{/* Prominent agent CTA — the way out of the manual toil */}
{remaining} items waiting. Reviewing them one at a time is slow.
The Review Agent scans all 8 categories at once, then triages what's left into three clear buckets.
Search by contract reference or contractor name
{items.length} in {catLabel(cat)}
{/* Transparent learning prompt — the agent asks before remembering anything */}
{learnPrompt && (
You've approved {patternCount[learnPrompt.patternKey]} {learnPrompt.pattern} in a row. Want me to handle these for you from now on?
I only ever auto-handle low-risk items, every rule shows up in Playbooks so you can see exactly what I've learned, and you can switch it off in one click. Nothing is learned silently.
)}
{items.length === 0 && (
All {catLabel(cat)} items cleared. Try another tab, or run the agent.
)}
{items.map(it => (
{it.who}{it.line} · #{it.ref}{it.amount}
))}
Severity is the agent's risk read. Only LOW items can ever be auto-handled. See setView("trust")}>Trust levels.
You've resolved sub-$5 currency-rounding expenses the same way. Add this as a rule?
Most teams auto-approve these; kept manual for you.
)}
{bucketLabel} · {counts[bucket]}
{bucket === "auto"
?
I checked these against your past decisions and your Trust Levels and found nothing to flag. Open a group to see why, or undo to send any of them back to the queue.
);
}
function PayoffStrip() {
const prevented = resolvedCount;
const mins = resolvedCount * 6;
return (
{Metric({ n: , label: "Issues prevented" })}
{Sep()}{Metric({ n: , label: "Value protected" })}
{Sep()}{Metric({ n: <>m, label: "Time saved" })}
See the full vision →
);
}
function Metric({ n, label }) { return
{n}
{label}
; }
function Sep() { return ; }
// ===== TRUST LEVELS (risk-tier × autonomy) =====
function TrustView() {
const tierMeta = {
low: { label: "Low risk", icon: "ShieldCheck" },
med: { label: "Medium risk", icon: "ShieldAlert" },
high: { label: "High risk", icon: "OctagonAlert" },
};
function setTier(t, v) {
setAutonomy(a => ({ ...a, [t]: v }));
if (t === "low" && v === "auto") flash("Low-risk items now auto-handled, with undo");
}
const sigRows = [
{ key: "frequency", title: "Approvals lower risk", desc: "If you've approved a pattern 3+ times, treat it as low-risk. This is how the agent learns what you trust.", star: true },
{ key: "novelty", title: "Novelty raises risk", desc: "First-time payees and documents are never auto-handled." },
{ key: "baseline", title: "Deviation from baseline", desc: "Flag anything that departs from the worker's usual amount or cadence." },
];
return (
{BackToReview()}
Trust levels
Two questions, both yours to answer: what counts as risky, and how much the agent may do at each risk level. Human-in-the-loop by default · reversible · audit-logged · never auto-pays.
{/* AXIS 1 — risk signals (editable) */}
1 · What counts as risky
You own the signals. Transparent and adjustable, not a black box. They define which items land in each tier.
Above this $-at-risk, never low-riskAED {signals.amount.toLocaleString("en-US")}
Set Low risk → Auto to watch the agent clear its low-risk items into "Auto-handled" (reversible). Then head back to the queue.
);
}
// ===== PLAYBOOKS — Rules + Reviewer Profile (memory) =====
function PlaybooksView() {
function saveEdit(i) {
setProfile(p => p.map((x, j) => j === i ? { ...x, text: editText || x.text, edited: true } : x));
setEditIdx(-1); setEditText(""); flash("Profile updated");
}
function addCustomRule() {
const t = newRule.trim();
if (!t) return;
setPlaybooks(p => [{ name: t, scope: "Your own rule", auto: "you", on: true, system: false, custom: true }, ...p]);
setProfile(pr => [{ text: t, source: "a rule you wrote", edited: true }, ...pr]);
setNewRule("");
// Adding your own rule is the final guided step — completing it closes the tour.
if (tourOn) { setTourOn(false); flash("That's the full walkthrough — you're all set. Rule added; the agent will follow it on its next run."); }
else { flash("Rule added. The agent will follow it on its next run."); }
}
return (
{BackToReview()}
Playbooks
Self-learning · you stay in control
Your decisions teach the agent. It keeps a plain-English profile of how you work, plus the structured rules that follow from it. All visible, all editable.
{/* Reviewer Profile (memory) */}
Reviewer Profilememory
I built this from your decisions. Edit anything. Nothing is learned silently, and I'll never act outside the Trust Levels you set.