Field notesARTICLE
What We Do Not Automate
Every assistant has a list of things it must hand to a person. Ours is short, it is written down before launch, and the reasoning behind each line is the same one sentence.
Costa4 min read
Short answer
An assistant should not decide anything whose cost of being wrong is irreversible or falls on a person rather than on the business: money leaving the account, commitments about dates outside your control, judgements about an individual, regulated advice, and an apology that carries compensation. Everything else is a candidate for automation, and the test is reversibility rather than difficulty.
Key facts
- The limits are written before launch and shipped as configuration, not discovered afterwards from transcripts.
- A request that hits a limit is handed over with the conversation attached, not answered with a refusal.
- The assistant states that a person is taking over and gives a time, rather than going quiet.
- The same list applies across chat, email and messengers, because the customer does not change when the channel does.
Buyers ask what it can do. The useful question, which almost always comes second, is what it will not do.
Here is our list. It is not long, it is written into the configuration before anyone switches anything on, and every line on it comes from one test.
The test that produces the list
Not is this hard? Difficulty is a poor guide, because the hardest sentences to write are often the safest ones to get wrong.
The test is: if this answer is wrong, who pays, and can it be taken back? A wrong opening hour costs a correction and a small apology. A wrong refund costs money that has already gone. Same model, same confidence, completely different class of mistake, and the second one does not belong to software.
Money that moves
Refunds, deposits released, discounts beyond a rule that is already written down, credit notes, cancellations with a fee waived.
The assistant can tell you what the policy says, gather what the finance person needs, and put the request in front of them with the order attached. It does not release the money. The gap between explaining a policy and executing a payment is the whole distinction, and it is worth keeping even when the amount is small, because the small amounts are how the habit of not looking gets established.
Dates you do not control
A delivery date that depends on a courier, a completion date that depends on a bank, a move-in date that depends on a landlord who has not answered since Tuesday.
The assistant reports the date the system holds and says where it came from. It does not commit on behalf of someone who has not been asked. Every business has an anecdote here, and in all of them the customer's anger came from the promise rather than from the delay.
The angry one
Not because the reply is beyond it. On a bad Friday a model writes a calmer, better-structured apology than most people manage, and that is exactly what makes this line worth defending.
A complaint is not a request for a well-worded message. It is a request for someone with authority to take it on. Automating the reply gets you a customer who received a good paragraph and still believes nobody read theirs. The assistant recognises a complaint in the first turn, says a person is taking it, and gets it there with the whole thread attached.
Judgements about a person
Who gets the tenancy. Whether a deposit is returned in full after a dispute. Whether a member stays in the community after a report. Whether someone is good for the money.
These have a subject, and the subject is a person. The reasoning behind such a decision has to be defensible by whoever made it, which means whoever made it has to be somebody.
Advice a regulator has a word for
Medical, legal, financial. The assistant does not interpret a contract clause, does not tell anyone what a symptom means, does not say whether a purchase is a good investment.
It says what it is, what the terms of your own business are, and who to ask. This one is not a judgement call; it is a compliance boundary, and it is the shortest conversation on the list.
Not on the list, and people expect it to be
Prices.
A price answer is automatable, on one condition: that your pricing is a rule and not a habit. Three real cases, a written formula, the exceptions named. If the number in your head depends on how the week has gone, that is not a limit for the assistant, it is a thing to write down first, and it is worth its own note later.
Where the line moves
Two of these have doors in them. A refund under a fixed amount, on an order that meets a written rule, with a daily log somebody actually reads, is a decision an owner is entitled to automate. A standard deposit return on a clean inspection is another.
What we will not do is move those lines quietly. A limit that gets relaxed after launch gets relaxed on purpose, in the configuration, with the person who owns the consequence knowing it happened. Across Omni AI, Voice AI and Social AI the mechanism is identical, and the list per business is not.
Most of the value in an assistant is in the ninety percent of messages that are questions with answers. The list above is how the other ten percent stops being a risk you find out about later.
Questions this raises
- Isn't a model good enough now to handle a refund request?
- It is good enough to write the message. That was never the question. The question is who is accountable for money that left the account, and the answer cannot be a configuration file.
- Can we automate an item on this list if we want to?
- Some of them, deliberately and with a ceiling. A refund under a fixed amount, on an order that meets a written rule, with a log a person reads daily is a business decision you are entitled to make. What we will not do is make it quietly.
- What does the customer see when a request hits a limit?
- One sentence saying it has gone to a named person and when they answer. What loses customers is not the handover; it is the twenty minutes of nothing that usually follows it.