← Glossary

Devancore Post-Trade Glossary

Human-in-the-Loop Execution

Human-in-the-loop trade execution is a controlled workflow where software can draft, enrich, validate, or prepare an order, but a human approval state is required before release.

Definition

Human-in-the-loop trade execution is the release-control model used when software prepares a trade instruction but a person must approve the action before it is routed. The software may parse intent, enrich fields, resolve instruments, run checks, stage an order, draft a FIX or API payload, or propose an exception path. The human gate decides whether the prepared action can leave the controlled workflow.

The key record is the decision state. A firm needs to know whether the instruction is a draft, pending review, approved, rejected, held for clarification, escalated, overridden, routed, partially filled, cancelled, or closed. Each state should have source evidence and a responsible actor.

Human release gate

Human release gate

The control is the point where a prepared draft becomes an authorized market or workflow action.

State What exists Control question
Source request Prompt, chat message, model output, rebalance file, blotter row, or portfolio instruction Can the reviewer see why the draft exists?
Prepared draft Instrument, account, side, quantity, order type, price, allocation, route, and settlement intent Are the generated fields complete and resolved?
Control result Entitlement, restrictions, risk, cash, margin, account eligibility, and destination validation Which controls passed, warned, or blocked the action?
Human decision Approve, reject, edit, hold, clarify, escalate, or override with reason What authority changed the draft state?
Release event OMS, EMS, FIX, brokerage API, allocation, settlement, or downstream workflow instruction What exactly left the controlled environment?
Evidence trail Source draft, visible context, reviewer, checker, timestamp, route, response, and outcome Can the release be reconstructed later?

A draft is not enough. A generated order may contain an instrument, account, side, quantity, price, and route, while still being unapproved, outside entitlement, blocked by a restriction, incomplete for allocation, or unsupported by settlement data. Human-in-the-loop design prevents the draft from becoming an uncontrolled state change.

The reviewer needs context, not only a submit button. The review screen should show the original request, parsed fields, resolved records, validation results, warnings, missing data, proposed route, downstream effect, and prior workflow history. If the reviewer changes the draft, the change should be stored as part of the approval record.

Approval is different from acknowledgement. Acknowledgement records that someone saw an alert. Approval authorizes a change: order release, route submission, allocation instruction, settlement action, exception closure, or override. Institutional workflows should record approval scope, user, role, time, reason, and affected records.

Supervision has to be designed into the release path. Broker-dealer and adviser workflows should be capable of supporting written supervisory procedures, including review queues, escalation paths, exception evidence, and supervisory visibility over AI-assisted order preparation. For broker-dealer environments, this is the practical operating layer around FINRA Rule 3110-style supervision.

The release event is the boundary. Once the instruction leaves the controlled workflow, downstream systems create new records: OMS order, EMS route, FIX message, brokerage API request, order ID, execution response, allocation, confirmation, settlement intent, and audit trail. The human approval should remain linked to those records.

How it works

Human-in-the-loop execution works by separating preparation from release. AI and workflow systems can assemble the draft. Controls test whether the draft is eligible. A human decision changes the state from prepared to approved, rejected, held, or escalated.

Execution control workflow

Execution control workflow

Human-in-the-loop design separates machine preparation from human release authority.

Step Record created Failure mode
Create draft AI or workflow output with source request, parsed fields, account context, and intended action The draft is treated as executable before required fields are complete
Resolve records Instrument master, account, portfolio, allocation, route, restriction, and settlement context The system maps to the wrong security, account, route, or strategy
Run controls Entitlement, pre-trade rules, restricted list, risk, buying power, margin, and destination schema The release path bypasses the same checks used by ordinary orders
Review context Visible source, generated fields, warnings, exceptions, differences, and proposed downstream action The reviewer sees a summary without enough evidence to approve
Record decision Approve, reject, hold, edit, clarify, escalate, or override with user, role, timestamp, and reason An unjustified control override is stored as a simple click
Release and reconcile Order ID, route, FIX or API payload, execution response, allocation, confirmation, and audit trail The released action cannot be linked back to the approved draft

Draft creation begins with a source request. The source may be a portfolio instruction, natural-language prompt, chat message, model output, rebalance file, blotter row, or exception workflow. The source record should stay attached because it explains why the draft exists.

Record resolution maps the draft to governed objects. Instrument master, account master, portfolio structure, strategy, allocation method, broker route, custody path, settlement instruction, and restriction data determine whether the draft points to real operating records.

Control validation runs before review. Entitlement checks whether the user may stage, approve, or release the order. Pre-trade rules test restrictions, suitability of account scope, risk limits, concentration, cash, margin, and market-access constraints. Destination validation tests whether the receiving system can accept the payload.

Review converts system output into accountable action. The human reviewer should see what was explicit, what was inferred, what failed, what changed, and what will happen if the instruction is released. Maker-checker applies when policy requires independent approval. If a warning, restriction, account eligibility issue, or risk limit is overridden, the reviewer should enter a reason that becomes part of the audit trail.

Release sends the approved instruction to the appropriate route: OMS, EMS, FIX session, brokerage API, allocation workflow, settlement instruction engine, or downstream operating queue. The response from that system should be captured immediately and linked back to the approved draft.

T+1 settlement pressure makes the interface design important. Human review should be fast because decisions often happen near allocation, confirmation, or settlement cutoffs. Speed should come from enriched context, grouped warnings, clear choices, and pre-filled evidence, not from skipping review or forcing manual re-entry.

Exception handling is part of the same model. A human reviewer may reject the draft, hold it for missing information, request clarification, escalate it to a supervisor, or override a warning with reason. Those decisions are records, not side notes.

In Devancore™

Devancore supports human-in-the-loop execution as a controlled operating-record workflow around generated instructions, trade capture, route release, exception handling, and post-trade evidence. The platform should be framed as the layer that preserves source intent, validation results, approval state, routed payloads, downstream responses, and audit trail.

Devancore should not be framed as an execution venue, broker, investment adviser, custodian, clearing broker, compliance officer, or autonomous trading system. Its role is to help the firm govern the transition from prepared instruction to approved action.

In a Devancore-style workflow, AI-generated output remains a draft until records resolve and controls pass. A user can review the source, fields, warnings, route, and downstream effect. A checker can approve where required. If the instruction is released, the order ID, FIX or API payload, execution response, allocation, confirmation, settlement state, and audit trail remain tied to the approval.

This page sits between the broader conversational finance pages and the narrower execution API pages. Conversational finance explains the interface. Natural language trading explains intent capture. AI-generated trade instructions explain the draft. Human-in-the-loop execution explains the release authority that makes the workflow governable.

The same control model applies when the action is traditional, API-based, or digital-asset related. The firm should be able to see who approved the action, what evidence was visible, what system received the payload, what happened next, and which records were affected.