The trust budget

The question

How much conversion are we willing to lose for calibrated trust — and how do we split it across the six levers that make us lose it?

Why it is a decision

Because the default is silence, and silence pays: “product metrics therefore reward silence”. There is no meeting in which someone chooses not to disclose. There are six different meetings in which, each time, the lever costs something and nobody has a budget to draw on. The point of this card is to turn six implicit micro-concessions into one declared allocation.

The six levers, as cost items

1. How much to disclose that a response is AI-generated.
2. When to show sources.
3. When to declare uncertainty.
4. When to require confirmation.
5. When to hand over to a human.
6. How much to explain how the system works.

Each one is a point of friction, and therefore a point of exit. The budget is split across these six: it is not decided in one block, and that is why today it is not decided at all.

What the literature says

• The five components the corpus attributes to the Trust Designer are the concrete form of levers 3, 4 and 5: delegation scale, confirmation points on the irreversible, situated uncertainty signals, traceability of actions and sources, recovery paths with human escalation.
• There is an asymmetry no interface declares today: the one between recommendation and fact. On factual accuracy the models show high performance; a list of products or suppliers, by contrast, is a probabilistic extraction that changes on every run. Trust built on facts gets transferred to recommendations, which work in an entirely different way.
• On levers 1 and 6 there is an external constraint with a date: AI Act art. 50 disclosure from 2 August 2026, marking of generated content from 2 December 2026. Which means: part of the budget will be spent anyway, by obligation. Deciding it in advance means choosing where.
• On disclosure there is a counter-intuitive effect worth keeping in mind: attributing consciousness to the system raises declared trust and lowers behavioural reliance. Surveys do not measure this decision.

The cost of either branch

Spending: friction at every step, measured on weekly metrics, without the figure that would justify the spend.

Not spending: the damage shows up late, and the only early signal — automation being switched off after a trial — comes from users who will not come back.

How it is measured, and where it is not

The benefit has four behavioural indicators already in the corpus: cancellation rate after an automatic action; frequency of external verification of the output; drop-off at confirmation points; share of users who switch automation off after trying it.

The cost has no measure. Nobody publishes the conversion delta per individual lever, and this has to be written down rather than worked around. Practical consequence: the budget can be decided and monitored, not yet balanced.

Who decides

The Trust Designer, if there is one. In most teams the role does not exist as a standalone figure, and it is one of the three figures the change requires and that are rarely there — see Who decides. Until there is one, silence decides.

By when

2 August 2026 for disclosure, 2 December 2026 for the marking of generated content. Levers 1 and 6 already have a date; the other four do not, and they stay where they are until somebody moves them.

Where it falls

Trust, AI literacy and calibration is the parent decision. Then Empathy and social presence, The regulatory perimeter, The user expects the system to act.