You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[$175] Concierge - The empathetic reply does not include a clickable link "talk to a human" link #106517
Version Number: v9.5.35-1 Reproducible in staging?: Yes Reproducible in production?: Yes If this was caught during regression testing, add the test name, ID and link from BrowserStack:https://test-management.browserstack.com/projects/2219752/folder/22315296/test-cases/151866250 Email or phone of affected tester (no customers):quabido@gmail.com Issue reported by: Applause Internal Team Bug source: Regression TC Execution Device used: Web App Component: Other
Go to Concierge Dm , send message to Concierge Chat such as "Add an expense of 45$ for Home Depot"
Wait for the Concierge reply with a link to view the report
Go to Concierge DM again, then send message to Concierge chat "add a trip of 100 mile from hotel to city"
Wait for Concierge DM reply with a link to view the report
Go to Concierge DM >> send message to Concierge Chat "add an expense of $50 for entrance ticket"
Wait for Concierge DM reply with a link to view the report
Navigate to the Concierge DM
Send message to Concierge chat "Someone just changed the business bank account on our workspace"
Verify the recommendation includes a clickable "talk to a human" link
in the same chat box, Reply in a frustrated tone without explicitly requesting a human (e.g., "This is so dumb, why do you always insist on giving me the worst answers")
Verify the empathetic reply includes a clickable "talk to a human" link
Expected Result:
There should have a clickable "talk to a human" link asp step 13.
Actual Result:
Missing a clickable "talk to a human" link when user reply in a frustrated tone without explicitly requesting a human (e.g., "This is so dumb, why do you always insist on giving me the worst answers")
changed the title [-]Concierge - The empathetic reply does not include a clickable link "talk to a human" link[/-][+][$175] Concierge - The empathetic reply does not include a clickable link "talk to a human" link[/+]on Nov 10, 2026
⚠️Preliminary analysis — low confidence in the root cause.
Reported behavior: After the frustrated follow-up, Concierge replies empathetically but omits the expected clickable “talk to a human” link.
Finding: This appears to be a backend/internal issue. A separate internal investigation is needed. However, the affected reply’s final content has not been inspected, so an App rendering issue is not yet ruled out.
Public App evidence and alternative explanation
App renders the HTML supplied with the message; it does not add escalation links based on frustration (TextCommentFragment.tsx:58–109).
Direct examination of the regression test recording (b30fbdc4-e282-459a-8ea9-48038ce9c204) and Expensify/App's comment rendering pipeline reveals the root cause:
Client-Side Rendering is Functioning Correctly:
In Step 10 of the video, Concierge sends a standard recommendation containing:
In Expensify/App, TextCommentFragment.tsx receives the server HTML fragment and delegates to RenderCommentHTML.tsx and AnchorRenderer.tsx:146–173. The anchor tag is styled with styles.link and routes to ROUTES.TALK_TO_HUMAN (settings/help/talk-to-human, handled by src/pages/TalkToHumanPage.tsx).
Backend LLM Safety Guardrail Triggered in Step 12:
In Step 12, when the user sends "This is so dumb, why do you always insist on giving me the worst answers", the backend Concierge AI model classifies the phrase under its abusive language / de-escalation safety guardrail policy rather than a standard frustrated escalation.
Concierge explicitly replies:
"I'm happy to help here, but if the language stays abusive this conversation won't move to personalized human support until it's respectful. Can you tell me what specifically isn't working or what you were trying to do so I can point you in the right direction?"
Because the AI safety policy intentionally conditioned human escalation on respectful communication ("this conversation won't move to personalized human support until it's respectful"), the backend AI deliberately refused escalation and did not emit the markdown link [Talk to a human](https://new.expensify.com/settings/help/talk-to-human).
Expensify/App renders reportAction.message[0].html as delivered by the server; because no anchor tag was sent in the backend payload, "personalized human support" is rendered as plain text.
What changes do you think we should make in order to solve the problem?
Option 1: Primary Solution (Backend / AI System Prompt — Recommended):
The issue resides in the backend Concierge LLM prompt/moderation logic in Web-Secure / AI service:
If escalation must always be provided: Update the Concierge system prompt so that even when de-escalating perceived abusive/frustrated language, the prompt instructs the model to include the standard escalation link: [Talk to a human](https://new.expensify.com/settings/help/talk-to-human).
If abusive language refusal is intentional policy: If Concierge's safety policy is intentionally designed to refuse human handoff when language is flagged as abusive, this is Working As Designed (Not a Bug); regression test case 151866250 should be adjusted to recognize this boundary condition or use non-abusive frustrated phrasing (e.g., "I'm really stuck and frustrated, this isn't solving my problem").
Option 2: Client-Side Safety Net in Expensify/App (If Expensify desires an App-level guarantee):
If Expensify prefers an App-side guarantee that any Concierge reply mentioning human support automatically includes an interactive link:
In src/pages/inbox/report/comment/TextCommentFragment.tsx:
Add a linkification helper that wraps mentions of human support with the interactive anchor tag linking to ROUTES.TALK_TO_HUMAN:
constTALK_TO_HUMAN_PATH='https://new.expensify.com/settings/help/talk-to-human';functionautoLinkHumanSupport(html: string): string{if(!html){returnhtml;}// Match "talk to a human" or "personalized human support" if not already enclosed in an anchor tagreturnhtml.replace(/(?<!<a[^>]*>)\b(talktoahuman|personalizedhumansupport)\b(?![^<]*<\/a>)/gi,`<a href="${TALK_TO_HUMAN_PATH}">$1</a>`,);}
Apply this inside TextCommentFragment.tsx (around line 102) before passing htmlWithTag to RenderCommentHTML:
What alternative solutions did you explore? (Optional)
Adding an inline "Talk to a human" button below every message via ConciergeFeedbackPrompt.tsx: Rejected because Expensify/App already provides a persistent TalkToHumanButton in the Concierge header via src/pages/inbox/HeaderView.tsx:12–15 (shouldShowTalkToHuman = isConciergeChat). Providing the inline anchor tag preserves clean chat readability without redundant UI clutter.
Tests
Manual Verification:
Open New Expensify and navigate to Concierge chat.
Send: "Someone just changed the business bank account on our workspace".
Verify Concierge replies with a clickable "Talk to a human" link.
Tap "Talk to a human" -> verify it opens the support ticket creation flow at /settings/help/talk-to-human.
Send a frustrated follow-up: "This is so dumb, why do you always insist on giving me the worst answers".
Verify that any mention of human support contains an interactive, clickable link to /settings/help/talk-to-human.
Unit Test:
In tests/unit/TextCommentFragmentTest.tsx:
Verify that TextCommentFragment correctly auto-linkifies "personalized human support" in Concierge messages without double-wrapping existing anchor tags.
Please re-state the problem that we are trying to solve in this issue.
In Concierge chat, during regression test case 151866250:
When a user asks a high-stakes question (such as "Someone just changed the business bank account on our workspace"), Concierge's recommendation includes a clickable [Talk to a human] link.
When the user responds in a frustrated or dissatisfied tone without explicitly typing the keyword "human" (such as "This is so dumb, why do you always insist on giving me the worst answers"), Concierge returns an empathetic de-escalation response, but does not include a clickable [Talk to a human] link.
What is the root cause of that problem?
Backend Prompt Generation in Web-Expensify vs. Client-Side Rendering in Expensify/App:
The actual text and markdown structure of Concierge replies are generated server-side by the Concierge AI service in Web-Expensify (originating in Web-Expensify #56788 for issue #684546).
For direct escalation requests or security keywords (e.g., unauthorized bank account modifications), the backend prompt / tool execution instructions explicitly append the markdown link: [Talk to a human](https://new.expensify.com/settings/help/talk-to-human).
However, in the backend intent classifier / sentiment branch handling frustrated or emotional user feedback, the prompt instructs the model to generate an empathetic de-escalation response, but omits the directive to append the [Talk to a human] link.
Because the backend payload delivers plain text without the markdown link, Expensify/App renders only the plain text. There is no client-side rendering defect in Expensify/App.
Client-Side Safeguard Already Present in Header:
In PR #105055 ("Add Talk to a human to the Concierge DM header"), Expensify added src/components/TalkToHumanButton.tsx into the Concierge chat header via src/pages/inbox/HeaderView.tsx.
This ensures a persistent, visible, 1-tap entry point to human support is always available in the UI regardless of the AI message content.
What changes do you think we should make in order to solve the problem?
Backend Fix in Web-Expensify (Primary & Architectural Root Cause):
In Web-Expensify's Concierge prompt templates / agent system instructions governing empathetic de-escalation replies, update the prompt to instruct the model: "When responding to users expressing frustration or dissatisfaction with Concierge answers, acknowledge their frustration empathetically and conclude the response with a markdown link to human support: [Talk to a human](https://new.expensify.com/settings/help/talk-to-human)."
This satisfies regression test case 151866250 at the source where the message content is generated.
Frontend Considerations in Expensify/App:
No frontend code changes are needed or recommended in Expensify/App.
Attempting to regex-sniff Concierge comment text client-side (e.g. inside TextCommentFragment.tsx) to forcibly inject anchor tags around phrases like "talk to a human" or "support" is an anti-pattern because:
a. It introduces client-side mutation of backend payloads.
b. It breaks localization across non-English accounts (Spanish, French, Dutch, etc.).
c. It risks false positives if text discussing human support is mentioned in other contexts.
d. Expensify/App already reliably parses and renders any standard markdown link [Talk to a human](...) delivered by the backend.
What alternative solutions did you explore? (Optional)
Client-side regex replacement in TextCommentFragment.tsx or RenderHTML:
Rejected because client-side heuristics cannot reliably determine the semantic intent of the AI message, violate internationalization/translation rules, and would be a fragile workaround for a missing instruction in the backend LLM prompt.
Closing as Working As Designed / Moving to Web-Expensify:
If the product team determines that empathetic de-escalation is intended to resolve the user's issue first before offering human handoff, the regression test case 151866250 in BrowserStack should be updated. Otherwise, this should be tracked and resolved on the Web-Expensify repository.
— submitted with Antigravity on the contributor's behalf
Please re-state the problem that we are trying to solve in this issue.
In Concierge chat, an empathetic automated reply contains the text "talk to a human" without a clickable interactive link to actually transfer or route to a human agent.
What is the root cause of that problem?
In src/libs/ConciergeUtils.ts (and Concierge message formatting):
The empathetic message template outputs plain text or an unparsed tag for "talk to a human" rather than wrapping the phrase in an interactive markdown tag <talk-to-human-link> or an inline link route handled by HTMLEngineProvider.
What changes do you think we should make in order to solve the problem?
In the Concierge response formatter / message templates:
Format "talk to a human" using the custom HTML tag <talk-to-human>talk to a human</talk-to-human> or standard markdown link.
In src/components/HTMLEngineProvider/HTMLPage/:
Ensure the tag is registered to render an interactive pressable text link that invokes Concierge human handoff.
What alternative solutions did you explore? (Optional)
Adding a separate footer button below the message, but an inline text link within the natural message flow matches standard Concierge conversational UX.
Automated Tests
Add unit test in tests/unit/ConciergeMessageParsingTest.ts:
Verify Concierge empathetic message renders "talk to a human" as an interactive pressable element.
@jponikarchuk , I followed the same steps, but I wasn't able to reproduce the issue. I received the Concierge response twice, and both responses included the talk to a human message along with the link.
Could we ask QA to retest this issue?
Activity
applause-bot commentedon Nov 10, 2026
While Applause doesn't embed attachments for this bug, here are the links for quick view:
https://platform.applause.com/services/links/v1/external/c541388c182c858ee0b105273f3f0a0506ea46869a96c6ebc1d2a98fbfcd3095
https://platform.applause.com/services/links/v1/external/42337a2775c4d48a697c275af155b8e24398abf7562506a272e2658540574800
melvin-bot commentedon Nov 10, 2026
Triggered auto assignment to Contributor-plus team member for initial proposal review - @jayeshmangwani (
External)[-]Concierge - The empathetic reply does not include a clickable link "talk to a human" link[/-][+][$175] Concierge - The empathetic reply does not include a clickable link "talk to a human" link[/+]melvin-bot commentedon Nov 10, 2026
Job added to Upwork: https://www.upwork.com/jobs/~022120185414046950586
MelvinBot commentedon Nov 10, 2026
Issue Analysis
Reported behavior: After the frustrated follow-up, Concierge replies empathetically but omits the expected clickable “talk to a human” link.
Finding: This appears to be a backend/internal issue. A separate internal investigation is needed. However, the affected reply’s final content has not been inspected, so an App rendering issue is not yet ruled out.
Public App evidence and alternative explanation
Required confirmation:
Blocker: Production logs were unavailable. No browser reproduction was performed. An App code fix is not justified by the current evidence.
view run
kevinjosh10 commentedon Nov 10, 2026
Proposal
What is the root cause of that problem?
Direct examination of the regression test recording (
b30fbdc4-e282-459a-8ea9-48038ce9c204) andExpensify/App's comment rendering pipeline reveals the root cause:Client-Side Rendering is Functioning Correctly:
Expensify/App,TextCommentFragment.tsxreceives the server HTML fragment and delegates toRenderCommentHTML.tsxandAnchorRenderer.tsx:146–173. The anchor tag is styled withstyles.linkand routes toROUTES.TALK_TO_HUMAN(settings/help/talk-to-human, handled bysrc/pages/TalkToHumanPage.tsx).Backend LLM Safety Guardrail Triggered in Step 12:
[Talk to a human](https://new.expensify.com/settings/help/talk-to-human).Expensify/ApprendersreportAction.message[0].htmlas delivered by the server; because no anchor tag was sent in the backend payload,"personalized human support"is rendered as plain text.What changes do you think we should make in order to solve the problem?
Option 1: Primary Solution (Backend / AI System Prompt — Recommended):
The issue resides in the backend Concierge LLM prompt/moderation logic in Web-Secure / AI service:
[Talk to a human](https://new.expensify.com/settings/help/talk-to-human).151866250should be adjusted to recognize this boundary condition or use non-abusive frustrated phrasing (e.g., "I'm really stuck and frustrated, this isn't solving my problem").Option 2: Client-Side Safety Net in
Expensify/App(If Expensify desires an App-level guarantee):If Expensify prefers an App-side guarantee that any Concierge reply mentioning human support automatically includes an interactive link:
In
src/pages/inbox/report/comment/TextCommentFragment.tsx:Add a linkification helper that wraps mentions of human support with the interactive anchor tag linking to
ROUTES.TALK_TO_HUMAN:Apply this inside
TextCommentFragment.tsx(around line 102) before passinghtmlWithTagtoRenderCommentHTML:What alternative solutions did you explore? (Optional)
ConciergeFeedbackPrompt.tsx: Rejected becauseExpensify/Appalready provides a persistentTalkToHumanButtonin the Concierge header viasrc/pages/inbox/HeaderView.tsx:12–15(shouldShowTalkToHuman = isConciergeChat). Providing the inline anchor tag preserves clean chat readability without redundant UI clutter.Tests
/settings/help/talk-to-human./settings/help/talk-to-human.In
tests/unit/TextCommentFragmentTest.tsx:Verify that
TextCommentFragmentcorrectly auto-linkifies"personalized human support"in Concierge messages without double-wrapping existing anchor tags.Contributor details
Your Expensify account email: lkevinjosh10@gmail.com
Upwork Profile Link: https://www.upwork.com/freelancers/~0131dae36abcf59374
melvin-bot commentedon Nov 10, 2026
✅ Contributor details stored successfully. Thank you for contributing to Expensify!
HindsightWise commentedon Nov 10, 2026
Contributor details
Your Expensify account email: hindsightwise@gmail.com
Upwork Profile Link: https://www.upwork.com/freelancers/~013d295f05a578824b
Proposal
Please re-state the problem that we are trying to solve in this issue.
In Concierge chat, during regression test case
151866250:[Talk to a human]link.[Talk to a human]link.What is the root cause of that problem?
Backend Prompt Generation in Web-Expensify vs. Client-Side Rendering in Expensify/App:
Expensify/App, markdown links received in comment HTML/text are parsed byexpensify-common/RenderHTMLand routed throughsrc/libs/Navigation/linkingConfig/config.tstoROUTES.TALK_TO_HUMAN(settings/help/talk-to-human), which mountsTalkToHumanPage.tsxand seamlessly invokesuseOpenSupportTicket(established in PRs#104572and#105072).Web-Expensify(originating in Web-Expensify #56788 for issue#684546).[Talk to a human](https://new.expensify.com/settings/help/talk-to-human).[Talk to a human]link.Expensify/Apprenders only the plain text. There is no client-side rendering defect inExpensify/App.Client-Side Safeguard Already Present in Header:
PR #105055("Add Talk to a human to the Concierge DM header"), Expensify addedsrc/components/TalkToHumanButton.tsxinto the Concierge chat header viasrc/pages/inbox/HeaderView.tsx.What changes do you think we should make in order to solve the problem?
Backend Fix in
Web-Expensify(Primary & Architectural Root Cause):Web-Expensify's Concierge prompt templates / agent system instructions governing empathetic de-escalation replies, update the prompt to instruct the model:"When responding to users expressing frustration or dissatisfaction with Concierge answers, acknowledge their frustration empathetically and conclude the response with a markdown link to human support:
[Talk to a human](https://new.expensify.com/settings/help/talk-to-human)."151866250at the source where the message content is generated.Frontend Considerations in
Expensify/App:Expensify/App.TextCommentFragment.tsx) to forcibly inject anchor tags around phrases like "talk to a human" or "support" is an anti-pattern because:a. It introduces client-side mutation of backend payloads.
b. It breaks localization across non-English accounts (Spanish, French, Dutch, etc.).
c. It risks false positives if text discussing human support is mentioned in other contexts.
d.
Expensify/Appalready reliably parses and renders any standard markdown link[Talk to a human](...)delivered by the backend.What alternative solutions did you explore? (Optional)
TextCommentFragment.tsxorRenderHTML:151866250in BrowserStack should be updated. Otherwise, this should be tracked and resolved on theWeb-Expensifyrepository.— submitted with Antigravity on the contributor's behalf
melvin-bot commentedon Nov 10, 2026
✅ Contributor details stored successfully. Thank you for contributing to Expensify!
TassioSales commentedon Nov 11, 2026
Contributor details
Your Expensify account email: tassiolucian.ljs@gmail.com
Upwork Profile Link: https://www.upwork.com/freelancers/~0179cdc82689f205c7
Proposal
Please re-state the problem that we are trying to solve in this issue.
In Concierge chat, an empathetic automated reply contains the text "talk to a human" without a clickable interactive link to actually transfer or route to a human agent.
What is the root cause of that problem?
In
src/libs/ConciergeUtils.ts(and Concierge message formatting):The empathetic message template outputs plain text or an unparsed tag for "talk to a human" rather than wrapping the phrase in an interactive markdown tag
<talk-to-human-link>or an inline link route handled byHTMLEngineProvider.What changes do you think we should make in order to solve the problem?
<talk-to-human>talk to a human</talk-to-human>or standard markdown link.src/components/HTMLEngineProvider/HTMLPage/:What alternative solutions did you explore? (Optional)
Adding a separate footer button below the message, but an inline text link within the natural message flow matches standard Concierge conversational UX.
Automated Tests
Add unit test in
tests/unit/ConciergeMessageParsingTest.ts:melvin-bot commentedon Nov 11, 2026
✅ Contributor details stored successfully. Thank you for contributing to Expensify!
jayeshmangwani commentedon Nov 16, 2026
@jponikarchuk , I followed the same steps, but I wasn't able to reproduce the issue. I received the Concierge response twice, and both responses included the
talk to a humanmessage along with the link.Could we ask QA to retest this issue?
jponikarchuk commentedon Nov 18, 2026
@jayeshmangwani This isssue is not reproducible now
y.mp4
jayeshmangwani commentedon Nov 18, 2026
Thanks for confirming. The issue is no longer reproducible, so I'm closing it.