The models propose.
A named person decides.
Institutions are now asked to account for every automated system that touches a person. This is the account for Yves: what the models are allowed to do, what they are structurally prevented from doing, what leaves your tenant, and where the whole arrangement sits under the AI regimes you answer to.
1
Human gate on the only route
No AI proposal becomes a real row without a named approval, audited every time.
0
Automated decisions about a person
Yves ranks messages and drafts responses. It does not judge individuals.
Text only
Reaches the model provider
Message body alone, contracted out of training. Nothing else leaves.
Beside every figure
Confidence interval
Wilson intervals, with thin samples marked thin rather than rounded up.
An assistant with no authority, by construction.
The governing question for any AI regime is what the system decides on its own. Yves decides nothing about a person. It translates, it clusters, it scores urgency, it flags manipulation, and it proposes. Every route from a proposal to a consequence runs through a person with a name, and that gate is audited on every pass.
That is not a policy we could relax under pressure to ship faster. The approval gate is the only path in the data model, so an unapproved proposal has nowhere to go. Enforcement itself is deterministic logic rather than a model, which keeps latency flat and keeps the prompt-injection surface off your request path.
The result is a system that stays firmly outside the highest-risk categories in the AI Act and its equivalents, and we would rather earn that by architecture than argue for it by classification.
Where this sits under the AI Act and its neighbours
The EU AI Act classifies by use rather than by technology, and the obligations that bite hardest attach to systems making consequential decisions about people. Yves is a classification and prioritisation layer feeding a human decision, and we document it as such rather than leaving your compliance team to infer it.
- Transparency: the portal tells senders in their own language that automated processing is applied to their message
- Human oversight: a named approval gates the only route from proposal to consequence, and every pass is audited
- Accuracy and robustness: the confidence scoreboard tracks how often a stated confidence matched the outcome
- Record keeping: trace-ID reconstruction rebuilds the full chain of events and agent runs behind any request
- Data governance: your taxonomy, your retention, your corpus, exportable in open formats at any time
- Where your own assessment requires it, we complete your AI or algorithmic impact template alongside the DPIA
One boundary, stated exactly
Model governance questions usually reduce to a single one: what did you send to a third party. The answer here is short, and it is the same answer we give a regulator.
- Message body text alone reaches the language model provider
- Contracted to exclude your data from model training, with the clauses recorded in the agreement
- Account data, device fingerprints and console activity never leave the platform
- United States by default, relocatable on Dedicated, or replaced entirely with your own contracted keys
- Any incident at the provider is notified to you inside seventy-two hours
A model that states its own uncertainty
The failure mode that damages an institution is not an AI that is sometimes wrong. It is an AI that is wrong with total confidence, in a board pack, with nothing beside the number to warn anyone. Yves is built to make that specific failure hard.
- A confidence interval printed beside every figure, computed with Wilson intervals
- Sample size stated next to every claim, and a trend held back until the data carries it
- Every category compared against its own baseline rather than against your largest site
- A public scoreboard tracking how often a stated confidence matched the outcome
- An integrity attestation on every evidence pack, stating what was excluded and on what signal
What the models do, and what they are not permitted to touch
01
Translate, from any of forty languages
The sender writes in their own language and your leadership reads it in yours, with the original held verbatim beside the translation so nothing rests on a paraphrase.
02
Classify against your taxonomy
Your categories, drafted with you before go-live and owned by you. Where the tail grows, the platform proposes additions from the cluster shapes it sees, and a person accepts or rejects each one.
03
Score urgency, independently of volume
Ensures critical single-incident safety disclosures are surfaced ahead of high-volume repetitive complaints. Urgency scores order triage queues for operator review and never trigger automated actions autonomously.
04
Flag manipulation, reversibly
Twelve documented detection rules acting on shape rather than identity. Every catch is bounded, recorded before it applies, and reversible. Quarantine never deletes.
05
Draft a response, for a person to send
Grounded directly in submission context and translated into the respondent's native dialect. Delivered as an editable draft requiring explicit authorization from a designated human operator before transmission.
06
What it never does
It does not identify a sender, score an individual, profile a member of staff, make an employment or clinical decision, or take an enforcement action without a human approval recorded in the audit trail.
“Agents propose. People decide. The gate is in the data model rather than in the handbook, which is the only version of that sentence worth anything.”
Yves platform team
Asked often enough to answer here
Bring your AI assessment.
We will fill it in.
Send the template your institution uses. It comes back completed, with the architecture references attached.