Platform
Automation & Intelligence

AI Compliance Assistant

An assistant that helps analysts work faster on the record

An assistant that helps compliance and risk teams summarise records, surface relevant context and draft narratives from platform data — always keeping a human in control of decisions and writing to the same auditable record. Much of an analyst's day goes into reading long records, gathering scattered context and writing up what they found. The assistant helps with exactly that: it drafts, summarises, retrieves and answers questions grounded in the firm's own data, so a person starts from a first draft rather than a blank page. It does not decide, close, approve or file anything. Every output is a suggestion a qualified person reviews, edits and owns, and the human judgement — and accountability — stays exactly where it was.

At a glance

How it works, visually

Where this sits in the platformSchematic
Data sources & inputsCustomers · vendors · transactions · documents · your systems of recordIngestion & screeningOnboarding · sanctions, PEP & adverse-media screening · capture & normalisationRisk, case & monitoring enginesRisk scoringCase & investigationOngoing monitoringControls, evidence & policyControls library · testing · evidence capture · policy mappingReporting & analyticsDashboards · regulatory returns · executive & board reportingIntegrations & audit trailAPIs & connectors · append-only, timestamped audit trailOne layered platform — every layer feeds the next, and every action lands in the audit trail.

The layered platform architecture — this module operates across ingestion, the engines and the audit trail.

The challenge

The problems this module solves

The operational realities that make this hard for compliance and risk teams today.

Analysts spend more time assembling than deciding

Before any judgement can be made, someone has to read the case, pull the customer's history, find the relevant policy and gather the adverse media. The decision itself is quick; the assembly around it takes the day, and that is where the hours go.

Writing up findings from a blank page is slow

Narratives, case rationales, review summaries and report sections all have to be written from scratch each time. Skilled people spend hours drafting prose that a first cut could have started, then refined — time that could have gone to the analysis itself.

Relevant context is scattered and easy to miss

What an analyst needs to make a call is spread across cases, records, policies and prior decisions. Finding all of it manually is slow, and a relevant piece — a prior alert, an applicable rule — is easy to overlook simply because no one thought to look there.

Answering 'what does our policy say' takes too long

A frontline question about an obligation or an internal interpretation sends someone searching through documents for the passage that answers it. The answer exists, but retrieving and citing it reliably is slow enough that people guess instead.

Fear of ungoverned AI in a regulated function

Firms are rightly wary of AI that acts on its own, can't show its sources, or leaves no record. An assistant that made decisions, hallucinated facts or worked off the record would import risk into exactly the function meant to control it.

The approach

How OnyxOne addresses it

Drafts the first cut, a person owns the final

The assistant drafts narratives, summaries and report sections from the firm's own data, so an analyst starts from a first version rather than a blank page. What it produces is always a draft: a qualified person reviews, edits and owns the result, and the decision remains theirs.

Gathers and summarises the context

It pulls together the relevant record — case history, related entities, applicable policy, adverse-media summaries — so the material an analyst needs is assembled in one place. The person still reads and verifies it; the assistant saves the hunting, not the judging.

Answers policy questions with citations

Asked what an obligation or internal interpretation says, the assistant answers from the firm's own knowledge base and cites the source passage, so the answer can be checked rather than trusted. Where it isn't confident, it says so rather than inventing an answer.

Surfaces anomalies for a person to investigate

It points out records or patterns that look unusual against their context, so an analyst's attention is drawn to what may warrant a look. It flags; it does not conclude. Whether something is actually a problem is always for a person to determine.

Grounded, cited and on the record

Every output is grounded in the firm's own data, carries its sources where it can, and is written to the same auditable record as any other work. The assistant makes no autonomous decisions, and its suggestions are attributable, reviewable and owned by the person who acts on them.

Capabilities

What's in the module

Turn on what you need and add more as your programme scales.

Narrative drafting

Drafts case narratives, rationales and report sections from platform data for a person to review, edit and own.

Record summarisation

Summarises long cases, entity histories and documents so an analyst can grasp the material faster, then verify it.

Adverse-media summarisation

Condenses adverse-media and open-source material into a summary a person confirms against the underlying sources.

Contextual retrieval

Gathers related cases, entities, prior decisions and applicable policy into one place so relevant context isn't missed.

Policy Q&A with citations

Answers questions about obligations and internal interpretations from the knowledge base, citing the source passage.

Anomaly surfacing

Highlights records or patterns that look unusual against their context for a person to investigate — it never concludes.

Drafting assistance

Suggests wording and structure for narratives and reports, which the author is free to accept, edit or discard.

Grounded outputs

Bases every response on the firm's own data and declines to answer, rather than inventing, where it lacks a basis.

Human-in-the-loop by design

Produces suggestions only; no output takes effect until a qualified person reviews and accepts it.

On-the-record and attributable

Writes assistant outputs and the human's edits to the same append-only audit trail as the rest of the work.

Dashboards

The views your team works from

Purpose-built dashboards and views, each answering a question a specific role needs to act on.

An executive viewIllustrative
ILLUSTRATIVE EXAMPLEOPEN CASES128SLA ADHERENCE96%SCREENING ALERTS1.2kOVERDUE REVIEWS14Cases by categoryAMLKYCFraudSanctionsConductOtherRisk mixby tierHighMediumLow

A representative layout of the KPI tiles and charts these dashboards present. Figures shown are illustrative examples, not real data.

Assistant activity

Where and how the assistant is being used across teams and modules, so its role in the workflow is visible.

Suggestion review

Assistant outputs alongside the human's final version, showing what was accepted, edited or discarded.

Citation coverage

How well grounded and cited the assistant's answers are, so the function can see where verification matters most.

Anomaly queue

Records the assistant has surfaced as unusual, presented for a person to investigate and disposition — never auto-actioned.

Oversight view

The audit-ready record of assistant suggestions and the human decisions taken on them, for model and compliance oversight.

Automation

What the platform automates

Rules, workflows, alerts and scheduling that run the routine so your team works the exceptions.

Draft-on-request

Generates a draft narrative, summary or report section when an analyst asks, saving the blank-page start — nothing is filed automatically.

Context assembly

Gathers the relevant records, entities and policy for a case into one place when invoked, so the analyst isn't hunting for context.

Adverse-media digestion

Condenses connected adverse-media and open-source material into a summary for the analyst to confirm against the sources.

Anomaly flagging

Surfaces records that look unusual against their context for human investigation, drawing attention without drawing conclusions.

On-record logging

Writes every assistant output and the human's edited final to the audit trail automatically, so the interaction is always recorded.

AI assistance

Where AI helps the analyst

Assistive, decision-support features that speed up the work on the record. Suggestions are always reviewable, and a person stays in control of every decision.

Grounded narrative drafting

Drafts case and report narratives from the firm's own data as a starting point, which the author reviews, edits and owns — the assistant never files or decides.

Cited policy answers

Answers questions about obligations and interpretations from the knowledge base and cites the source passage, and says when it isn't confident rather than inventing an answer.

Anomaly and context surfacing

Summarises long records and points out what looks unusual for a person to investigate, informing human judgement without making any decision of its own.

The workflow

The enterprise workflow

A defined, end-to-end process with clear ownership at every stage.

The workflow, step by stepSchematic
01Ask or invokeAn analyst asks a question or invokes the assistant on a case, record or documentfrom within their existing workflow.02Gather from the recordThe assistant retrieves the relevant context — history, related entities, applicablepolicy, source material — from the firm's own data.03Draft or answerIt produces a draft narrative, a summary or a cited answer, grounded in that dataand marked clearly as a suggestion.04Review & verifyThe analyst reads the output, checks it against the cited sources and the underlyingrecord, and edits as needed.05Decide & actThe person makes the decision and takes the action — the assistant neither decides,closes, approves nor files anything itself.06RecordThe assistant's suggestion and the human's final, edited version are written to theauditable record, attributable to the person who owns them.

Every result, decision and override is captured against the record it belongs to.

01

Ask or invoke

An analyst asks a question or invokes the assistant on a case, record or document from within their existing workflow.

02

Gather from the record

The assistant retrieves the relevant context — history, related entities, applicable policy, source material — from the firm's own data.

03

Draft or answer

It produces a draft narrative, a summary or a cited answer, grounded in that data and marked clearly as a suggestion.

04

Review & verify

The analyst reads the output, checks it against the cited sources and the underlying record, and edits as needed.

05

Decide & act

The person makes the decision and takes the action — the assistant neither decides, closes, approves nor files anything itself.

06

Record

The assistant's suggestion and the human's final, edited version are written to the auditable record, attributable to the person who owns them.

The value

What your team gains

Assistive

A first draft, not a decision

Drafting narratives and summaries from the record lets analysts start from a first version and spend their time on judgement, while the decision stays entirely theirs.

Grounded

Answers you can check

Responses are grounded in the firm's own data and cited to their sources, so an analyst can verify an answer rather than take it on trust — and the assistant says when it doesn't know.

In control

Human always in the loop

No output takes effect until a qualified person reviews and accepts it; the assistant makes no autonomous decisions, so accountability stays exactly where it was.

On the record

Suggestions that are auditable

Assistant outputs and the human's edits are written to the same append-only trail as any other work, so what the assistant proposed and what the person decided are both on the record.

Less hunting, more analysis

By gathering scattered context into one place, the assistant removes the assembly that eats an analyst's day without touching the judgement that only a person can make.

Governed AI for a regulated function

Grounded, cited, human-in-the-loop and on the record by design — the assistant is built to help inside the controls a regulated function needs, not to work around them.

Built for

Industries it serves

BankingFinancial ServicesFintechInsuranceInvestment FirmsAsset ManagementLendingGamingRegulated Enterprises
Integrations

Works with your existing systems

Described as capabilities — OnyxOne connects to the systems your deployment requires, configured per implementation.

Platform modules
  • Works on the cases, screening results, due-diligence records and documents already held on the platform, as a helper within them
Regulatory knowledge
  • Answers policy and obligation questions from the regulatory knowledge base and cites the source passage
Case management
  • Drafts case narratives and summaries for the investigator to review, edit and own within the case record
Reporting & analytics
  • Drafts report narratives from live programme data, which a person reviews before the report is finalised
Audit trail
  • Writes every suggestion and the human's edited final to the same append-only audit record as other work
Assurance

Security, compliance & reporting

Security & data handling

  • The assistant operates only on the firm's own data, within the same role-based permissions as the user — it can surface nothing a person could not already see.
  • Every output is a suggestion; nothing the assistant produces takes effect, closes, approves or files anything without a qualified person's review and action.
  • Assistant outputs and the human's edits are attributed and written to an append-only audit trail, so what was suggested and what was decided are both recorded.
  • Outputs are grounded in the firm's data and cite their sources where possible; the assistant is designed to decline rather than fabricate where it lacks a basis.
  • The assistant makes no autonomous regulatory decisions and holds no authority — accountability remains entirely with the human user.
  • Data handling, residency and retention for assistant interactions are configurable to your regulatory obligations.

Compliance support

  • Positioned as decision-support only — the firm's staff make and own every decision
  • Supports explainability by grounding outputs in the firm's data and citing sources
  • Keeps a human in the loop consistent with supervisory expectations for AI in regulated functions
  • Records assistant suggestions and human edits for audit and model-oversight review
  • OnyxOne is a technology vendor; the assistant informs the firm's judgement and never replaces it

Reports & exports

  • Assistant-usage reports by team, module and task type
  • Suggestion-versus-final reports showing where human edits changed the draft
  • Citation and source-coverage reports for grounded answers
  • Time-in-workflow reports showing where the assistant is used most
  • Audit-trail reports of assistant outputs and the human decisions taken on them
  • Anomaly-surfacing activity reports for review and tuning
Best practice

How to get the most from it

Treat every output as a draft

Read, verify and edit what the assistant produces before it goes anywhere. Its value is a faster first cut, not a decision — the moment a suggestion is accepted unread, the control is gone.

Check the citation, not just the answer

When the assistant cites a source, confirm the passage says what the answer claims. Grounded and cited makes verification possible; it is the person's job to actually do it.

Keep the human on every decision

Use the assistant to gather, draft and summarise — never to close, approve or file. Autonomy is exactly what a regulated function should not delegate to it.

Review where it helps and where it doesn't

Watch how much human editing outputs need. Where the assistant reliably saves time, lean in; where it is often rewritten, treat its output with more caution and feed that back.

FAQ

Questions, answered

Does the assistant make compliance decisions?

No. It drafts, summarises, retrieves and answers questions, but it never decides, closes, approves or files anything. Every output is a suggestion that a qualified person reviews, edits and owns, and the decision — and the accountability for it — remains entirely with the human.

How do you stop it inventing facts?

Outputs are grounded in the firm's own data and cite their sources where possible, and the assistant is designed to say it doesn't know rather than fabricate where it lacks a basis. Even so, the person is expected to verify every output against the cited sources and the underlying record before relying on it.

What can the assistant actually see?

Only the firm's own data, and only within the same role-based permissions as the user invoking it. It cannot surface anything a person could not already access, and it does not reach outside the platform's governed data to answer.

Is there a record of what the assistant did?

Yes. The assistant's suggestion and the human's final, edited version are both written to the same append-only audit trail as any other work, attributed to the person who acted on them — so what the assistant proposed and what the person decided are both on the record for oversight and model review.

Isn't AI risky in a regulated function?

Ungoverned AI would be. This assistant is deliberately assistive only: grounded, cited, permission-bound, human-in-the-loop and on the record, with no autonomous decisions and no authority of its own. It is built to help inside the controls a regulated function needs, informing the firm's judgement rather than replacing it.

See AI Compliance Assistant in your programme

Book a walkthrough and we'll show how this module fits your policy, workflows and obligations — then scope an implementation.