diff --git a/src/api/config.js b/src/api/config.js
index ede5219..9c62cb0 100644
--- a/src/api/config.js
+++ b/src/api/config.js
@@ -52,6 +52,11 @@ export const STAGES = [
   // policy nobody paid for. It gets its own band on the overview so the role
   // that owns it does not have to find it in a list of things mostly handled
   // by somebody else.
+  // WHERE THE AI PUTS DOWN WHAT IT CANNOT HONESTLY FINISH. An AI employee has
+  // no way to raise a ticket and no way to send a message; the only move it
+  // has for "a person should take this" is Flag for Review, and it lands here
+  // with the reason it wrote. Nothing is running on these leads.
+  { uid: 'zk-state-review',    name: 'Needs a Person',           kind: 'needs',    need: 'Flagged by the AI',                by: 'operations' },
   { uid: 'zk-state-payment',   name: 'Payment Pending',          kind: 'needs',    need: 'Premium confirmation',             by: 'operations', approval: 'ops_admin', approvalLabel: 'Premium awaiting your confirmation', approvalNote: 'Cover cannot start until the premium is received — section 64VB. Only operations can confirm it.' },
   { uid: 'zk-state-issued',    name: 'Policy Issued',            kind: 'auto',     doing: 'Closing the file',                by: 'Engage AI' },
   // Nurture. There is no scheduler yet, so Resume Outreach is a button and it
@@ -160,6 +165,10 @@ const STATE_ACTIVITIES = {
     { uid: 'zk-act-uw-decline',    label: 'Decline the risk',            by: 'you',              role: 'do' },
     { uid: 'zk-act-uw-prepare',    label: 'Prepare Referral',            by: 'Referral AI',      role: 'force' },
   ],
+  'zk-state-review': [
+    { uid: 'zk-act-review-resume', label: 'Send back to the workflow',  by: 'you',              role: 'do' },
+    { uid: 'zk-act-review-refer',  label: 'Send to underwriting',       by: 'you',              role: 'do' },
+  ],
   'zk-state-payment': [
     { uid: 'zk-act-realise',       label: 'Confirm premium received',    by: 'you',              role: 'do' },
     // Not offered: the chase is a SCHEDULED activity now (86), booked 24h
diff --git a/src/screens/Lead.jsx b/src/screens/Lead.jsx
index f42f71e..655ff00 100644
--- a/src/screens/Lead.jsx
+++ b/src/screens/Lead.jsx
@@ -45,8 +45,22 @@ const MONEY = new Set(['quoted_premium','quoted_od_premium','quoted_tp_premium',
   'sme_burglary_si','sme_ee_si','sme_bi_gross_profit','sme_claims_36m_amount','sme_stock_si'])
 
 /** Field groups, in the order the lead was actually worked. */
+/* What the AI said stopped it, turned into a sentence an operator can act on.
+   Keyed by the stored value so an unknown one degrades to the generic line. */
+const BLOCKER = {
+  missing_evidence:   'It says the evidence it needed was not there.',
+  contradictory_data: 'It says the evidence it had contradicts itself.',
+  tool_failed:        'It says a lookup or a tool it depends on failed.',
+  outside_my_remit:   'It says the judgement is a person\u2019s to make, not its own.',
+  other:              'It could not complete its step.',
+}
+
 const GROUPS = [
   ['Source',        ['lead_ref','product_line','source_channel','partner_code','partner_branch','rm_or_agent_id','consent_artefact']],
+  // First, because when a lead is flagged this is the only thing anyone opens
+  // it to read. Empty on every lead that was never flagged, and an empty group
+  // does not render.
+  ['Flagged for review', ['review_blocker','review_reason','review_return','review_note']],
   ['Customer',       ['customer_name','entity_name','mobile','email','pan','gstin','udyam_no']],
   ['Intake',         ['lead_score','attribution_status','attribution_reason','dedupe_match_ref','eligibility_outcome','eligibility_reason']],
   ['Contact',        ['contact_outcome','outreach_window','contact_notes','call_transcript']],
@@ -430,6 +444,31 @@ export default function Lead() {
           </div>
         ) : null}
       </div>
+    ) : stage?.uid === 'zk-state-review' ? (
+      /* THE AI PUT THIS DOWN. It could not finish its step honestly and said
+         so in the workflow, which is the only way it has of reaching a person.
+         Its own words lead, because deciding where the lead resumes is a
+         judgement made on them. */
+      <div className="attn" role="status">
+        <div>
+          <h3><span className="dot dot--amber" />The AI stopped here and asked for a person</h3>
+          {row.review_reason ? <p className="attn__said">“{row.review_reason}”</p> : null}
+          <p>
+            {BLOCKER[row.review_blocker] ?? 'It could not complete its step.'}{' '}
+            Nothing is running on this lead. Read the trail, fix what is missing, then send it
+            back to the workflow at the right step — or hand it to an underwriter.
+          </p>
+        </div>
+        <div className="attn__acts">
+          {actions.filter((a) => a.role === 'do').map((a) => (
+            <button key={a.uid} type="button"
+                    className={'btn ' + (a.uid === 'zk-act-review-resume' ? 'btn--navy' : 'btn--ghost')}
+                    onClick={() => { setOpen(a.uid); goToAction() }}>
+              {a.label}
+            </button>
+          ))}
+        </div>
+      </div>
     ) : primaryAction ? (
       <div className="attn" role="status">
         <div>
diff --git a/src/screens/screens.css b/src/screens/screens.css
index 1f97571..0e96415 100644
--- a/src/screens/screens.css
+++ b/src/screens/screens.css
@@ -388,3 +388,12 @@
   .kpis { grid-template-columns: repeat(2, 1fr); row-gap: 16px; }
   .facts { grid-template-columns: repeat(2, 1fr); row-gap: 16px; }
 }
+
+/* The AI's own words, when it flagged a lead. Quoted, because deciding what
+   to do next is a judgement made on the sentence it wrote. */
+.attn__said {
+  margin: 8px 0 0 !important; padding: 8px 14px;
+  border-left: 2px solid var(--zk-amber-bar); background: #fff;
+  border-radius: 0 var(--r-sm) var(--r-sm) 0;
+  font-size: var(--fs-md) !important; color: var(--zk-ink) !important; max-width: 78ch;
+}