Academy

From a vague message to the right structured action

Humans speak vaguely, incompletely, messily. The hard job of an intelligent system is to pull the right action out of that mess — without asking the user to speak like a form.

The Felesh team 4 min

Humans don’t speak like forms. A person’s message is often vague, incomplete, and messy: “About that thing from last week, I think we should do something.” To a human familiar with the context, this sentence means something. To a system, it is a puzzle. We believe the hard job of an intelligent system is exactly this: extracting a correct, structured action from that mess — without asking the user to speak like a form.

Why this is hard

The difficulty comes from the gap between human language and a structured action. An action requires clear fields: what, for whom, at what priority. Human language does not state these fields explicitly; it hides them in context, allusion, and assumption. Bridging that gap — not by forcing the user into structure, but by the system understanding the mess — is what makes a good experience.

Three fields hidden in one vague messageThe message “about that thing from last week, I think we should do something” is split in three phrases. “That thing from last week” hides what, in context. “We” hides for whom, in allusion. “Should do something” hides the priority, in assumption. None of the three fields is stated.About that thing from last week, I think we should do something.whatcontext:the earlier quotefor whomallusion:who is “we”?priorityassumption:how urgent?none of the three is stated; each is read from context
The three fields an action needs are not stated in the message; each one hides in context, allusion or assumption.

The path of conversion

This conversion happens in a few distinct steps. First, finding context: the system must work out what “that thing from last week” refers to and retrieve it from shared memory. Second, organising: breaking what was said into meaningful parts — a new fact, a note, a state change. Third, the stage decision: should this matter advance to the next stage? Finally, action: turning all of this into a clear, structured proposal.

The path from message to proposal, step by stepThe same message goes through four steps. Find context: that thing is supplier quote Q-204, recalled from shared memory. Organise: the fact is a one-week delivery slip, the note says the sender wants action, the state stays open. Decide stage: advance to purchase approval, yes, because the quote is still valid. Propose: reorder Q-204, 3 pallets, to purchasing, priority this week, waiting for approval.“About that thing from last week,”“I think we should do something.”find contextthat thing = supplier quote Q-204recalled from shared memoryorganisefact: delivery slipped a weeknote: wants action · state: opendecide stageadvance to purchase approval?yes: the quote is still validproposereorder Q-204 · 3 pallets to purchasingpriority this week · waits for approval
Each step turns more of the mess into structure, until the last one is a proposal a person can read.

A human in the loop

Crucially, the output of this path is not an automatic, final action; it is a proposal. The system presents what it understood in structured form, and a human reviews and approves it. This loop catches errors and builds trust: the system performs the heavy lifting of extraction, and the human remains the final judge.

A person reviews the proposal before it runsA proposal card lists what: reorder Q-204, amount 3 pallets, for: sales team, priority this week. The reviewer reads it and corrects for whom from sales team to purchasing team. After approval, the order is sent. Nothing runs before approval.proposal · to reviewwhatreorder Q-204amount3 palletsforsales teampurchasing teameditedprioritythis weekreviewerapproveorder sentonly after approval
The agent proposes in structured form, and a person corrects what it got wrong before anything runs.

Domains are only examples

This pattern is not tied to one particular domain. It could be a sales note becoming a file update, a support message becoming a structured ticket, or a vague request leading to a specific action. These are only examples. What stays constant is the pattern itself: a vague message on one side, a clear action on the other, and an intelligent path in between.

Three domains, one pathA table of three domains. A sales note, met Dara wants it by spring send numbers, becomes a file update: Dara, proposal, task send pricing. A support message, it happened again please look, becomes a delivery ticket, high priority, linked to quote Q-204. A vague request, can we do that after the holiday, becomes a reschedule of the kickoff call to the week after the break.domainvague message inclear action outsalesnote“met Dara, wants it byspring, send numbers”file: Dara · proposaltask: send pricingsupportmessage“it happened again,please look”ticket: deliveryhigh · quote Q-204vaguerequest“can we do that afterthe holiday?”reschedule: kickoffto week after breaksame path every time
Whatever the domain, a vague message goes in and an action of the same shape comes out.

Putting it together

We believe the mark of a truly intelligent system is not answering clear questions, but understanding vague messages. The user should not have to speak like a machine to be understood; it’s the system that should understand human language. And it’s right here that the gap between a rigid tool and a real colleague shows itself.