AI Doesn’t Fix Bad Data, It Inherits It
The structural debt, the finance data model, and the question to ask before scaling
AI investment in finance is now macro-scale. Finance teams still spend close to half their time fixing data. And the organisations getting real value from AI are those that entered the wave with cleaner data, tighter controls and assurance-ready governance. The binding constraint is not algorithmic sophistication. It is accumulated structural debt in the finance data model.
If your finance systems produce impressive reports that controllers still do not trust, this article is for you.
The Data
Recent studies from Gartner, FP&A Trends and KPMG point in the same direction: AI spend is rising fast, while finance data foundations remain weak.
Gartner’s 2026 forecast indicates that total AI-related spending will reach around 2.5 trillion dollars by 2026 (Gartner, 2026). At the same time, roughly a third of finance leaders cite poor data quality as the single biggest barrier to adopting AI in finance (Gartner, 2025).
The 2025 FP&A Trends Survey provides granular evidence (FP&A Trends Group, 2025). Only 2 percent of organisations consider their FP&A teams optimised. More than 60 percent are constrained by manual processes and inconsistent data. Only 17 percent rate their data quality as good. Most strikingly, 46 percent of FP&A time is still spent on data collection and validation rather than analysis and decision support.
KPMG’s 2026 Global AI in Finance report, based on 1,013 senior finance leaders across 20 countries and 13 sectors, adds a third angle (KPMG, 2026). Active AI use in finance has more than doubled in two years, and around 71 percent of organisations report that AI is meeting or exceeding ROI expectations. The strongest gains concentrate in judgement-heavy work. Data quality, integration and assurance readiness emerge as key differentiators between those seeing performance at scale and those not.
Three findings hold across all three sources:
AI investment in finance is now macro-scale, not experimental.
Finance teams still spend a large share of their time fixing data instead of analysing it.
Organisations with cleaner data and stronger governance see disproportionate gains from AI.
AI is amplifying the existing architecture rather than compensating for it.
What AI Actually Learns in Finance
AI models in finance learn whatever patterns the data allows them to see, including structural inconsistencies. When an experienced analyst encounters an odd variance, they apply context and professional judgement. They know that a reorganisation took place in Q2 last year, that a business unit shifted cost centres, or that a new revenue recognition policy was introduced.
A forecasting model trained on three years of P&L data does not possess that contextual buffer. If cost-centre structures were reorganised twice during the period and historical data was not consistently remapped, the model will treat those structural breaks as part of the normal pattern. The outputs can appear plausible, even statistically well-calibrated, while encoding distortions that a controller would have discounted almost instinctively.
This is not primarily a failure of the algorithm. It is a reflection of the data-generating process. The model has inherited the organisation’s data debt and turned it into parameters.
One effective response is to carve out restated history domains, where past reorganisations are systematically remapped, and expose those through governed data products. Models then learn from a stable slice of history rather than the full, inconsistent landscape. But this only works when breakpoints, remappings and lineage are visible in the underlying data structures. Where those foundations are absent, each new AI use case spreads the same structural flaws to more decisions.
Two Decades of ERP and the Finance Data Model
Enterprise resource planning systems were rolled out across the 1990s and 2000s with a promise of integration: one system, one truth. A substantial body of research has since documented a more ambivalent reality: in many organisations, ERP reshaped fragmentation rather than eliminated it.
Charts of accounts that are coherent within each business unit but difficult to compare across entities. Cost-centre structures redesigned multiple times without systematic restatement of historical data. Integration landscapes that evolved into layered combinations of ERP modules, consolidation tools, treasury systems and spreadsheet-based shadow systems.
Any group finance team trying to rebuild a ten-year trend across three reorganisations recognises the practical consequences immediately: the data exists in multiple systems; the comparability does not.
AI Removes the Human Buffer
For much of the last two decades, finance functions have managed data debt through human buffers rather than structural fixes. Controllers and analysts acted as intermediaries between imperfect systems and decision-makers, spotting anomalies and applying judgement where systems fell short.
AI changes the economics of this arrangement. When models are embedded into forecasting, performance measurement or decision support and run at scale, the scope for manual intervention shrinks. Errors that previously surfaced as visible anomalies now live in model outputs that look coherent until they are acted on.
KPMG’s 2026 report highlights this shift directly. Organisations that describe themselves as ‘assurance-ready’ report three to six times the rate of significant improvement in core finance metrics compared to those that are not. Trust, operationalised through governance, controls and human oversight, is not a brake on AI. It is a precondition for scaling it safely.
AI does not primarily change what good governance looks like. It changes the cost of not having it.
Three Practical Moves Before You Scale
Pick one critical domain and make the data genuinely comparable. Start with a single, high-impact area such as P&L by segment or cash flow by business unit, and ensure that reorganisations, chart changes and historical series are consistently restated. Appoint a temporary data product owner for that domain with a clear mandate. The goal is not a perfect data estate, but one end-to-end domain where models can learn from structurally coherent history.
Link AI deployment to governance and assurance from the outset. Treat auditability, explainability and human oversight as part of the design brief, not as after-the-fact compliance. If you cannot produce a clear lineage from source data to model output at reasonable cost, the model is not ready to influence consequential decisions.
Shift FP&A capacity from data firefighting to data architecture. Nearly half of FP&A time is absorbed by data collection and validation. That is the wrong problem to keep solving manually. Redirecting a fraction of that capacity into improving the underlying finance data model, through common definitions, unified sources and documented transformations, produces more durable results than another dashboard layer on top of the same broken foundations.
A Structural Question, Not a Tool Question
The persistent pattern across these studies points to architecture, not tooling. Ask a structural question, not a technical one:
If a model learned everything your current finance data can teach it today, would you trust the decisions it is empowered to influence?
If the honest answer is ‘not yet’, the next investment decision is not primarily about which AI tool to choose. It is about which elements of your finance data architecture, governance and assurance need to change so that the answer can plausibly become ‘yes’.
#FinanceTransformation #AIinFinance #DataQuality #ERP #FP&A #KaldLabs