exposure. There is nothing to sell, nothing to retain, nothing to learn, and a wide range of ways to cause harm. The downside is not proportionate to anything: a careless sentence produced automatically, under your company's name, to somebody at their worst moment, is the kind of thing that gets screenshotted and stays screenshotted.

Which is why this is the one item on this list that has to be a hard rule that fires before anything else happens, rather than a preference expressed in an instruction. Everything else here is a judgement about where value sits. This one is a decision about which conversations your business is willing to have automatically at all, and the answer should be that this is not one of them.

The uncomfortable part of this one is that your material may genuinely contain the answer, and the assistant may quote it correctly. Being right is not sufficient here, for three reasons.

The first is that the correct answer depends on facts about the person that are not in the conversation. Which product they hold, when they bought it, where they live, what they have already done, what else is going on. A person asks the clarifying question that changes the answer. An automated reply answers the question as asked, which in this category is frequently the wrong question competently answered.

The second is where the liability sits. It does not attach to the tool, it attaches to you, and a written answer given in your name is a written answer given by your business. The third is that in regulated trades the act of answering can be the breach regardless of accuracy: certain advice may only be given by certain people, and a correct answer from the wrong source is still a problem, sometimes a bigger one than a wrong answer would have been.

The version that catches people out is the question that does not look like this at all. A product question that is really about an interaction with a medication. A delivery question that is really about a contractual right. A pricing question from somebody describing a dispute. The signal is not vocabulary, it is that the right answer depends on the individual, and that is the test worth teaching whoever writes your escalation rules.

This is the strongest business case on the list and the one most often lost by accident, because a sales question sitting in a support widget looks like a support question and gets handled like one.

The arithmetic is not close. In any considered purchase, the value of one qualified conversation is large relative to the cost of a person spending fifteen minutes on it. Deflection optimises the small side of that ratio. A business that would never let a sales call ring out will happily let an assistant handle the same conversation in text because it arrived through a different box on the same website.

There are things a person does here that are not available to anything else. Judging whether this buyer is worth an exception, and making it. Hearing the sentence that reveals the opportunity is five times larger than the question implied. Asking what they are comparing you against. Deciding to bend something. An assistant can do none of these, and the ones it cannot do are the ones that decide considered purchases.

There is also the signal the channel itself sends. Somebody deciding whether to spend serious money with you is partly deciding what dealing with you will be like, and being answered by a person is direct evidence about that. The right design is not an assistant that sells. It is one that recognises the shape early, answers the factual part if it can, and hands over fast to somebody whose job this is.

None of these announce themselves. Almost nobody types the word complaint, and nobody at all writes that they are a high value prospect. They are recognisable by shape rather than by vocabulary, which means a keyword list is a starting point and never the whole mechanism.

The shapes are reasonably consistent. A complaint is past tense with a gap stated between what was expected and what happened. A decided cancellation uses the present or the perfect rather than the conditional: I am cancelling, I have cancelled, not how would I cancel. A sales conversation contains at least one of quantity, timeline, budget or a comparison. A legal or medical question is one whose correct answer changes depending on facts about the person, which is a test rather than a pattern and is the most useful line to give whoever writes the rules.

Because classification will miss cases, the safety net matters more than the rules do. Keep a visible route to a person in every conversation, not offered after three failed attempts but present from the first message. The visitor is a better classifier of their own situation than any rule you will write, and the cheapest way to catch a valuable conversation is to make it easy for the person having it to ask for somebody.

Doing all of this raises handovers and lowers deflection, and it does so on purpose. If the report is a single percentage, the report will show that your assistant got worse in the month you made it better.

Say that out loud before it happens rather than after. Whoever reads the number needs to know that five categories were deliberately routed to people, that the routing is expected to move the figure down, and why each one earns its place. Otherwise, in a quarter or two, somebody optimising the metric in good faith will take them off the list one at a time, and each removal will look like an improvement.

The better report is two lines rather than one. Conversations handled without a person, and handovers broken down by category. The second line is the one that shows the design working: complaints reaching people, sales conversations reaching people, distress firing its rule. A rising handover count in those categories is not leakage from your automation. It is the automation doing the most valuable thing it does, which is recognising the conversations that were never yours to absorb.