← Glossary

Devancore Post-Trade Glossary

Conversational Finance Ops

Conversational finance for investment operations is a permissioned natural-language interface over operating records: the question, the entitled sources, the cited evidence, and a workflow action that still needs approval.

Definition

Conversational finance for investment operations is a permissioned natural-language interface over the operations book. The desk asks why a trade is still open, which cash breaks aged past a cutoff, which positions do not match custody, or which evidence sat behind a report. The interface may answer only from records the user is entitled to see, with citations, and may start a workflow only under the same permissions that apply outside the chat.

This page opens a cluster. Later articles define the parent concept, agents that propose multi-step work, natural-language order intent, books-and-records queries, and copilots. The lane here is the operations desk: positions, cash, breaks, settlement, controls, and reporting evidence. Order tickets, unsupervised agents, and retail advice are out of scope.

Four records every ops answer must carry

Four records every ops answer must carry

If any one is missing, the interface is a chat window, not an operations control.

Record What it holds Control question
Question User, time, intent parsed to positions, cash, breaks, settlement, or evidence Can the intent be mapped to a named operating object?
Entitlement Role, account scope, desk, and whether retrieve or act is allowed Would this user see the same rows in the source system?
Citation IBOR, cash, custody, break file, settlement status, report pack, as-of Can the answer be replayed to the source rows?
Action None, annotate, escalate, or draft a controlled workflow Did state change only after permission and, where required, a checker?
Trail Prompt, retrieved keys, proposal, approver, outcome Can a supervisor reconstruct what was asked, shown, and done?

The control problem is time-to-record, not chat novelty. Grids and overnight packs hide the current fail, the aged break, and the missing citation. A fluent model that cannot point at the investment book, the cash recon, or the break file creates a new operating risk: people act on an unsourced sentence.

Query and action must stay split. "Show aged settlement fails" is retrieval. "Mark these matched" is a state change. The second path needs entitlement, maker-checker where the firm requires it, and a trail of prompt, sources, proposal, and outcome. A single chat box that does both without a branch will post.

Entitlement is the same map the source system uses. If a user cannot open another desk's break queue in the recon tool, the conversational layer must not summarize that queue. Scope is a control, not a display preference.

How it works

The workflow is parse, authorize, retrieve, cite, then branch. Natural language is mapped to an operating object. The permission layer decides retrieve versus act. Sources are read at a named as-of. The answer cites keys. Only then may a draft action enter maker-checker.

From question to cited answer or draft action

From question to cited answer or draft action

Query and action are different paths. Mixing them is how a chat posts a break without a checker.

Step Data required Failure mode
Parse Intent: position, cash, aging, fail, control, report evidence A vague prompt is answered from a model prior, not from a book
Authorize User role, account scope, retrieve versus act The chat shows another desk's breaks or cash
Retrieve IBOR, recon, break, settlement, or pack at a named as-of The answer mixes yesterday's snapshot with today's fail file
Cite Source system, key, as-of, and materiality A fluent summary cannot be tied to a row
Branch Query-only versus draft workflow An 'approve matched breaks' line posts without maker-checker
Record Prompt, citations, proposal, decision, outcome Operations cannot show what the interface did

Source systems are the investment book of record, cash and custody recon, trade-break and fail files, settlement status, control exceptions, and reporting packs. The conversational layer does not replace those systems. It is an access and draft path over them. If the book is stale, the sentence is stale.

Retrieval must name grain and time. Positions as-of last close are not the same as intraday failed trades. Mixing them in one paragraph is how operations trusts a number that never existed on one run.

Human-in-the-loop sits on the action branch. Annotating a break, escalating a fail, or requesting a checker is allowed when policy says so. Generating a settlement instruction or rewriting a reportable figure is a later, stricter workflow — and still not autonomous execution. Segregation of duties does not relax because the request arrived as a sentence.

Evidence is the prompt plus the rows. Supervisors need to see what was asked, which keys were returned, what was proposed, who approved, and what posted. That trail is the operations audit object for the interface. Conversational finance as a cluster will reuse this pattern; this page only requires it for ops query and ops draft.

In Devancore™

Devancore supports conversational finance for investment operations as a post-trade operating-record layer around questions, entitlements, cited retrieves, draft actions, and trails. It can help teams ask current-state questions over positions, cash, breaks, and settlement without dropping permissioning or source evidence.

Devancore does not act as an autonomous agent, executing broker, administrator, or supervisor. It does not answer from outside the entitled books, post a break silently, or certify a report. Those functions remain with the source systems and the responsible parties. The product boundary is the controlled operating copy of question, citation, and draft that operations already depends on.

In a Devancore-style workflow, a question, retrieve, citation, draft, approval, or block enters as an event. The record is mapped to user, object, source keys, as-of, and workflow state. Query-only, drafted, checked, posted, or blocked remains visible with actor, timestamp, reason, and source reference.

That structure is what the desk uses. The book may hold the position. The recon may hold the break. The pack may hold the report line. The operating question is whether a natural-language answer still joins those records under the same entitlement as the screens.

Conversational use depends on the same chain. A user can ask which fails aged past cutoff, which cash breaks lack an owner, which positions miss custody, or which draft actions sat unapproved. The answer should resolve to keys, as-of, owners, and evidence.