Batch windows. Deterministic rules. Licence meters that charge by the master row. Stewardship queues nobody can clear. Traditional MDM solved the 2010 problem beautifully — and it is now the single largest bottleneck between enterprises and their agentic AI ambitions.
This is a practitioner's argument for what replaces it, how it is architected, what it costs, and how you get there without a big-bang rewrite.
Six months of published research tells a coherent story: MDM was written off as legacy plumbing, then abruptly reinstated as strategic infrastructure — because generative and agentic AI cannot function without it.
After retiring the MQ, Gartner reinstated it in April 2026 (Kennedy, Robison, Radhakrishnan), naming Profisee, Salesforce (Informatica), Reltio, Stibo Systems and Semarchy as Leaders. The stated rationale is the arrival of agentic and augmented data management — MDM as the trusted context layer that prevents enterprise AI hallucination.
Gartner puts the average annual cost of poor data quality at $12.9M and warns that a majority of AI projects risk abandonment without AI-ready data practices; only about 37% of organisations express confidence in their data management for AI. A Harvard Business Review Analytic Services survey cited by Reltio finds just 15% possess the data foundation agentic AI needs.
Salesforce completed its acquisition of Informatica on 18 November 2025, absorbing catalog, governance, quality, metadata and MDM into the Salesforce platform. For an installed base of 9,000+ organisations running heterogeneous SAP / Oracle / Snowflake estates, roadmap independence is now a live procurement risk, not a hypothetical.
Profisee 2026 R1 ships automated agents, semantic matching and an MCP server; Reltio's AgentFlow Unstructured ingests PDFs, Word and HTML to surface "dark" data. Directionally right — but these are capabilities layered onto hub architectures, commercial models and stewardship paradigms designed for a rules-and-batch world.
On standard record-linkage benchmarks, classic fuzzy matching lands near 37% F1; pure embedding retrieval improves recall but suffers precision (~35%). Hybrid designs — embedding-based candidate generation with an LLM adjudicating only borderline pairs — reach 95%+ precision at negligible per-pair cost. This is the single most consequential fact in modern MDM.
Independent commentary (notably from the open-source entity-resolution community) argues the reinstated MQ under-weights ML-native resolution, cost-per-record economics and the rise of composable, lakehouse-resident mastering. We include the counter-argument deliberately: a point of view that only cites confirming evidence is marketing, not analysis.
These are not product defects. They are the logical consequences of an architecture designed when compute was expensive, data was structured, and "real time" meant overnight.
Every match rule, survivorship hierarchy and validation constraint must be anticipated and encoded before the first record lands. Reality then arrives in a shape nobody predicted — a new source format, a default value stuffed into a key field, a merger that doubles your naming conventions — and the rule set degrades silently.
Batch match-and-merge windows mean operational systems consume a version of truth that is hours to days stale. Fine for a monthly report. Fatal for an AI agent making a credit decision, routing a claim, or answering a customer live.
When you are metered per master record, per domain, per profile or per opaque processing unit, every instinct to master more data — more domains, more history, more external enrichment — becomes a budget conversation. The commercial model is structurally hostile to the data completeness AI requires.
Consumption units that are not clearly defined in compute or storage terms make annual budgeting a negotiation rather than a calculation. Public analysis of Informatica's IPU model notes it "isn't clear what precisely an IPU is," with third-party commentary citing budget overruns of 40–60% against initial projections. Treat vendor-adjacent figures with appropriate scepticism — but the pattern is consistent across buyers we have worked with.
Low-confidence matches fall to a human queue. Because rules cannot generalise, the queue fills faster than any team can clear it — and it fills with near-identical tasks. Vendor patent literature itself acknowledges "a large number of very similar tasks in the queue for the attention of a data steward," causing downstream delay in master data updates.
Contracts, onboarding PDFs, support transcripts, regulatory filings, emails and scanned KYC packs contain the strongest identity evidence in the enterprise — and traditional MDM cannot read any of it. Mastering runs exclusively on structured columns while the corroborating evidence sits unused in a document store.
Traditional enrichment means buying a reference file and joining on a key. It cannot reconcile that a customer's registered entity changed name after acquisition, that two addresses are the same building, or that a "new" prospect is an existing subsidiary — facts freely available in public registries, filings and web sources.
Deep rule chains are technically deterministic yet practically inscrutable. Under GDPR, DORA, BCBS 239 or model-risk review, "rule 4,412 fired" is not a lineage narrative. Ironically, a well-instrumented AI pipeline that records the evidence and confidence behind each decision produces a better audit trail than the rules it replaces.
Data model workshops, source profiling, rule authoring, UAT, cutover. By the time the hub is live the sponsoring executive has moved on and the business case has been re-litigated twice. Enterprise MDM programmes routinely spend $150K–$300K on implementation services in year one alone, before licence.
Match rules, survivorship, workflows and cleansing logic live in vendor-specific configuration stores with no meaningful export. Combined with platform consolidation — Salesforce absorbing Informatica being the clearest recent example — this converts a technology choice into a decade-long strategic dependency.
The critique above is not theoretical. Below is a real multi-domain Informatica MDM programme R2 Digital delivered — customer, membership, certification and product domains across a dozen source systems. It was a well-executed implementation, faithful to the vendor's reference architecture. Its constraints are exactly the constraints of the era it was built in.
The design included a documented alternate solution: "if there are performance or HA issues, deploy an MDM/RDM cache on every business application, refreshed at low latency." That is a sophisticated, correct engineering answer — and it is also an admission that a batch-anchored hub cannot reliably serve real-time reads at enterprise concurrency. In the AI-native design, real-time resolution is the primary path, not the fallback, and the cache disappears along with the six replicas of truth it created.
Landing and staging were both full daily reloads to keep MDM decoupled from source systems. Entirely reasonable in 2019. It also means the golden record's freshness ceiling is 24 hours no matter how much real-time integration sits on top — and no agentic use case tolerates that. Streaming CDC into an open table format removes the constraint without giving up the decoupling.
Informatica's Trust Framework assigns a trust factor per column per source — genuinely ahead of its time, and the conceptual ancestor of learned survivorship. But those factors were set by hand in workshops and then rarely revisited. The AI-native version keeps the concept and makes the weights observed: measured against outcomes, recalibrated continuously, and explainable per attribute.
Hub Console, Data Director, DQ Analyst, DQ Developer, Admin Console, Address Doctor, plus Collibra and Tableau. Each is competent; collectively they demand a specialist per console. This is why MDM programmes carry the headcount they do — and why the single steward cockpit in the AI-native design is a genuine operating-model change, not a UI preference.
Guiding principles, a target operating model spanning consumers, capabilities, technology, processes, governance and roles, a Business Rules Spreadsheet under version control with impact analysis, defined authorised sources, and a change-management chain of accountability. None of this is made obsolete by AI. It is precisely the control framework that makes autonomous mastering defensible. Modernise the engine; keep the governance.
The programme was structured as domain-by-domain shippable increments, prioritised by business value, with a PoC preceding the MVP. That sequencing maps almost one-for-one onto the Phase 1–3 migration pattern later in this page. Organisations that already ran MDM this way migrate fastest, because the hard organisational muscle — incremental domain delivery — is already built.
All five Gartner Leaders are competent products built by serious engineering teams. The critique below is architectural and commercial, not a quality judgement — and it reflects what surfaces in the third and fourth year of an implementation, not the demo.
| Dimension | Informatica MDM (Salesforce) | Reltio | Profisee | Stibo Systems | Semarchy |
|---|---|---|---|---|---|
| Licence metric | Hybrid IPU consumption + per-domain record counts. Enterprise entry commonly quoted around $200K/yr. High friction | Subscription by consolidated profile count, tiered, invoiced annually in advance. High friction | Volume-based subscription; cited by Gartner as transparent. Lower friction | Record/SKU and module-based; PIM-weighted commercials. Medium | Deployment + data-volume based; relatively legible. Medium |
| Master-row & overage exposure | Growth in records and processing both consume budget. Third-party analyses report 40–60% overruns vs. initial projections. Severe | Profile-count tiers create hard cliffs; teams routinely exclude domains or history to stay under tier. Severe | Still volume-metered — mastering "everything" remains a cost decision. Medium | Multi-domain expansion is a re-contracting event. Medium | Scales more gracefully but still couples volume to price. Medium |
| Pricing predictability | IPU definition is not publicly precise in compute/storage terms — budgeting becomes negotiation. Opaque | No public price list; multiple product lines, support tiers and add-ons. Opaque | Most legible of the Leaders. Good | Quote-driven. Medium | Quote-driven, generally simpler. Medium |
| Matching engine | Deterministic + probabilistic rule configuration; tuning is a specialist skill and a standing cost centre. Rules-first | ML-assisted matching with a strong graph model; still fundamentally configured rather than learned end-to-end. Hybrid | 2026 R1 adds semantic matching — genuinely forward-looking, layered on an existing hub. Hybrid | Rules-centric, PIM-optimised. Rules-first | Rules with ML assists. Hybrid |
| Processing model | Batch-anchored heritage with streaming bolted alongside; full reprocessing is expensive. Batch-anchored | API-first and the most genuinely real-time of the incumbents. Strong | Near-real-time; strong Azure/Fabric alignment. Good | Batch-oriented for large catalog syndication. Batch-anchored | Continuous, incremental. Strong |
| Unstructured & external evidence | Limited native use of documents as identity evidence. Gap | AgentFlow Unstructured ingests PDF/DOCX/HTML — the most advanced incumbent answer. Leading | MCP server exposes master data to AI tools; ingestion of unstructured evidence less mature. Partial | Rich asset handling, weaker identity reasoning from text. Partial | Limited. Gap |
| Stewardship experience | Functional workbench; queue-driven, minimal learning from steward decisions. Dated | Modern UI, still task-queue centric. Dated paradigm | Some administration still requires the FastApp Studio desktop client (web migration signalled for 2026). Friction | Deep but complex. Complex | Clean, developer-friendly. Good |
| Deployment constraints | Multi-cloud, but heavy footprint and dedicated administrators required. Heavy | SaaS-only; limited control over residency specifics. Medium | SaaS runs on Microsoft Azure only — a hard constraint for AWS/GCP-standardised enterprises. Constraint | Flexible. Good | Flexible, incl. customer-managed cloud. Good |
| Connectivity breadth | Broadest connector library in the market. Strength | API-first; good coverage. Good | Smaller out-of-the-box connector library than integration-led vendors. Gap | Strong retail/manufacturing ecosystems. Good | Native integration capability noted by Gartner as lagging competitors. Gap |
| Strategic / roadmap risk | Salesforce completed the acquisition Nov 2025; independent commentary expects R&D to prioritise Agentforce/Data Cloud integration over standalone multi-platform capability. Elevated | Independent; capital-intensive growth model. Medium | Independent; deep Microsoft dependency. Medium | Privately held, stable. Low | Independent. Low |
| Portability of your logic | Proprietary metadata; exit cost is effectively a re-implementation. Locked | Proprietary config. Locked | Proprietary config. Locked | Proprietary config. Locked | More code-like, better versioned. Partial |
The traditional MDM business case forced a choice: tighter matching meant more steward hours; broader coverage meant a bigger licence. AI-native design breaks all three constraints at once — and the mechanism is specific, not hand-waving.
An embedding of "Intl. Business Machines Corp, Armonk NY" sits near "IBM Corporation, 1 New Orchard Rd" in vector space. No edit-distance function achieves that. Layer an LLM adjudicator over only the ambiguous band and precision on published benchmarks moves from the ~37% F1 range of fuzzy matching to 95%+ precision — at pennies per thousand adjudications.
Each steward decision becomes a labelled training example. Active learning routes the most informative ambiguous cases to humans rather than an arbitrary confidence cut-off, so a few hundred decisions retune the model where thousands of rule edits previously could not. The queue shrinks structurally rather than through headcount.
The largest saving is not model efficiency — it is the removal of a per-master-row licence meter. When mastering runs on infrastructure you already own (lakehouse compute, your own vector index, open-weight or API models on measured inference), cost tracks usage, not the size of your customer base. Enrichment you would previously have bought as a feed becomes a reasoning step over public sources.
Illustrative, normalised to 100 at year-1 traditional spend. Shape, not magnitude — actual numbers depend entirely on record volume, domain count and estate complexity.
The crossover typically lands in the second year. The strategic point is the direction of the curve: traditional MDM gets more expensive precisely as you succeed at it.
The design principle: master data stops being a destination system and becomes a trust and context service — resolved continuously, exposed through APIs and MCP, and consumed as readily by autonomous agents as by a BI dashboard.
Agentic systems fail on ambiguity, not on reasoning. If an agent cannot answer "is this the same customer?" deterministically and instantly, every downstream action inherits the doubt. Tier 5 exists precisely so that agents consume resolved entities rather than re-deriving identity from raw tables.
Gartner has cautioned that a majority of agentic analytics efforts relying solely on protocol-level access (MCP) will fail by 2028 without semantic foundations. Exposing tables to an agent is not the same as exposing meaning. The ontology in Tier 2 and the resolved entities in Tier 5 are what make MCP useful rather than merely convenient.
Policy, risk appetite, autonomy thresholds, edge-case adjudication and accountability. The architecture increases human leverage; it does not remove human authority. Any design that cannot name what remains human is not ready for a regulator.
This is where the argument becomes concrete. Below, each classical MDM capability is set against its AI-native equivalent — what changes mechanically, and what it changes commercially.
Nobody replaces a production MDM hub in one cutover, and nobody should try. The pattern that works is a strangler-fig migration: run the AI-native layer in shadow, prove it domain by domain, and retire the legacy hub only when it has nothing left to do.
| Diagnose & baseline | ||||||||||||
| Shadow resolution | ||||||||||||
| Domain 1 production | ||||||||||||
| Coexistence waves | ||||||||||||
| Legacy decommission | ||||||||||||
| AI activation (MCP/API) | ||||||||||||
| M1 | M3 | M5 | M6 | M8 | M9 | M11 | M12 | M14 | M15 | M17 | M18 | |
Permanent hybrid. Coexistence without a committed decommission date becomes two platforms, two cost lines and two versions of truth. Set the legacy exit date in Phase 2 and defend it.
Sequence the migration against your renewal calendar. The single largest financial lever in the programme is arriving at renewal with credible, demonstrated optionality — not arriving with a slide deck.
The instinct is to port thousands of legacy match rules. Resist it. Use them as test cases for the new resolver, not as its specification. Porting the rules ports the problem.
Architecture diagrams do not win transformation programmes; the people who have to live inside them do. Here is the same steward, same organisation, before and after.
Up from 380 yesterday. It has never gone down. She sorts by age and starts at the top, knowing she will not reach the bottom.
Eleven variations of the same regional distributor. Same judgement, eleven times. The system will present the same eleven again next month when new records arrive.
A rural Indonesian address the postal library rejects. She emails the local sales team and moves on. It will sit unresolved for nine days.
Asked why two entities merged in March. The answer is somewhere inside a 4,000-condition rule set. She books time with the MDM developer, who is on another project.
A newly acquired subsidiary's CRM. Configuration effort is estimated at six weeks. She schedules it for next quarter.
Last night's run merged 240 records on a default value stuffed into a tax-ID field. Unmerging is manual and takes the rest of the afternoon.
She cleared 61 tasks. 97 arrived. The backlog grew again.
Not the least-confident cases — the most informative ones. Each arrives with the model's reasoning, the evidence considered and the alternative it rejected.
The eleven distributor variants are clustered into a single decision. She sets the policy once; the system applies it to this cohort and to every future instance.
The profiling agent detected drift in a new source's date format and drafted a validation rule in plain language. She amends one condition and approves it. No developer involved.
The lineage view shows the exact evidence, confidence score and reasoning behind the March merge, with citations to the source documents. She exports it as an audit artefact.
The profiling agent maps and proposes standardisations; she reviews and ratifies. Live in three days instead of six weeks.
Sits with the commercial team on customer hierarchy definitions for the new market entry, and tightens the autonomy threshold on a high-risk attribute after reviewing the week's sampling report.
Her decisions retrained the resolver. Tomorrow's queue will be smaller than today's — structurally, not heroically.
A free 45-minute consultation with R2 Digital. We will look at your current MDM footprint, your renewal timeline and your AI roadmap, and give you a straight answer on whether an AI-native approach is worth pursuing — including when it is not.
Published analyst commentary, vendor material and research cited in this analysis. Several sources are commercially interested parties; they are included because they are the primary record of the claims made, not because they are disinterested.
A 30-minute working session with R2 Digital. Tell us a little about your current state and we'll come prepared.