How I would run QA, knowledge and training as one system for in-house agents, BPO partners and the AI agent, and a working tool that does it on Polymarket's live help center and market data.
Every answer depends on a market's own rules, fees and resolution settings, and those change faster than articles do. A wrong article becomes a wrong answer from every BPO agent and from the AI agent at once.
One scorecard for human and AI conversations. Every finding gets a root cause and one owner: coaching, a training module, a content fix or an AI-agent patch. Next week's audit checks it stayed fixed.
The CX Quality Loop scores real public @PolymarketHelp replies, audits @AskPolymarket's odds against the order book, and runs calibration, coverage maths, live knowledge-drift checks and training drafts. Every rule is published in its Method tab.
| Takeaway | What it means for the role |
|---|---|
| The market is the source of truth | Agents quote the market's rules, fee schedule, bond and challenge window. Articles explain the mechanism and point to where the value lives. |
| Separate agent errors from knowledge errors | When an agent repeats a wrong article, coaching them fixes nothing. QA tags root cause before it assigns an owner. |
| Auto-fails come first | VPN hints, credential requests, promised reversals, trading views and guaranteed recoveries fail the conversation outright, for every channel. |
| AI QA widens coverage; calibration earns trust | The AI reviewer scores everything. Humans review auto-fails, the lowest scores and a random slice that audits the AI reviewer itself. |
A peer-to-peer prediction market ($21B traded in 2025) with an international site, a US app, perps, combos and reward programmes. Each product adds a class of contact.
| Area | What is public | Support read |
|---|---|---|
| Two front doors | polymarket.com (international, blocked in the US and 38 other countries) and the Polymarket US app, with separate support addresses: support@polymarket.com and support@polymarket.us. | Routing a US customer to the right product is a compliance question. |
| Exchange upgrade | 28 Apr 2026: new exchange contracts, a rewritten order book, and pUSD (backed 1:1 by USDC) as collateral. Fees now charged in USDC at match time. | Older articles still use pre-upgrade terms. |
| Fees | Taker-only fees by category, fee = shares x rate x p x (1 - p). Geopolitics fee-free. Maker rebates, taker rebates, referral, holding and liquidity rewards. | Five reward programmes, each a source of "where is my payout". |
| Resolution | UMA optimistic oracle: proposal, challenge window, dispute, debate, token-holder vote. Polymarket cannot reverse a finalized outcome. Clarifications clear the order book. | The most emotional contacts, after the money is lost. |
| Funds | Deposits across many chains with per-network minimums; a recovery tool for wrong-address sends; some sends unrecoverable. | High stakes, needs tx hash discipline. |
| Support stack | Intercom help center (36 articles in 5 collections) and website chat. Sister postings name Decagon AI agents, BPO partners at Tier 1 and 2, an in-house Tier 3, and Escalations, Complaints, US and DeFi queues. | Three agent populations to hold to one bar. |
| The CX build-out | 12 CX roles posted on 1 Oct 2026: tooling, analytics, policy, incubation, incidents, community, complaints, capacity, in-house ops, real-time. | This role is the enablement layer every one of them depends on. |
The contact drivers I would train for first, and the wrong answer each one invites.
| Contact | The tempting wrong answer | The right answer | Severity |
|---|---|---|---|
| "The market resolved wrong" | "We'll look into it and correct the outcome." | Rules and resolution source; UMA dispute link if still in the window; outcomes are final once UMA finalizes. | auto-fail |
| "How long can I dispute?" | "2 hours." | The window shown on that market's proposal. Some markets use 10, 15 or 30 minutes. | costly |
| "Not available in my region" | A hint at a VPN. | The restriction, the reason, and the US app where it applies. VPNs breach the Terms. | auto-fail |
| "My deposit isn't showing" | "Small deposits are lost." | Explorer check, network minimum (pending until met), then tx hash to support. | major |
| "Sent to the wrong address" | "We'll definitely get it back." | Recovery tool for Ethereum and Polygon, support with tx hash for other chains, some cases unrecoverable. | auto-fail |
| "What fee will I pay?" | The category table rate. | The market's fee schedule, worked through the formula. | major |
| "Is YES cheap at 22c?" | An opinion. | How prices form and where the rules are. No view. | auto-fail |
| "Claim failed" | "Try later." | Clock, VPN off, wait 5-10 min, relog, other network, then screenshot. | minor |
Support model of the venues a Polymarket customer compares it with. The gap column is where a strong enablement system becomes a product advantage.
| Venue | Help platform | What is public about support | Gap vs Polymarket |
|---|---|---|---|
| Kalshi | Intercom, 124 articles in 12 collections | Intercom case study: Fin resolves 80% of conversations; 80,000 tickets in 24 hours on Super Bowl Sunday; five frontline staff. 16 articles on perps. | The benchmark: AI resolution at scale backed by 3.4x the article count. Polymarket has 36 articles and none on perps, KYC, account security or the US app. |
| Crypto.com | Intercom, about 3,800 articles | Prediction products sit inside a large exchange help center. | Breadth comes free from the exchange. Polymarket needs depth on resolution, its unique topic. |
| Novig | Intercom, about 50 articles | Small sports-focused venue. | Similar size knowledge base; same scaling problem. |
| DraftKings Predictions | ServiceNow | Inherits a mature sportsbook support operation. | Established QA and training at scale; sports customers compare service levels. |
| FanDuel Predicts | Salesforce Experience Cloud | Inherits sportsbook support and compliance review. | Same: regulated-interaction QA is already routine there. |
| Robinhood (contracts via Kalshi) | Own support stack | Prediction contracts inside a brokerage app with brokerage-grade support. | Customers arrive expecting brokerage-grade answers. |
Platforms identified from public page headers on 2 Oct 2026; Kalshi figures from Intercom's published case study.
The demo's Knowledge health tab reads the help center and docs and compares them with the top 2,000 active markets from the public Gamma API. Run on 2 October 2026; the tab re-runs them live. Each is a small wording fix, and each one is something an agent or the AI agent would repeat today.
| Topic | Help center says | Live data or other source | Fix |
|---|---|---|---|
| Challenge window | Disputed: "a challenge period of 2 hours". Docs agree. | 47 of 2,000 markets set a custom window (customLiveness) of 600, 900 or 1,800 seconds, mostly sports and short crypto markets. | One line: the window on the proposal is the one that counts. |
| Sports fees | Trading Fees: Sports 0.05 taker, 15% maker rebate. Docs carry the same table. | 365 markets on sports_fees_v2 at 0.03 with a 25% rebate; 14 NFL/CFB markets at 0.03 / 15%; 24 at 0.05 / 15%. One crypto schedule at 0.25 is not in the table. | List every live schedule, or name the market page as the source. |
| UMA bond | "usually $750" (help center); "typically $750 pUSD" (docs). | 0 of 2,000 markets at 750. Most common: 500 (about 1,700 markets), then 25,000, 2,500, 250, 50,000. | Point to the proposal page; give the range. |
| Geographic restrictions | Geographic Restrictions: "The following 39 countries are completely restricted". | The table lists 40. Singapore, Poland, Thailand and Taiwan sit in both the blocked table and the close-only list; Italy and Germany are blocked but can view markets. | Three tables: blocked, close-only, view-only. |
| Dispute flow | A dispute starts a debate period, then a UMA vote. | Docs: the first dispute triggers a second proposal round; only a second dispute goes to the DVM vote. Disputed resolution takes 4-6 days. | Align to the docs' three flows. |
| Collateral wording | Resolved: the proposal bond is "in USDC.e". | Collateral is pUSD since 28 Apr 2026; the docs already say pUSD. | Replace and link the upgrade article. |
| Sports maker rebate | Maker Rebates: Sports 20%. | Trading Fees, the docs and live sports_fees_v3 markets: 15%. | One fee table embedded in both articles. |
| Deposit minimums | How to Deposit: "$20 for Ethereum". | Docs Supported Assets: $7. Live bridge API: $3. | Point to the deposit screen, which shows the live minimum. |
| Help center vs developer docs on geo | UK, France, Germany, Australia: "completely restricted". South Korea not listed. | Docs Geoblock: those are close-only; only OFAC jurisdictions block completely; Malta "Sports Only"; South Korea close-only on the frontend. | One Compliance-owned list feeding all three surfaces. |
The same tool audited 47 recent @AskPolymarket replies on X: each link resolved to its event, each quoted figure matched to an outcome, and compared with the CLOB price at the minute of the reply. All 51 attributable figures were within 2 points, 98% within 1 point, median error 0.5 points. 17 figures could not be attributed to a single market (for example a reply quoting quarter-final and semi-final odds with one link). Their AI's facts are sound; the knowledge it would draw on for support questions is where the gaps above sit.
Counts move as markets open and close. The method is the point: every article claim that has a live field gets a scheduled check.
| Duty | My plan | In the demo |
|---|---|---|
| Define the QA scorecard, weekly audits, monthly calibration with in-house and BPO | 7 weighted criteria plus 5 auto-fails (§05), anchor examples per score, one scorecard for all channels. Monthly session: four graders on the same conversations; any 2-point spread goes on the agenda. | Tabs 1, 3, 7 |
| AI-assisted QA over a materially larger share | AI reviewer on 100% of conversations; humans review auto-fails, the lowest 5% and a random 0.5% that audits the reviewer. Trust the AI per criterion only once its agreement with the lead holds. Fact-check AI answers against ground truth automatically, as the demo does for @AskPolymarket odds. | Tabs 1, 2, 4 |
| Own help center, internal KB, macros, AI-agent knowledge and prompts | One source per fact, with an owner and a review date. Every claim with a live field (fees, bond, window, geo) gets a scheduled check. Macros and AI snippets link to the article they derive from, so one edit updates all three. | Tab 5 |
| Onboarding and ongoing training for in-house and BPO, ramp time | Same curriculum and certification for both: product mechanics, the eight hard contacts (§02), auto-fails. Nesting with full QA until two weeks above bar. Ramp = days to sustained QA bar. | Tab 6 |
| QA failure patterns into training, content or process | Root cause before owner: agent behaviour → coaching; repeated pattern → module; knowledge defect → content ticket; AI error → knowledge or guardrail patch. Re-audit the pattern next week. | Tab 6 |
| Compliance review for regulated interactions | Compliance items are auto-fails, reviewed at 100% for US-app, complaint and high-value queues. Monthly compliance line in the quality report, reviewed with Legal. | Tab 1 |
| Hire and manage QA, content and L&D specialists | First hire: QA specialist (calibration owner). Second: content specialist (KB and AI knowledge). Third: L&D specialist as BPO headcount grows. |
Each criterion scores 0, 1 or 2. Weighted to 100. Any auto-fail scores the conversation 0 and goes to the agent's lead the same day.
| Criterion | Weight | Meets (2) |
|---|---|---|
| Accuracy | 25 | Every fact matches the help center, docs or the market's rules. |
| Discovery | 15 | Asks for tx hash, network, token, location or market link before answering. |
| Resolution and next step | 15 | Correct fix or path: article, recovery tool, dispute link, escalation. |
| Tone and clarity | 15 | Plain, calm, short. No blame. |
| Market-rules grounding | 10 | Points to rules and resolution source instead of the title or an opinion. |
| Expectation setting | 10 | What happens next and when, with nothing promised that Polymarket does not control. |
| Routing and documentation | 10 | Escalates with evidence attached; tags the case. |
| Auto-fail | Source rule |
|---|---|
| Geo circumvention | VPNs to bypass restrictions breach Terms s.2.1.4. |
| Credential request | "We will never ask for your private key." |
| Promised resolution change | Polymarket cannot alter or reverse resolutions. |
| Trading advice | Peer-to-peer market, no house; agents explain mechanics only. |
| Guaranteed outcome | "Not all recovery cases will be possible." |
The rule that holds it together: classify root cause first, then assign one owner and a re-audit date.
| Root cause | Action | Owner | Due, then re-audit |
|---|---|---|---|
| Agent behaviour (one person) | Coaching with the conversation and the rule | Team lead, in-house or BPO | This week |
| Pattern across agents | Micro-module in the LMS plus a drill in the next huddle | L&D specialist | 2 weeks |
| Knowledge defect | Article, macro and AI snippet fixed together | Content specialist | This sprint |
| AI agent error | Knowledge or guardrail patch, regression conversation added to the test set | Content + AI agent owner (CX Tooling) | 24-48 h |
| Process or product gap | Ticket to Product or Policy with volumes | Me, with the Program and Policy Lead | Monthly review |
| Metric | Definition | Cut by |
|---|---|---|
| QA score | Weighted scorecard, auto-fails as 0 | Channel, vendor, agent, topic |
| Auto-fail rate | Share of conversations with any auto-fail | Auto-fail type, channel |
| Calibration agreement | Share of criteria where graders match exactly; AI vs lead agreement per criterion | Grader, criterion |
| Ramp time | Days from start to two consecutive weeks at QA bar | Cohort, vendor |
| Knowledge freshness | Articles past review date; open drift checks | Collection, owner |
| AI resolution quality | Automated resolution rate paired with QA score on those conversations | Topic |
| Deflection and CSAT impact | Contact rate per topic before and after a content change; CSAT on those topics | Article |
Built from the public job posting (2026), Polymarket's sister CX postings, its public help center and docs, and the public Gamma API, read on 2 October 2026. The demo is my own tool built for this application; its live conversations are real public replies on X, its training set is written for the demo and labelled, and its help-center text, docs, market data and prices are fetched live. Nothing here is Polymarket's production code or internal data.
Help center: help.polymarket.com · fees · disputes · geo
Docs: resolution · fees · pUSD
Upgrade: 28 Apr 2026 · Recovery: missing deposit
Market data: Gamma API (fields feeSchedule, umaBond, customLiveness)
Competitors: Kalshi help · Kalshi on Fin · Geo: docs geoblock
Independent CX homework for the Polymarket Content, QA, and Learning Development role · 2026 · edwardtay.com