# CDQ Genius Capabilities ## Overview **CDQ Genius** is a digital data steward for business-partner data. It orchestrates CDQ's products, reference data, data-quality rules and trusted sources through natural language, so a user can find, verify, explain and act on a business partner, with the evidence behind every answer. Genius is not a generic chatbot. It is grounded in CDQ's business-partner data expertise, product capabilities, reference data, source provenance and customer-specific insights. It connects conversational interaction with governed execution: it explains what it is doing, respects entitlements and policies, and involves a human wherever a decision calls for one. **This article covers:** - What Genius is, and how it decides which CDQ capability a question needs - The capabilities Genius orchestrates behind the conversation, and when each one is reached - How evidence, provenance and confidence grading make an answer defensible Available in: * CDQ Genius Chatbot — the conversational surface inside CDQ Cloud Apps Before you start Using Genius requires the **Genius User** role — without it, the Genius chatbot and app are not shown. Each organization also has its own daily Genius credit allowance, which resets every day; CDQ sets and can adjust this limit per organization. If a question can't be answered because the daily limit has been reached, reach out to your CDQ admin. ## What is CDQ Genius? Genius is an orchestration layer over CDQ's existing services, not a replacement for them. A request in plain language is classified into an intent, routed to the agent that owns that intent, and answered from a named source. The same lookup, curation, validation and screening services that power CDQ's products do the work. Genius decides which of them a question needs, in what order, and how to present the result so it can be reused and audited. | | ![How Genius resolves a question: intent classification, routing to the owning agent, a CDQ trusted source checked first with web research as fallback, and an answer that always carries its source and confidence](/assets/genius-img001.0f1214763a78429523243686727d28d0ece4f547038e94f9554c6a7ef32833a4.60a3ad26.png) | | It behaves according to three properties: - **CDQ-first, source-grounded.** Trusted CDQ sources are consulted before anything else, and every returned fact carries the source it came from. The open web is reached only when CDQ's sources cannot settle the question. - **Explainable.** Genius reports which agent it routed to, which source it checked, and why one candidate was preferred over another. Where a fact cannot be cited, it is marked unconfirmed. - **Human where it matters.** Today Genius suggests, explains and hands the decision to the steward on anything consequential: duplicates, compliance findings, borderline identity matches. As confidence in the agents and their evidence grows, more routine decisions will be automated; the evidence trail behind each one is what makes that possible. Genius is scoped by entitlement. A capability the customer has no contract for is unavailable and never silently substituted with a different source. ## The two layers of today's experience The current experience has a conversational layer (the Genius Chatbot) and, behind it, the capabilities that answer, verify and explain. ### The Genius Chatbot The chatbot makes each answer reusable and the work visible: - **Copy option.** Any answer, whether a resolved entity, a set of identifiers or an explanation, can be copied in one click for pasting into an MDG/ERP screen, a ticket or an email, without re-typing or losing formatting. - **Status and progress information.** While a request runs, Genius shows what it is doing step by step: which agent it routed to, which source it is checking, whether it is falling back to web research. The user sees the trace, not a silent spinner. - **Feedback collection.** A response can be rated or flagged as incorrect, incomplete or unclear. Feedback is recorded as a structured signal against that interaction and feeds prompt and agent refinement. | | ![The Genius Chatbot answer view, showing the routing/status trace, the source list behind an answer, and the Copy and feedback icons](/assets/genius-img002.f2300efd4cba1d274e5640c26a9e2aecb134d4e43129095e9a0851afc084d3f5.60a3ad26.png) | | ### The onboarding showcase skill A showcase skill demonstrates what end-to-end onboarding looks like inside the chat. Activated with `onboard` plus company details, it chains the integrated capabilities into a single guided flow: 1. Duplicate check against the customer's Data Mirror 2. Identity resolution in reference data sources 3. Curation of the business partner 4. Validation against data-quality rules 5. Sanction-list screening The flow returns a ready-to-use business partner with the evidence behind each step. ## Identity: lookup and duplicate detection Identity comes first. Before a record is curated, validated or screened, Genius establishes which legal entity is being talked about and whether the customer already has it. ### Lookup and fetch Genius resolves a business partner from a natural-language description (by name, address or identifier) across CDQ's connected sources, returns the most likely intended match first, and fetches the full record details. It explains why a candidate was chosen. When CDQ has no confident match, the request is extended to web research instead of returning an empty result. ### Extended lookup capabilities Real input is rarely clean, and the highest raw match score is not always the right answer. Several capabilities sit on top of lookup and do the work a data expert would otherwise do by hand: | Capability | What it does | | --- | --- | | Smart re-ordering | Ranks candidates by likely intent: prominence, headquarters versus branch or subsidiary, legal-entity type, source authority, and a preference for domestic sources. A search for a large group returns the group headquarters from the national register first, with subsidiaries below. | | Identifier detection | Recognises what an identifier is from the string alone (VAT ID, LEI, DUNS, register or tax number, Peppol ID, GLN), infers the country, normalises the value, and switches to an identifier search instead of a name search. | | Alternative strategies | Treats a failed query as unfinished. Retries differently: drop or expand the legal form, search the identifier instead of the name, relax street to city to country, restore diacritics or transliterate, and reports which strategy produced the hit. | | Messy input detection | Spots unclean input before searching: name and address in one field, ERP abbreviations, values in the wrong field, `c/o` and attention lines, operational noise. Splits it into search terms and states what was ignored. | | Commercial source lookup | Targets a lookup at a commercial source the customer has licensed (Dun & Bradstreet or Bureau van Dijk) and returns that provider's own key, the DUNS number or the BvD ID, taken from the search result. Retrieving the full provider entry is not available through Genius, so the identifier is handed back for the customer to fetch where the cost is justified. | ### Reaching the improvements from an existing lookup integration The intent-based ordering is not limited to the chat. Customers who already integrate CDQ's lookup endpoint (SAP MDG integrations among them) can apply smart re-ordering to their existing lookup results without building anything new. It is switched on through the augmentation configuration, under the lookup settings, by adding the corresponding entry to `featuresOn`. The result is the same ordering contract on both surfaces: identifier matches dominate fuzzy name hits, the prominent legal entity outranks the unrelated small match, and address agreement and source trust count. No candidate is dropped or hidden, only reordered. ### Duplicate detection Before a record is created, Genius checks the customer's own Data Mirror for the same business partner, asked in plain language, for example *"is this already in our data?"*. Genius forms its own verdict field by field from what lookup returns, without repeating the API's duplicate flag, and places each candidate in one of three tiers: no match, possible duplicate, or confirmed duplicate. It returns where the duplicate actually lives (the external ID plus the real data-source name) and shows the matching and conflicting fields side by side, so the reasoning behind the tier is visible. Borderline cases stay at *possible* and go to the steward with the evidence. A false alarm costs one review, while a missed duplicate propagates into master data. ## Web research: answers where official registers stop Not every business partner is in a register. Sole traders and other non-legal entities have no register entry to find, and in a number of countries there is no public source to query at all. Web research covers that gap. When CDQ's trusted sources cannot settle the question (no authoritative match, several candidates with none decisive, or evidence too weak to conclude), Genius resolves the entity from public sources instead of returning a dead end. The benefit for the steward is a defensible answer where there previously was none, and a record that can be contributed back to CDQ's data. Every fact is tied to the exact page it was read from, re-verified against that page or dropped, and graded. What arrives is not a paragraph of prose but a set of evidence items, each carrying its source and confidence level. Renames, mergers and trading names are noted. The steward reads which facts are solid and which are thin, and decides accordingly. Grading uses the CDQ trust framework, which weighs source authority (official registers outrank generic pages), corroboration from independent sources, and match strength across name, address and identifiers: | Confidence | What it reflects | | --- | --- | | `LOW` | Weak signals only: no authoritative source, little or no corroboration. | | `MEDIUM` | Partial evidence: some agreement, but not from an authoritative source or not independently confirmed. | | `HIGH` | Official sources, corroborated by independent evidence. | This is where Genius differs from a general-purpose language model. A plain model produces a fluent answer with no way to tell how much of it holds. Genius adds purpose-built research skills that know which sources to try and how to retry when one fails, CDQ's domain knowledge of registers, identifiers, legal forms and country specifics, and a confidence framework that grades each finding against its source. The output is evidence a steward can act on and defend in an audit. ## Data quality: curate, validate, qualify Once identity is settled, Genius runs the data-quality capabilities and explains each result. ### Curate Genius curates a provided business partner through Partner Guard and explains why the partner is curated the way it is. Every change is reported with its concept, field, action and value, together with the source that produced it: a trusted source such as VIES or Peppol, or a system-derived value. Unchanged fields are stated as unchanged, so the full picture of what curation did and did not touch is visible. ### Validate Genius validates business-partner attributes against CDQ's data-quality rules, flagging invalid, missing or inconsistent values. It shows which fields fail and why (for example an incomplete legal address as the common cause behind several errors) and offers the next step, such as running address curation before continuing. ### Qualify Genius qualifies a business partner by bringing identity, tax and address evidence from authoritative sources into one summary, cross-checked against VIES. The summary names the authoritative source behind each value and when it was last synchronised. Where the qualification engine cannot complete a formal score, Genius reports that explicitly alongside the source-backed data. ## Bank account verification Genius verifies a bank account from an IBAN in the same conversation as the rest of onboarding, instead of a separate tool consulted just before payment. A verification returns: - the account's trust score - the account and owner details - the associated payment information - whether the account matches a known fraud case Where a fraud case matches, the answer leads with a prominent warning not to pay and includes the case details behind the flag, so the finding can be checked rather than taken on faith. Format problems in an IBAN or BIC come back explained in plain language, not as an error code. ## Compliance: sanction screening Genius screens a single business partner against compliance and sanction lists on request. Every hit is returned with its list type, list name and match probability, plus what the steward can do next. Where the screening response points to a likely false positive (a different country, no identifier overlap), Genius says so and explains why, so effort is not spent on irrelevant matches. When nothing is found, it states that explicitly. Warning The compliance judgement stays with a human: Genius reports what the lists returned and separates plausible matches from noise, but does not conclude whether a partner is sanctioned. Screening is available to users entitled to AML Guard, and batch screening runs outside Genius. ## Knowledge: reference metadata and product guidance Not every question is about a specific business partner. Genius also answers questions about CDQ itself. ### Reference metadata Genius answers questions about CDQ's curated reference metadata: data-quality rules, data sources, country and regional reference data, identifiers, legal forms, registration authorities, bank-account formats and data-model concepts. This is the same knowledge published in the CDQ Metadata Wiki, scoped by topic and country and read live, so answers such as *"which data sources support update monitoring?"* return the current list with technical keys. ### Product guidance Genius explains CDQ products themselves: what each feature does, the value it delivers, and when to use it. Users get product guidance in the conversation instead of searching documentation. When a topic is not covered, Genius says so. ## Summary CDQ Genius connects natural language with CDQ's business-partner data services. Identity is resolved first, from CDQ's trusted sources, with web research covering the cases official registers cannot: non-legal entities and countries without public sources. Curation, validation, qualification, bank account verification, screening and knowledge follow, each answer tied to a named source and each finding carrying its confidence, so a steward can decide quickly and defend the decision later. Existing lookup integrations can adopt the intent-based ordering through the augmentation configuration, without a Genius integration of their own. ## Related Resources - [CDQ Developer Portal](https://developer.cdq.com) ## Your opinion matters! We are constantly working on providing an outstanding user experience with our products. Please share your opinion about this tutorial!