Your Finance AI Is Only as Good as Your Data: What SAP's Acquisition Spree Reveals About Why Autonomous Finance Stalls

Here is a pattern that repeats across every finance AI implementation review I have read in the past eighteen months: the AI works. The data does not. The model produces sensible outputs within weeks, but the inputs — vendor records, GL mappings, cost centre hierarchies, intercompany reconciliations — take months or years to clean to the point where the model's outputs are trustworthy. The bottleneck is never the algorithm. Between 2021 and 2026, SAP spent several billion dollars acquiring companies that, at first glance, seem unrelated to its core ERP business. Signavio, a process mining and intelligence platform. LeanIX, enterprise architecture management. Taulia, supply chain finance. Reltio, master data management. Dremio, a data lakehouse. Prior Labs, foundation models for tabular data.

The common narrative around these acquisitions — including the May 2026 wave — is that SAP is building out an AI story and needed adjacent capabilities to compete with Salesforce, ServiceNow, and the hyperscalers. That narrative is accurate but incomplete.

The more revealing interpretation is this: SAP's acquisition strategy is a multi-billion-dollar acknowledgment that the data problem in enterprise finance is so foundational that no AI capability — not Joule, not Business AI, not any agent layer — can deliver on its promise until the underlying process and data infrastructure is coherent. SAP bought what its existing platform could not provide: visibility into what processes actually run, how data actually flows, and where the gaps between intended and actual behaviour exist.

That acknowledgment has a direct implication for any finance leader evaluating autonomous finance or AI-in-finance investments. The bottleneck is not the AI. It has not been the AI for some time.

What the acquisitions actually address

Signavio, acquired in 2021 for approximately $1.2 billion, does process mining and process intelligence. Its core function is to read event logs from transactional systems — including S/4HANA — and reconstruct what processes actually look like in execution, not what they are supposed to look like in design documentation. The delta between the two is, in most enterprises, significant.

The reason SAP paid that price for that capability is instructive. SAP's own configuration tooling, its best-practice content, and its accelerators all assume a level of process coherence that most enterprises do not have. Before you can automate a process, you need to know what the process actually is. Before you can train an AI system on financial process data, you need data that reflects a consistent, well-governed process — not a decade of workarounds, manual overrides, and system-to-system reconciliations masquerading as a closing cycle.

LeanIX, acquired in 2023 for approximately $1.3 billion, addresses a different but adjacent problem: enterprise architecture visibility. What applications exist, what data do they own, how do they connect, and what is the authoritative source of record for each data domain. For a finance AI to work correctly, you need to know what system holds the truth about a vendor master record, a cost centre hierarchy, or a bank account. In most mid-to-large enterprises, this is not obvious.

Taulia adds real-time working capital and supply chain finance data, which feeds directly into the cash flow forecasting and liquidity management use cases that finance AI vendors routinely demo first because the data is relatively clean. WalkMe addresses adoption — the gap between a system being configured correctly and users actually using it correctly, which affects data quality at the point of entry.

Reltio, absorbed as part of the same multi-year program, targets master data management directly — the quality and consistency of the foundational records on which every other system depends. And the May 2026 acquisitions bring the strategy to its logical conclusion. Dremio is designed to unify SAP and non-SAP data into a coherent lakehouse layer, collapsing the fragmented source-of-truth problem that has historically required custom integration work. Prior Labs, with its TabPFN foundation models, addresses the quality of the AI layer that operates on that data once it is unified. Together they complete the stack: process visibility, enterprise architecture mapping, adoption, working capital data, master data, unified data access, and a model layer that can handle the irregularity of real enterprise tables.

Read together, these acquisitions describe a platform strategy aimed at closing the gap between enterprise data reality and the idealised data environment that AI systems require to perform as advertised.

The May 2026 additions — Dremio and Prior Labs, following the earlier Reltio move — confirm the pattern rather than change it. Dremio targets the specific problem of unifying SAP transactional data with external sources in a coherent analytical layer. Prior Labs addresses the quality of the models that process tabular enterprise data. Reltio had already taken aim at the master data problem directly. These are not adjacent bets; they are investments in the layers that sit between raw enterprise data and a working AI system.

The real reason finance AI underdelivers

Finance AI projects that stall — and the majority of enterprise AI projects do stall or significantly underdeliver — share a consistent pattern. The AI is not the problem. The data is.

The failure modes are well-documented by this point:

Inconsistent master data. Vendor master records with duplicate entries, inconsistent naming conventions, missing tax identifiers, or conflicting payment terms across systems. A payment matching AI trained on this data learns to replicate the inconsistencies rather than resolve them. An autonomous payment release agent operating on this data will make systematic errors that are costly to unwind.

Process variation that looks like data noise. The actual GL posting logic used by the Czech subsidiary differs from the German one, which differs from the shared service centre. The differences are not documented anywhere; they exist in the institutional knowledge of two finance managers and a series of manual journal entries that appear in the data as anomalies. An AI model reading this data cannot distinguish meaningful variation from noise without that context.

Reconciliation gaps between systems. The ERP holds one version of revenue; the CRM holds another; the revenue recognition system holds a third. In the hands of experienced finance staff, these are reconciled through a close process that is partly manual and partly understood. Handed to an AI system without fixing the underlying source-of-truth problem, the reconciliation either fails or produces numbers that pass automated checks but are wrong.

Stale reference data. Cost centre hierarchies that reflect an organisational structure from three years ago. Profit centre assignments that were correct when the company had two divisions and are incorrect now that it has five. Chart of accounts structures that have accumulated accounts no one uses but no one has deactivated. AI systems that rely on this reference data propagate the staleness into their outputs.

None of these problems are new. They predate AI by decades. What AI does is expose them faster and at greater scale. A human finance team working through a close cycle encounters the bad data and compensates — slowly, with effort, through experience accumulated over years. An autonomous finance system encounters the same data and either fails visibly or produces output at speed that appears correct until it is not.

What SAP is actually solving, and what it is not

SAP's data foundation acquisitions improve the visibility and governance layer. Signavio can map where your close process deviates from best practice. LeanIX can document which system owns which data domain. Together they make the problem legible in ways it was not before.

What they do not do is fix the underlying data. Process mining can show you that your three-way match process has 47 variants and that 12% of invoices follow none of the intended paths. It does not fix the vendor master, retrain the finance team, or resolve the ERP-to-CRM revenue mismatch. It creates visibility; the remediation is still organisational and technical work that requires human ownership.

This matters for the timeline and the investment case for finance AI projects. A Joule-powered autonomous close is not a technology purchase; it is a data remediation program followed by a technology purchase. The AI configuration is often the smallest part of the effort and the fastest to implement. The data readiness work — master data governance, process standardisation, source-of-truth clarification, reference data cleanup — is the long lead-time item.

SAP's positioning rarely acknowledges this sequence explicitly, because acknowledging it accurately would lengthen every deal cycle in the pipeline. But the acquisition pattern acknowledges it implicitly, which is the more reliable signal.

What finance leaders should actually do first

For a CFO or CIO evaluating when and how to invest in AI for finance processes, the actionable implication is a readiness assessment before a technology selection.

The assessment should address three questions.

First: Do you know what your processes actually do? Not what the design documentation says they do — what they actually do, in terms of system events, touchpoints, exception rates, and manual interventions. If the answer is no, process mining is the right first investment. It is not glamorous and it does not appear in AI pilot demos, but it is the dependency that determines whether any AI investment will deliver.

Second: Do you have authoritative data sources for the domains AI will touch? For payment automation: vendor master integrity and bank account data. For cash flow forecasting: consistent, system-of-record revenue and receivables data. For autonomous close: a chart of accounts and cost structure that reflects current organisational reality. Mapping these against what you have is unglamorous work that IT and finance rarely do together. It should be done before any AI vendor is invited in.

Third: Do you have governance for AI outputs? If an autonomous matching agent releases a payment incorrectly, who owns that error? How is it detected, how quickly, and what is the remediation path? Autonomous finance is not a lights-out operation; it is a supervised operation with different supervision requirements than manual processing. The controls and escalation paths need to be designed before the automation runs, not after the first incident.

The durable takeaway

SAP spent billions to acknowledge, indirectly, that enterprise data is not ready for the autonomous finance it is selling. The acquisitions are a genuine attempt to close that gap over the next several years through tooling that makes process and data reality more visible and manageable.

For finance leaders, the takeaway is to use that acknowledgment as calibration. The vendors pitching autonomous close, intelligent payment, and AI-powered planning are selling the end state. The end state is real and achievable. The timeline is determined almost entirely by data readiness — a variable that the vendor cannot control and that the finance organisation can.

The most strategically valuable thing a CFO can do before signing an AI-in-finance commitment is know their own data honestly. That is not a technology question. It is a governance and organisational question that the Signavio dashboards can surface but that only leadership can resolve.

SAP's acquisition spree made the problem visible at the platform level. Making it visible at the company level is a different, more local, and more urgent piece of work.