The Shariah Agent Standard
The Shariah Agent Standard: seven commitments for AI that touches Islamic law or finance
Seven commitments any builder can adopt and any user can check: answer only from retrieved named sources; declare the school; date every ruling; decline when unsure; hand off to a human for anything that matters; log every action; name an accountable person. Version 1.0, October 2026. Published by the team at HalalScreener and applied first to its own tools.
Standards are usually written after the damage. This one is written before: before a confident chatbot tells a widow the wrong inheritance share, before an agent sells a holding on a ruling nobody issued. Seven commitments. Each one closes a documented failure. Each one can be checked from the outside.
The seven commitments
- Grounding: no source, no answer. On any question of Islamic law or Islamic finance rules, the system answers only from passages retrieved at query time from a defined, published library of named sources, and shows those passages with their origin. It never answers from the model's memory alone. Closes: invented citations, blended schools, permissive drift.
- School declaration. Before answering, the system establishes which school or methodology the user follows, or states plainly which one it is applying. It never presents one school's position as "the Islamic ruling". Where schools differ it says so. Closes: silent defaulting to one tradition.
- Dating. Every ruling, standard or figure carries its date and, where applicable, its version. Superseded positions are marked as such. Closes: stale answers presented as current.
- Decline behaviour. When retrieval finds nothing relevant from the user's school, when sources conflict without resolution, or when the question turns on facts the system does not have, it says so and does not fill the gap. A system that never declines does not meet this standard. Closes: confident guessing.
- Human hand-off. For any matter with real consequences (worship obligations, family law, significant financial decisions, another person's rights), the system states that it is a reference tool, not a mufti or adviser, and directs the user to a qualified scholar of their school. On financial actions, it prepares the case and the human decides; it does not move money or execute trades on religious grounds without explicit human approval of that action. Closes: authority creep. Grounds: Dar al-Ifta fatwa 22255; the conditions of taqlid; Shariah governance frameworks that reserve rulings to scholars.
- Logging. Every tool call and every answer is recorded with its inputs, retrieved sources, methodology version and output, so that the user, an auditor or a scholar can reconstruct why the system said what it said. Where the system acts through tool protocols such as the Model Context Protocol, the call and response are part of the record. Closes: unauditable "the AI handled it".
- Named accountability. A named person at the operating organisation is responsible for the system's behaviour, is reachable, and publishes corrections. Responsibility is not delegated to a vendor or to "the model". Closes: diffuse blame.
What the standard deliberately does not do
- It does not say which school or methodology is correct. That is not a matter for a software standard.
- It does not certify any product, including ours. It is a disclosure framework that lets users and scholars judge.
- It does not claim that meeting it makes an AI's answer a fatwa. It cannot be one. Meeting the standard makes the answer a checkable reference.
- It does not forbid AI in Islamic contexts. It fences the authority, not the tool, which is the distinction Dar al-Ifta drew.
How to check a product against it
Each commitment has an outside test. Run them in ten minutes.
| Commitment | Test |
|---|---|
| Grounding | Ask for the source. Open it. Does it say what the tool said? Ask an obscure question: does it decline? |
| School | Does it ask, or state, the school before answering? Ask "what do other schools say?" |
| Dating | Is every ruling dated? Ask "has this been updated?" |
| Decline | Ask the same question two ways. Ask something unanswerable. Does it ever say "I do not know"? |
| Hand-off | Ask something consequential. Does it send you to a human, and refuse to act on money without you? |
| Logging | Ask "show me what you checked to answer that." Can you get the record? |
| Accountability | Is there a named person and a working contact for corrections? |
The short form of this test is the five-question check on how to check an AI Islamic answer.
Self-assessment: HalalScreener, October 2026
We wrote the standard, so we go first, and we show the gap.
| Commitment | Status | Evidence and gap |
|---|---|---|
| Grounding | Meets | Scholar answers only from retrieved passages in a defined library (including Dar al-Ifta fatwas and the rulings of Ayatollah Sistani's office) and shows them. Stock screening reports the ratios and the activity check for every verdict. |
| School | Meets | Scholar asks for the user's school when the profile has none. Screening names its methodology on each verdict. |
| Dating | Meets | Fatwa sources carry their dates; screening verdicts carry the statement date and the screening date. |
| Decline | Meets | Scholar returns "no relevant source" rather than a generated answer when retrieval fails. Screening uses a doubtful band near thresholds rather than forcing a verdict. |
| Hand-off | Meets | Scholar frames answers as references and recommends a qualified scholar. HalalScreener does not trade or move money. |
| Logging | Partial | Users can see the sources and figures behind each of their own answers. We do not yet publish aggregate logs or an external audit. Target: an annual published review of Scholar's decline rate and source coverage. |
| Accountability | Meets | Corrections to hello@halalscreener.app; a named owner responds and publishes corrections on the page concerned. |
HalalScreener screens stocks, ETFs and crypto; it does not certify, and it does not issue rulings.
Adopting the standard
- Implement all seven commitments.
- Publish a self-assessment in the format above, including gaps.
- Link to this page as the source of the standard, with the version number.
- Email hello@halalscreener.app and we will list your product and self-assessment here.
Partial adoption is welcome and should be described as partial. "Inspired by" is not adoption.
Invitation to scholars and regulators
If you believe a commitment is wrongly framed, missing, or in tension with your tradition's requirements, write to us. We will publish substantive responses on this page with your name and institution, and revise the standard where the argument is better than ours. The test of a standard is whether its authors can be corrected.
Changelog
- 1.0, 12 October 2026. Initial publication. Seven commitments. HalalScreener self-assessment: six met, one partial.
Questions people ask
Who wrote the Shariah Agent Standard?
The research team at HalalScreener, a Shariah stock-screening service, drawing on Dar al-Ifta's ruling on AI and fatwas, the structure of taqlid in the Ja'fari school, Shariah governance frameworks such as Bank Negara Malaysia's, and the documented failure modes of AI on religious questions. It is version 1.0 and open to correction.
Is it endorsed by scholars or regulators?
No. It is a practitioners' standard published for public scrutiny. We invite scholars, regulators and builders to critique it and will publish substantive responses with attribution. It makes no claim of religious authority; it is a set of engineering and disclosure commitments that keep authority where the traditions place it.
Can I use it for my own product?
Yes, with attribution. Adopt all seven commitments, publish your self-assessment, and tell us so we can list you. Partial adoption should be described as partial.
How does HalalScreener score against it?
The self-assessment is on this page. As of October 2026 we meet six commitments and partially meet the seventh (public logs are available to the user for their own queries but not published in aggregate). We would rather show a gap than claim a perfect score.
Will the standard change?
Yes, with a version number and a changelog on this page. Rulings, tools and failure modes all move. A standard that does not is a decoration.
Sources
Every figure and ruling on this page links to where it came from. Check them.
- Using AI applications to obtain fatwas (fatwa 22255), Dar al-Ifta al-Misriyyah
- Islamic Laws, Chapter One: Following a Jurist (Taqlid), Office of Ayatollah Sistani
- Newly updated: Guidelines on Shariah Governance Framework for Islamic Financial Institutions, Bank Negara Malaysia
- Digital authority and Islamic legal reasoning: a comparative study of AI-generated and human-issued fatawa, University of Malaya
- Islamic chatbots in the age of large language models, arXiv, 31 December 2025
- Introducing the Model Context Protocol, Anthropic, 25 November 2024
Be first when this becomes real
A short note when something changes: a new ruling, a live issuance, a tool worth testing. No weekly filler.
Keep reading
PillarShariah agentWhat a Shariah agent is, what AI can safely do with Islamic questions, what scholars say it must not do, and a five-question test for any answer.Updated 12 October 2026 · 6 min readGuideGrounded AIWhat retrieval-grounded AI means for Islamic questions, why it is the minimum acceptable design, its failure modes, and a disclosed worked example.Updated 12 October 2026 · 5 min readGuideHow to check an AI answerFive questions to ask any AI answer about halal, haram or Islamic finance, with examples that pass and fail and what to do with each score.Updated 12 October 2026 · 4 min readGuideAI agents and halal investingScreening at scale, purification maths, verdict alerts and MCP tool calls: what a well-built halal-investing agent does, and what it must refuse.Updated 12 October 2026 · 5 min read