Can AI market agents tell missing data from low risk?

The financial industry's debate over artificial intelligence is moving from what models can write to what systems may be allowed to do. On 31 August, the Financial Stability Board warned that frontier models were showing more sophisticated autonomy and called for safer deployment, stronger recovery planning and resilience among critical technology providers.

That warning was directed chiefly at cyber and operational risk, not at a science-fiction trading desk. Yet it captures the practical issue facing any financial firm that connects a general-purpose model to tools: an articulate answer is not the same thing as an accountable one.

FINRA has told member firms that existing supervisory obligations continue to apply when generative AI is used, whether the technology is built internally or supplied by a third party. European securities regulators have made a similar point: management remains responsible for decisions made with AI-based tools.

Fahali, a market-intelligence startup, is trying to turn that governance problem into a product. It describes the system as a market-risk conscience for humans and supervised AI agents. The metaphor is deliberate. Fahali is not offering an agent permission to trade; it is building a layer intended to make evidence, uncertainty and authority harder to blur.

The danger in a quiet field

Many consequential software errors are mundane. A data field is absent. A source times out. An integration maps null to 0. A downstream model then treats “not measured” as “measured and calm”.

No interface can force another model to reason correctly. Fahali's narrower claim is that the contract can stop doing the first part of the damage. It keeps measured-empty, event-gated quiet, stale, no-coverage, partial, withheld and error states distinct. Callers still have to preserve those states.

Missing is reported as missing. Event-gated quiet remains quiet. Neither becomes zero, safe or a win.

This sounds less exciting than predicting the next crash. It is also closer to the problem regulators describe: reliability, data integrity, model risk, oversight and clear accountability.

A protocol, then a product contract

Fahali reaches AI applications through the Model Context Protocol, or MCP. The protocol is an open standard for connecting language-model applications to external tools and data. MCP provides the transport and capability structure; Fahali supplies the market-specific response contract.

As of 1 September, Fahali's MCP 1.21.0 advertised 49 canonical tools. The count is less important than the proof boundary. Production qualification exercised 38 independently callable reads and a separate 11-call watch, delta and subscription lifecycle. Because two calls overlapped, those tranches covered 47 distinct advertised tools. The memory-write and agent-authored alert utilities were outside that live production proof.

The current autonomous workflow is long-only and MONITOR_ONLY. It can maintain a watch, decide whether eligible evidence supports a nonbinding monitoring judgment, preserve history and deliver an event. It cannot provide personalized advice, select a security, rebalance, generate or route an order, access a broker, hold custody or move capital.

Beyond a price feed

The tool catalogue covers familiar market-risk jobs, but several names require restraint. Fahali's public contagion surface maps lower-tail co-movement—assets that have tended to be weak together. It does not establish that one asset caused another to fall, or that one reliably leads the other.

The shock tool is similarly narrower than a forecast. It applies a user-supplied signed move to supplied positions and reports deterministic impact. A short position gains under a direct negative shock; an equal long and short in the same instrument offset in the linear case. The result is scenario arithmetic, not a probability, investment recommendation or autonomous portfolio action.

That distinction is commercially awkward but analytically important. Financial software often bundles a measurement, an inference and a proposed action into one field. Fahali's newer contract separates the observation from the authority to publish a canonical direction—and separates both from any financial action, which the system does not have.

Cryptography is not compliance

Fahali also uses cryptographic controls, though not in the sweeping way an earlier draft of this article suggested. Persisted watch verdicts bind their decision, evidence and content with SHA-256 hashes. Webhook and agent-callback deliveries are separately HMAC-signed.

Those mechanisms make alteration detectable and support replay and receiver verification. They are not a universal signed-verdict attestation, do not make records tamper-proof and do not establish regulatory compliance.

The product contract also encodes identity, freshness, source completeness, expiry, tenant and no-execution controls. These can support a firm's supervision. They cannot replace firm-specific governance, legal analysis, human review, complaint handling or the duties that apply in a particular jurisdiction.

A record that keeps the misses

Fahali's strongest trust feature may be its judged record. Eligible predictive claims are registered with their identity and horizon before the outcome exists, then resolved after that horizon closes. Misses remain in the record.

The scorecard does not support one platform-wide accuracy number. Results have to be read by exact stratum, with the instrument class, horizon, base rate, sample, precision, recall and lead-time definition attached. When a result fails its evidence gate, the headline is withheld rather than replaced with a better-looking aggregate.

This is an internal measurement system, not an independent audit. Instrument-identity contamination has previously forced Fahali to quarantine outcome records and recompute affected aggregates. That history is a reason to inspect the methodology, not a reason to assume the latest score is infallible.

The commercial test is still ahead

Fahali has production evidence for a bounded monitoring lifecycle, but not yet evidence of sustained multi-customer operation. The company's next programme, Gate 6, remains a written control specification. Its recurring controls are not built and the clean-day observation clock has not begun.

The earliest controlled customer pilot follows 90 consecutive clean days after those controls are deployed, consumed and read back active, plus the required legal, security and model-risk approvals. That timetable is an internal release policy, not a regulatory waiting period or safe harbour. It does not put advice or capital authority on the roadmap.

The unanswered business question is whether teams will pay for evidence discipline that is intentionally narrower than an execution platform. Fahali's case is that a persistent monitor can reduce synthesis work and make the handoff to a human more defensible. It makes no claim that the product prevents a bad trade or protects a portfolio from loss.

The bet

Bloomberg and other terminals provide broad market data, news and workflow infrastructure. Fahali is positioning a narrower monitor-only evidence layer that can sit alongside existing tools. This is not a feature-equivalence or performance claim.

The wager is simpler: as software becomes more agentic, evidence boundaries will matter as much as additional data. Missing must stay missing. Expired authority must stay withheld. A noncausal association must not become a transmission story. And the human who owns the financial decision must remain visible in the system design.

Whether Fahali can preserve that discipline under real customer load is the next test. For now, its ambition is not to become the agent's hands. It is to make the agent show its work.