Blogs

Our most recent community show took a deep dive into Reltio Reference Data Management (RDM) and introduced RDM Intelligent Mapping, also known as Autopilot — a new AI-powered engine designed to eliminate manual mapping backlogs. The session, led by Abhilasha Sabharwal, Product Manager at Reltio, covered RDM fundamentals, core capabilities, and a live demo of Autopilot in action.

Why Reference Data Quality Matters

Reference data — the shared classifications and lookup values that make master data and transactional data consistent across systems — may seem small, but its impact is far-reaching. Poor reference data quality can lead to operational drag from manual rework, revenue leakage from incorrect classifications, and compliance risk from issues like wrong tax codes or failed audits. Traditional approaches relying on spreadsheets and IT-driven remediation simply don't scale.

Core Capabilities of Reltio RDM

Abhilasha walked through the foundational vocabulary and capabilities that make up Reltio RDM:

• Transcoding: Simple, reverse, and hierarchical transcoding automatically convert source system codes into standardized canonical values, and back again for outbound integrations

• Spreadsheet-like UI and drag-and-drop remediation for resolving unmapped values

• Language localization for managing display labels across multiple languages without touching underlying codes

• Reference data hierarchies supporting multi-parent, multi-child relationships

• User-defined attributes allowing up to 20 custom fields per canonical value

• Auto-code generators, allowing users to generate codes automatically

• Data change requests, version control, audit trails, and role-based access to support governance at scale

Introducing RDM Autopilot

The centerpiece of the session was RDM Intelligent Mapping, or Autopilot — built to solve the persistent challenge of unmapped values. Autopilot uses a multi-signal AI scoring pipeline combining vector-based semantic search, fuzzy matching, and configurable LLM reasoning to automatically map values, generate data change requests (DCRs) for steward review, or recommend new lookup values entirely.

The live demo showed Autopilot resolving unmapped medical specialty values in real time — auto-mapping high-confidence matches like "cardiovascular medicine" to "Cardiology," generating a DCR for a lower-confidence variant, and flagging a misspelling before it could enter the canonical dataset.

Key features include:

• Two operating modes: Preview mode scores values silently with no action taken, while Active mode allows Autopilot to auto-map, create DCRs, or ignore values based on configurable thresholds

• Frequency-based gating to prevent one-off or erroneous values from being auto-mapped prematurely

• Configurable thresholds for auto-mapping, DCR suggestions, and new lookup creation — adjustable globally or per lookup type

• Full audit trail showing the reasoning behind every decision Autopilot makes

Q&A Highlights

Attendees asked detailed questions about how Autopilot fits into real-world governance needs:

• Source-level control: Autopilot doesn't currently filter by source system, but thresholds can be tuned per lookup type to manage risk from less-trusted sources

• Reasoning logs: Available at all times, in both preview and active mode, by clicking into any value

• Licensing: No separate license required — Autopilot is available to all RDM customers

• Learning from history: Autopilot learns from prior mapping decisions, and adding descriptions to lookup types improves AI context and accuracy

• Hierarchies: Not yet supported for hierarchical relationships, but planned as a future enhancement

• Import functionality: A long-requested CSV/JSON import capability is in development and coming soon

Getting Started

Abhilasha recommended a phased rollout: consolidate reference data into RDM as a single source of truth, start Autopilot in preview mode to build confidence, then move to active mode — beginning with one or two lookup types before expanding tenant-wide. Full documentation and a best practices guide for Autopilot are available on the Doc Portal.

Please note: If you asked for enhancements during this community show, or have any further enhancements you’d like to see for Reltio products, please use our Idea Portal, “Aha!” to submit your enhancement request.  In the meantime, Abhilasha has responded to the questions that were asked below. If you have questions that weren't addressed live, feel free to post them in the comments section!

FULL Q&A:

• When you say ‘no action’ does that mean RDM will function as it does today? i.e. manual mapping?

  • Yes. If the score does not meet either threshold, no action is taken. That value simply stays an unmapped value, exactly as today — it's still available for your normal manual/drag-and-drop remediation. Autopilot is additive: it only steps in when it's confident enough to auto-map or suggest. Anything it isn't confident about is left untouched and behaves exactly like RDM does now.

• Can the 'Create New Lookup' be configured to be applicable only for a particular source ?

  • Not today. Autopilot configuration is scoped at the tenant level and per lookup type — not per source system. You can enable/tune create-new-lookup for a specific lookup type, but you can't restrict it to a single named source. Source diversity is used the other way — as a guardrail: a value generally needs to appear across multiple distinct sources before it qualifies for auto-creation, which stops one system's junk value from spawning a new lookup. If per-source scoping is something you need, that's good roadmap feedback. The customer can create a detailed enhancement request for this on Reltio Idea Portal & we will be happy to pick this up once we have the clarity on the ask.

• Do we still get the reasoning logs after configuring Auto-Pilot in Active Mode? If so, where can we find them, and how can we access those logs?

  • Yes - every decision is logged in Active mode.  Each auto-mapped value carries a full record: the confidence and component scores, the model version, the policy version, who/what made it (mappedBy: S/ystem), a timestamp, and its state.  You can access these two ways:

    • Autopilot match-statistics APIs — Get / Scan match statistics (including by lookup type and by status) return the per-value decisions and scores.

    • The audit trail — every add/change is captured for compliance.

  • The human-readable AI reasoning explanation ("why Autopilot suggested this") is surfaced on the DCR suggestions in the UI. So: structured decision + score logs persist for auto-maps; the narrative reasoning is most visible on suggestions.

• Will Autopilot provide any suggestions  on the "No Action" scenario in Preview Mode/Active mode for all the RDM error/Unmapped values in the tenant.

  • Two Parts:

    • On the "no action" bucket: No. Autopilot only creates a suggestion (DCR) for values scoring ≥ the suggestion threshold (0.75). Below that = no action, no suggestion — they remain unmapped for manual handling. In Preview mode it scores and raises suggestions silently (never auto-maps); in Active mode it auto-maps ≥ 0.90 and suggests the 0.75–0.90 band.

    • "For all unmapped values in the tenant": Autopilot evaluates values as they flow through transcode for an enabled lookup type — it's forward-looking, not a one-time sweep of your entire existing backlog. Recurring backlog values get picked up as they recur; a bulk retroactive pass over historical errors isn't part of the current behavior.

• Is this enabled for all customers ? Or require separate license 

  • Access is role-gated: enabling Autopilot requires create/read/update/delete on the RDM Configuration resource, which ROLE_ADMIN_TENANT has by default. On commercial availability / licensing, it is included in RDM by default, no separate licensing required.

• After configuring Autopilot, do we need to run the Reindex unresolved RDM lookups first?  Does it apply to the existing RDM data in the tenant?

  • It's not a prerequisite to enable Autopilot — it's a downstream cleanup step, and yes, it applies to existing entities. The sequence: once a previously-unmapped value gets resolved (by Autopilot or a steward), you run the Reindex unresolved RDM lookups job so the existing entities that had errored pick up the now-resolved canonical value and stop showing as transcode errors in search. You can run it for All Entities or a selected list of URIs, via the Console (Tenant Management → Jobs) or the API. So: configure Autopilot first; use reindex afterward to refresh existing data once mappings are resolved.

• Does it understand current mappings, and then suggest mappings based on how we have done it earlier i.e. suggested canonical codes and values align to organisation standards?

  • Yes. Autopilot scores against your existing canonical values and known mappings — the semantic index is built from your canonical values, and the alias layer captures your known variations. So its candidates resolve to your existing canonical codes/values and align to your standards rather than inventing arbitrary ones. It also improves over time by learning from your stewards' approvals and rejections. 

• How long will it take Autopilot to read the newly imported data?

  • Effectively near-real-time. The transcode call returns immediately (zero latency) and Autopilot evaluates the unmapped value asynchronously in the background right after. Two timing nuances to set expectations:

    • Auto-mapping a recurring value waits for frequency gating — a value must recur enough times within the sliding window (default ~10 occurrences) before it's auto-mapped, so it's deliberately not instant on first sighting.

    • Once a value is auto-mapped, it appears in MDM dropdowns after the ~10-minute cache refresh (direct API is real-time)

#Featured

#CommunityWebinar

#RDM