How Amanah works
Amanah triages incoming crowdfunding campaigns for fraud signals, policy violations, and zakat-eligibility claims. AI reads every submission and lays out what it found — but it never decides. A human reviewer approves, rejects, or escalates every campaign. This page explains how that works and where the line between the two is drawn.
From submission to decision
Public
Campaign submitted
Title, story, goal, category, organizer profile, zakat claim. Anyone can submit; submissions are rate-limited.
AI · first pass
Screening
A fast model scores fraud risk against a seven-category rubric, reports how confident it is, and writes a one-line summary — in a few seconds, for every campaign.
AI · closer look
Deep review
Risky or uncertain campaigns get a second pass that weighs the organizer's history. Every concern it raises must quote the exact campaign text that triggered it — evidence, not vibes.
Human · the decision
A reviewer decides
Reviewers work a triage queue sorted by risk and approve, reject, or escalate. Each decision is written to an immutable audit log with the reviewer's name and a snapshot of what the AI was recommending at that moment.
Who owns what
The AI owns
- Reading every submission, instantly
- Scoring risk against a fixed rubric
- Quoting evidence for every flag
- Saying “I'm not sure” out loud
- Ordering the queue so people look at the right things first
Humans own
- Every approve, reject, and escalate
- Judgment on genuinely ambiguous cases
- Zakat-eligibility calls that need documentation
- Accountability — names in the audit log
- The rubric itself
“Insufficient information to distinguish an urgent legitimate need from fabricated urgency… this is exactly the kind of case a human should decide.”
— an actual deep-review output from the demo queue. Amanah is built to admit what it can't verify. Calibrated uncertainty is a feature, not a failure: when the AI isn't sure, it says so and hands the case up, rather than guessing.
The line is drawn in code, not in the UI
- The AI cannot change a campaign's status. Approvals, rejections, and escalations happen only behind reviewer authentication. The AI's only output is a recommendation and its evidence — there is no code path for it to decide.
- Submitters and reviewers see different worlds. The public status page never shows a risk score, a flag, or the AI's reasoning — because teaching submitters what the model looks for would teach bad actors how to evade it. Assessment detail is reviewer-only.
- Every decision is attributable and permanent. The audit log records who decided, what they decided, and what the AI was recommending at that moment. Those rows are only ever added, never edited or removed.
We measure where it fails
The hard case is subtle fraud — a calm, plausible story with one quiet inconsistency. A fast model can read past it. That is exactly why no campaign is ever approved without a human, and why Amanah runs against a labeled test set and shows reviewers its own error rates — including the fraud it would have missed — after every run. The dataset, the metrics, and the failure modes we still wrestle with are documented openly.
Built in the open
Want the engineering perspective — the two-stage pipeline, the cost model, the schema, the boundaries enforced in code? It all lives in the repository, not bolted onto this page. The docs cover the architecture, cost model, and evaluation in depth.