Fraud & Quality Shield
If you pay respondents, every bot and every repeat submitter costs you money and dirties your data. The Shield decides, for each submission to a public survey, whether it is a real person's answer. It holds the doubtful ones for review before any payout goes out.
Built for research panels, market-research firms, and programs that give gift cards for feedback. It runs inside Salesforce, on the surveys you already run with AutoSurvey. Einstein reviews every answer through the Einstein Trust Layer, with the data masked and under Salesforce's zero-data-retention terms with its model providers.
What it does
The Shield scores each submission from 0 to 100 on signals you choose. The score lands in one of four decisions. You set the thresholds.
Submitted as normal.
The respondent sees the completion screen. Most real answers end here.
Submitted and marked for attention.
Counted in your analytics. The respondent sees the completion screen.
Submitted, but kept out of analytics and out of any payout.
A reviewer approves or rejects it in the queue. The respondent sees the completion screen and learns nothing.
Not submitted.
The run closes as invalid. The respondent sees a polite stop screen with no reason and no count.
Why it is useful
A held submission never reaches your payout Flow until a reviewer approves it. You set the incentive value per form, and the dashboard shows an estimated payouts-avoided figure. The label says estimated, because it is.
Held and blocked submissions stay out of your reports and dashboards. Duplicate emails, duplicate devices, copied answer patterns, and straight-line responses are caught before they skew a result.
Every held submission shows why it was held: each rule, its measured value, and its weight. A reviewer approves or rejects in bulk. No spreadsheet export, no manual de-duplication.
Einstein scores each submission for risk and for answer quality, and writes a one-paragraph rationale for the reviewer. It runs through the Einstein Trust Layer: Salesforce masks sensitive data, keeps an audit trail, and holds zero-data-retention agreements with the model providers it uses. Which model answers is set by your org's Einstein configuration, not by the Shield. It needs Einstein generative AI (Agentforce) enabled; without it, the rest of the Shield runs unchanged.
Pick reCAPTCHA v2, reCAPTCHA v3 or Cloudflare Turnstile. Set the challenge to never, always, or only when the policy raises risk. Most real respondents never see it.
Monitor mode scores everything and holds nothing. Run it for two weeks, replay a new policy against the last 30 days, and switch holds on only when the numbers look right to you.
The screens
The Shield lives in your AutoSurvey app as three tabs: Shield, Shield Review and Shield Settings. Purple reason codes come from Einstein.
01 · Shield dashboard
Screened, blocked, held, and an estimated payouts-avoided figure from the incentive value you set per form. A 30-day chart shows the decisions by day, so a spike of blocks is visible the morning after it happens. The top-reasons table shows how often Einstein raised a reason on its own.
02 · Review queue
A reviewer sees the risk score, the reason codes, and a hashed email prefix instead of an address. Bulk approve or reject feeds your own payout Flow. Einstein's one-line verdict sits next to the selection, so the reviewer does not open each record.
03 · Policy configuration
A policy is a set of numbers you can read. A monitor-only switch for the first two weeks. The Einstein review section shows whether Einstein generative AI is enabled in the org, when Einstein runs, and the two AI thresholds. A replay shows what the new numbers would have done to the last 30 days before you save them.
04 · Submission risk detail
Each rule's measured value and weight, Einstein's risk and quality scores with its rationale, the pace per screen against the form's median, the answers, and the same actor's history. In this example the text heuristic missed the problem and Einstein caught it. IP address, location and device fingerprint are not captured.
05 · The respondent's view
The challenge appears only when the policy raises it. A blocked attempt gets a neutral stop screen with no reason and no count, so a bot operator learns nothing from it.
The first version
Kept for a later version: device fingerprinting, behaviour analytics, IP reputation, email verification services, blocklists and allowlists, cross-flow risk profiles. AI review through any model that the Shield calls on its own, outside the Einstein Trust Layer, is not planned.
Privacy posture
Design partner wanted
The design and the proof of concept are complete, down to the data model and the effort by phase. We finish the build with one organization that pays respondents and runs its surveys on Salesforce. You run the Shield on your own surveys, monitor mode for the first two weeks, and you tell us what the queue and the policy get wrong.
The Shield is included in the AutoSurvey price. There is no add-on charge, and you keep the price you sign at.
AutoSurvey already gives every guest survey CAPTCHA and rate limiting today. The Shield adds the queue, the policy, the duplicate checks and the Einstein review on top. See Security.
Status. The Fraud & Quality Shield is under development. The screens on this page are design previews rendered from the completed design, not screenshots of shipping software. Every organization, person and number on them is fictional sample data, and no figure is a customer result. The Einstein review needs Einstein generative AI (Agentforce) enabled in your org and is subject to Salesforce's own terms for it. Scope may change before release.