Contents6 sections
Field notesARTICLE
The Things an Assistant Structurally Cannot Notice
Not a list of current limitations that will close next year. A class of problem that has no message attached to it, and therefore nothing for any assistant to react to.
Costa4 min read
Short answer
An assistant reacts to messages, so it cannot see anything that is not one: a customer who stopped writing, a pattern spread across separate conversations, a policy that is still being applied correctly and has stopped being right, or anything that happened outside the chat. These need a scheduled human review and a small number of deliberate triggers, because absence never arrives as an event.
Key facts
- The assistant acts on messages; a conversation that stops produces nothing to act on.
- Recurring questions are countable across conversations, which is how a cross-conversation pattern becomes visible at all.
- Scheduled checks are configured deliberately, because silence has no natural trigger.
- Anything happening off-channel stays invisible unless a person puts it into the record.
There is a category of problem no assistant will ever catch, and it is not the category people worry about.
The worry is usually about a hard question answered badly. That is a solvable problem: you write a rule, or you route it to a person. The problems below are different in kind, because there is nothing for the assistant to be wrong about.
Non-events have no trigger
A customer wrote three times over two weeks, was going to confirm on Friday, and has now been silent for nine days.
Nothing happened. No message arrived, so nothing ran, so nothing decided anything. Every part of the system behaved correctly and the outcome is a lost deal that nobody has looked at. This is not a gap in the model, it is the shape of a system that acts when spoken to.
The pattern that is not in any one conversation
Three separate people this week mentioned a competitor by name. Nobody complained. Each conversation, read on its own, was completely normal, and each was handled well.
The information exists only in the aggregate, and the aggregate is not a conversation. The same is true of a question that got slightly harder to answer over a month, or a category of request that has been quietly declining since a page changed.
Correct, and no longer right
The most expensive version. A price is applied exactly as written, from a rule everybody approved, and it stopped making sense in July when a supplier changed terms.
Every answer is correct. Nothing fails a test, nothing triggers a handover, and there is no error anywhere in the transcripts, because the transcripts measure conformance to the rule and the rule is the thing that is wrong. An assistant will apply an out-of-date policy with exactly the same fluency as a current one, and that is the strongest argument for the review below.
What happened off the record
A customer who was short on the phone. A regular who came in and did not buy anything. A conversation at a trade fair that changed how somebody feels about you.
None of this is in the channel, so none of it is in the system, so none of it exists to any software you run. It is worth saying because businesses forget it quickly once the transcripts get good, and start treating what is in the record as what happened.
What a person still has to do
A weekly hour, and it is a short list.
The handovers, which are the documented edges of your rules. The questions nothing could answer, which is the edit list. And one query that has to be built on purpose: conversations that stopped mid-flow and never resumed. That last one is the only item here that nothing will hand you unless somebody asks for it, and it is usually where the money is.
Add to that a monthly reading of a handful of ordinary, successful conversations, specifically looking for answers that are correct and out of date. It feels like a waste of time until the first one is found.
Why this list is not a roadmap
Everything above could be described as a feature we have not built yet, and describing it that way would be dishonest.
Absence can be given a trigger, and we do that. A pattern can be counted, and Omni AI counts what it can. But the decision that nine days of silence means something, or that a competitor being mentioned three times matters, or that a correct price is now the wrong price, is a judgement about the business made from outside the conversation. That is not the model's to make, and no version of it will be.
Questions this raises
- Will better models fix this?
- Not this part. A stronger model reads a message better; none of these problems have a message. The fix is a schedule and a query, and it costs almost nothing next to the model.
- Can it flag customers who went quiet?
- Yes, if you configure it as a rule with a threshold, which makes it a scheduled job rather than a reaction. That distinction sounds pedantic and it is exactly why the feature does not appear on its own.
- So what does a person read every week?
- The handovers, the unanswered questions, and a list of conversations that stopped mid-flow. The third one is the only item on that list that nothing will produce for you unless you ask for it.