Handling Multi-Format Royalty Data With AI

The mechanism matters more than the headline number. When an ISRC is wrong, when a songwriter's name has a typo, when an IPI number is missing, DSPs cannot match the stream to a verified owner. The money goes into a holding account. Over time, those funds either expire or get redistributed to major rights holders by market share. The independent artist with bad metadata loses twice: the royalty itself and any practical leverage to recover it.

The financial scale of this failure is not abstract. Industry estimates put unclaimed royalties globally somewhere between $2 billion and $4 billion annually, all of it attributable to missing or inaccurate metadata. The Mechanical Licensing Collective has distributed over $2 billion since its 2021 inception, but it also holds close to $1 billion in unmatched funds. At roughly $60 million received per month and an approximately 85% identification rate, around $9 million flows into that unmatched pool every single month. Muso.AI estimates that 83% of songwriters have an MLC mismatch issue. The black box is not a fringe problem. It is the default condition.

The harm is asymmetric, and that asymmetry is the actual point. Independent artists and small publishers lack the back-office infrastructure to audit statements, file disputes, or chase mismatches across jurisdictions. The people most likely to have metadata errors are precisely the people least equipped to catch them. Large rights holders have teams. Small ones have none. Money moves from one pocket to another, quietly, every month.

Why human-led reconciliation breaks at the volume streaming creates

Paul Goldman at Muserk has said plainly that there is no way for artists or publishers to handle this work manually anymore, regardless of their size or resources. That is not rhetorical. It is an operational reality, and it follows directly from the math.

Luminate's 2025 data puts new track uploads at roughly 106,000 per day. Each one is a new data point entering an already saturated system. The revenue streams attached to those tracks are fragmenting further: performance royalties around 42% of publishing revenue, mechanicals around 35%, sync around 15%, and emerging categories like gaming, VR, and user-generated content claiming a growing share year over year. Each category runs on different contractual and reporting logic. CISAC's 2025 report puts total global music creator royalties at €12.59 billion in 2024, up 7.2%. More money at stake means more transactions to process. The data volume problem compounds annually, and the back-office hasn't kept pace.

Cross-industry research from McKinsey and Deloitte has found that finance teams spend roughly 20 to 30% of their time on manual data processing. Royalty back-offices face that same drag, plus an additional layer of format inconsistency that most finance departments never encounter. Manual reconciliation requires ingesting statements in mismatched formats, normalizing column headers, date formats, and territory codes, then matching ISRCs and ISWCs across sources that may carry the same song under different identifiers, then flagging anomalies across all of it. Multiplied by catalog size and active income streams, the throughput requirement exceeds what human capacity can address before transactions expire or get redistributed.

Prager Metis put it precisely: with the increasing size of royalty statements, which contain huge numbers of low-value transactions, AI is going to play a significant role in the royalty audits of the future. Volume is the ceiling, not complexity. A skilled analyst can handle a complicated royalty dispute. No analyst can cross-reference billions of low-value transactions before they expire.

How AI handles format ingestion before reconciliation begins

A catalog receiving statements from dozens of DSPs, multiple CMOs, and social platforms will get data in incompatible schemas. A column labeled "territory" in one feed appears as "market" in another and "region" in a third. Date formats vary. Rights type taxonomies do not align. Traditional ETL pipelines built on fixed rules require a custom parser for each source. When a DSP silently changes its export schema, the rule-based pipeline breaks silently. The error is invisible until someone notices a discrepancy downstream, often well after the window to dispute has closed. I have seen catalogs lose entire quarters of recoverable income because a format change went undetected for two reporting cycles.

AI-based ingestion handles this differently. It processes unstructured and semi-structured inputs, including PDFs, non-standard CSVs, and varying XML schemas, without requiring a new parser for each one. It classifies ambiguous fields by learning from labeled examples across ingestion runs. When a source format drifts, the system flags the anomaly rather than absorbing the error. Error rates in AI-driven financial data automation can fall below 0.1%, compared to the compounding human error that produces the metadata mismatches described above.

The practical output is a normalized data layer: streams from different sources mapped to a common schema with consistent ISRC, territory, period, and rights type fields, ready for matching. One thing to understand clearly: ingestion AI still requires initial configuration. Someone must define the target schema and provide labeled examples of each source format. The learning is supervised, not automatic. But that configuration is a one-time investment, not a per-statement cost, which is exactly where the efficiency gain compounds over time.

Matching and identity resolution across incompatible catalogs

The identity problem is structural. The same song can appear under a different ISRC on a re-release, a different ISWC at a CMO that registered it separately, a variant spelling of a songwriter's name, or no identifier at all in a social platform's reporting. Traditional matching relies on exact-match lookups against ISRCs and ISWCs. Those lookups fail silently when the identifier is wrong, missing, or near-duplicate. The work goes unmatched, and the money follows it into the black box.

Blokur built a different approach grounded in graph theory. Rather than treating works and identifiers as rows in a lookup table, the system treats them as nodes in a relationship graph. Works, contributors, identifiers, and registrations all carry relationships to each other. A work registered under two different ISWCs at two different CMOs can be linked through shared contributor IPI numbers and phonetic name similarity, even when no direct identifier match exists. Music Reports integrated Blokur's technology to create Trakdex, and that combined system was selected for the MLC's Supplemental Matching Network. That institutional selection matters. It signals that graph-based matching has crossed from experiment to operational infrastructure.

For an independent publisher or self-managed artist, this is concrete. The matching problem previously demanded human expertise: knowing which CMO holds which registration, understanding how to file disputes across jurisdictions, maintaining relationships with collection societies. Graph-based AI can increasingly handle that reasoning systematically. Small catalogs with thin registration histories give the graph less evidence to work with, so the returns scale with catalog depth, but the access to systematic matching itself is the meaningful shift. That capability used to belong exclusively to organizations with dedicated staff.

Anomaly detection and forecasting once the data is clean

BMG's StreamSight, built with Google Cloud, is a production example of what this layer looks like at scale. Machine learning models trained on historical streaming data forecast future royalty income by market and rights type, then flag anomalies against those baselines. Problems that would previously require hours of manual review get surfaced automatically, often before a human analyst would have opened the statement.

What anomaly detection catches is specific. A DSP reporting zero streams for a catalog in a territory where the prior three periods showed consistent volume is almost certainly a reporting failure, not a real drop in listening. A spike in revenue tied to an unfamiliar rights claim could indicate an unauthorized sub-license or a split dispute. A revenue figure inconsistent with known contractual rates signals rate misapplication, which turns up more often than it should in complex co-publishing deals. Each of these gets buried in statement volume when the review process is human-led and working in arrears. The irregularity is there; no one has time to find it.

The forecasting side has operational consequences beyond auditing. Knowing projected royalty income by period and territory helps you plan cash flow, prioritize catalog investment, and negotiate advances from an informed position. Instead of reviewing last quarter's statement for errors after the fact, you can track live anomalies and project next quarter's income in real time. If you currently operate entirely in arrears, that shift is not marginal. It is a fundamentally different way of running your business.

Recovering royalties that have already slipped through

Muserk is the clearest real-world example of recovery AI operating at scale. Founded by Paul Goldman, the company uses patented technology called Blue Matter to search billions of lines of streaming data across major platforms. It has recovered more than $100 million in royalties globally, fees that were not being reported through traditional channels. Clients have seen a 100% increase in first-year royalties and 422% cumulative growth over three years.

What recovery AI is actually doing is cross-referencing usage data against rights registrations to find instances where a work was streamed or monetized but no corresponding royalty payment was triggered or received. The volume problem that defeats human-led processes becomes an asset here. No human team could cross-reference billions of data points against a rights database at the speed needed to catch low-value, high-frequency transactions before they expire or get redistributed.

The recovery opportunity is not shrinking. With approximately $9 million added to the MLC's unmatched pool every month and the broader global unclaimed pool estimated between $2 billion and $4 billion, the recoverable income is large and continuously replenishing. If you are an independent publisher or self-managed artist, recovery AI gives you access to systematic royalty auditing that was previously available only to rights holders large enough to fund manual audits. The structural inequality does not disappear. But the gap narrows in a way that simply was not possible before these tools existed.

AI fraud detection and the emerging royalty tier problem

Fully synthetic tracks are entering DSP catalogs and accruing royalties under the same structure as human-created work. Deezer launched AI detection technology in June 2025 to filter fully AI-generated tracks for royalty purposes, with stated capability to identify 100% AI-generated content. The direction other major platforms are moving points toward three-tier royalty structures: fully human-made content in the first tier, AI-assisted content in the second, fully AI-generated content in the third, carrying minimal or zero royalty rates.

For rights holders and publishers, this introduces another metadata field that must be accurately captured and reported: how a given work is classified under the emerging tier system. For the royalty data pipeline, AI detection outputs become a new input format. The multi-format problem gains another dimension rather than getting simpler. CISAC's 2025 report flags AI as a threat to creator revenue even as global collections grow. The fraud detection layer is one institutional response, but the regulatory and contractual frameworks governing AI-generated content are still forming. If you are building a pipeline today, you have to build it knowing that this piece of the schema is unsettled, and it will keep moving.

What a working AI royalty pipeline looks like end to end

The stages follow a clear sequence, even if the implementation is not simple.

Ingestion comes first. Statements from all sources are normalized into a common schema, and format anomalies are flagged before any reconciliation begins. Matching follows: works are identified across incompatible identifiers using graph-based or ML matching, and orphaned registrations are surfaced for resolution. Anomaly detection runs continuously against historical baselines, routing irregularities for review rather than burying them in statement volume. Recovery search cross-references usage data against rights registrations to surface unreported or uncollected income. Forecasting projects income by market and rights type, informing cash flow planning and catalog investment decisions. Classification flags AI-generated content for emerging tier treatment as detection standards and royalty structures develop.

Each stage depends on the one before it. Bad ingestion produces bad matching. Bad matching creates blind spots in anomaly detection. Bad detection means recovery searches are working against incomplete data. Every failure at ingestion echoes through every subsequent layer. This is why the format problem has to be solved first, before anything downstream can function reliably, and why organizations that treat ingestion as a commodity step tend to be the same ones still puzzling over anomalies they cannot localize.

If you treat this as a technology investment rather than a back-office cost, you build a durable advantage in catalog management. The tools exist. The royalty pipeline problem is solved in pieces, by people who have deployed these systems in production and measured the results. The only real question is whether you deploy them before the unmatched pool grows larger and the tier problem adds another layer of complexity to a system that is already difficult enough.