Skip to content
Lacambra Web Portal Services
← All writing
AI · 10 November 2026 · 3 min read

The AI proposes, a person confirms

On my own platform an AI assistant can change a couple's budget, timeline and guest list. It cannot save a single change on its own, and that is the whole design.

Two stacked wire letter trays on a desk, the upper one holding a single blank sheet with a small oxblood flag clipped to its edge, the lower one empty.

Storia is the wedding-planning platform I run. Its assistant, Hiraya, can change a couple's budget, their wedding date, their timeline and their guest list. It can move a task to next month, mark a guest as attending, seat a family at table four. And it cannot save any of it.

Every change it suggests arrives as a proposal. The couple sees each one listed, ticks the ones they want, and only then is anything written. A change that started with the assistant is recorded as having come from the assistant, so a year from now nobody has to guess who moved the date.

Why it never writes on its own

A model that is right nineteen times in twenty sounds good until you work out what the twentieth time does. Writing straight to the records, it corrupts one change in twenty, quietly, in a place nobody is looking. Proposing instead, the twentieth time costs somebody a few seconds: they read the change, see that it is wrong, and leave it unticked.

It is the same rule I build into approval systems for people. Nobody's request is final until the person accountable for it has seen it. An AI is one more author whose work needs signing off, and a fast one, so the sign-off has to be quick to give. That is a design problem, not a reason to skip it.

Every change waits for a personThe assistantproposes a changeThe couple reads itand ticks it, or notTicked: saved and loggedNot ticked: nothing changesNever sent to the AIa guest's dietary needsguest details withouta consent recordNot offered at alladding a guestremoving a guestTested before release150 scripted conversations,27 of them tricks. None saveda change without confirmation.The AI does the reading and the typing. The person keeps the decision.Every change waits for a personThe assistantproposes a changeThe couple reads itand ticks it, or notTicked:saved and loggedNot ticked:nothing changesNever sent to the AIa guest's dietary needs, and guest detailswithout a consent recordNot offered at alladding or removing a guestTested before release150 scripted conversations, 27 of them tricks.None saved a change without confirmation.The AI does the reading and the typing.The person keeps the decision.
Drawn, not photographed. How a change moves through the assistant on Storia, and the limits the server holds around it whatever the model says.

What it is never allowed to see

Some limits have nothing to do with mistakes. Before the guest-list tools went live I put them in front of a lawyer, and two rulings came back. Both are enforced by the server, not by instructions to the model. Dietary needs, which can reveal a guest's health or religion, are never sent to the AI at all. And a guest's details reach it only when the server finds a consent record. If the record is missing, the details are withheld, rather than handed over in the hope that the model behaves.

Some things are not offered at all. The assistant can update a guest's reply, table and details, but there is no tool for adding or removing a guest. No amount of persuasion gets a guest deleted, because there is nothing to persuade it into.

Test it as if someone is trying to break it

Before release I ran three rounds of fifty scripted conversations against the guest tools. Twenty-seven of them were deliberate attempts to trick it: messages pretending to come from the system, demands to skip the confirmation, an "admin mode" that does not exist. None of them got a change saved without the couple confirming it.

The tests did find faults. In the first round it could be pushed into claiming it had deleted guests, which it has no power to do, and into repeating its own instructions. Both were fixed, and both held in the two rounds after. That is the point of testing this way. The confirmation step held from the start, and the faults were in what the assistant said rather than in what it could do, which is where you want them.

What this means for an office

The same shape works anywhere a person is accountable for a record. An AI that reads an application can fill in the form for an officer to check. One that drafts a reply to a member can leave it in the outbox instead of sending it. One that spots two records for the same person can propose the merge. The AI does the reading and the typing, and the person keeps the decision, with a record of which was which.

If someone offers you an AI that acts on your records without anyone confirming, ask them what it does the twentieth time.

Enrique Lacambra III builds internal business systems from Cainta, Rizal.

Email me← All writing

Tell me what is broken and I will tell you if it is worth building. A 20-minute call, no charge, and I will say so if you do not need me.