Machine Learning in Spend Analysis: How the Analysis Actually Gets Done

Written by Philip Solomon and Anna Pilipiuk, Head of Growth Strategy and Product

Once your procurement data is clean, machine learning is what turns it into a spend view you can act on. It classifies transactions to a taxonomy, resolves supplier identities, flags anomalies and forecasts what is coming next, all at a scale no manual team can match.

This is Part 2 of our series on getting real value from spend data. Part 1 covered the work that happens before analysis begins: consolidating sources, standardising formats, removing duplicates and enriching thin records. If you have not read it yet, start with how to improve procurement data quality before spend analysis, because everything below depends on it. This article picks up at the next question: once the data is good enough, how does machine learning actually conduct the analysis?

Why this matters now

Procurement leaders are under pressure to answer harder questions faster. Deloitte’s 2023 Global Chief Procurement Officer Survey found that more than 70% of CPOs reported increased procurement risk or supply chain disruption in the prior 12 months, with cost and inflation risk cited as a high or moderate impact issue by 89% of respondents. Yet only 25% said they could identify and predict disruptions in a timely manner to a large extent. That gap is not a strategy problem, but a data and analysis problem, and it is exactly where machine learning comes into play.

Manual spend analysis does not scale to millions of transaction lines across dozens of systems. Machine learning does. The point of this article is to explain how the main techniques work, where they are reliable, where they are not, and how to keep the outputs defensible to the business.

How machine learning classifies spend

Classification is the core of spend analysis. It is the process of mapping every transaction to a category in a taxonomy, whether that is a standard such as UNSPSC or eClass, or a custom hierarchy built around how your organisation sources.

Most production systems use supervised learning. The model is trained on labelled examples, transactions a human has already categorised correctly, and learns the patterns that predict each category from features such as supplier name, line description, general ledger code, amount and business unit. Once trained, it maps new transactions automatically, typically to three or four levels of depth (for example, IT, then Software, then SaaS Subscriptions).

A few practical points matter here:

• Training data drives everything. Industry guidance suggests a minimum of around 10,000 correctly categorised transactions for a basic taxonomy, with more needed for taxonomies of 100-plus categories. Quality beats volume: a smaller set of consistently labelled transactions outperforms a large set of inconsistent ones.
• Confidence scoring decides what to automate. Every prediction carries a confidence score. High-confidence predictions (commonly cited thresholds sit around 90 to 95% and above) can be auto-classified, medium-confidence items go to quick review, and low-confidence items get closer human analysis. New or unusual suppliers naturally score lower until the model has seen enough similar examples.
• Accuracy is real but context-dependent. Peer-reviewed research classifying procurement text to UNSPSC reported around 93% accuracy on a train-test split using a support vector machine model. Vendor platforms frequently claim 95% or higher on cleaned data; those are supplier claims, not independent benchmarks, and they typically assume good input data.

A note on rules versus learning. Older rules-based classification (if supplier is X and GL code is Y, assign category Z) works reasonably on structured, purchase-order-backed spend but tends to leave a large share of card, services and tail spend in “miscellaneous”. Machine learning is what recovers that unmanaged spend.

Reading messy free text with NLP

A great deal of spend detail lives in unstructured text: line item descriptions, supplier names and invoice narratives. Natural language processing (NLP) is how a model reads it.

In practice, NLP converts text into numerical representations (from simpler term-frequency features such as TF-IDF through to context-aware embeddings from transformer models) so the model can recognise that “mgmt consultancy”, “management consulting” and a specific consultancy’s name point to the same thing. This is what lets a system interpret a cryptic card descriptor or a vague “professional services rendered” line rather than giving up on it. NLP is also the first line of defence on the messiest inputs, correcting spelling, expanding abbreviations and standardising terms before classification runs.

Supplier normalisation and hierarchy detection

The same supplier routinely appears under many guises: different legal entities, trading names, abbreviations and manual typing errors across finance systems. Entity resolution is the machine learning discipline that decides which records refer to the same real-world supplier and links them together.

Modern matching goes beyond exact string comparison, using techniques such as edit distance, n-grams, phonetic matching and probabilistic scoring, then applying confidence thresholds. High-confidence matches are merged automatically; ambiguous ones are routed to a reviewer. The best systems learn from each decision a steward makes and retrain, improving match rates over time.

The same approach reconstructs parent-child hierarchies, so spend with a dozen subsidiaries of one group rolls up to the true parent. This is what makes questions such as “what is our total exposure to this corporate family?” answerable. It matters because hierarchies are perishable: mergers, rebrands and restructures constantly change them, so treating the hierarchy as living data rather than a one-off mapping is what keeps supplier-level analysis accurate.

Finding what should not be there: anomaly and outlier detection

Once spend is classified and suppliers are resolved, machine learning can flag transactions that do not fit the pattern. This is where analysis starts protecting cash directly.

Common signals include duplicate or near-duplicate invoices, price variance against contract, off-contract and maverick spend, split orders designed to dodge approval thresholds, payments to suppliers with no delivery history, and unusual round-number amounts. Models learn the normal behaviour for each category, supplier and requester, then surface deviations, and unlike rule-only checks they can catch novel patterns rather than only known ones.

The financial case is substantial:

• Duplicate and erroneous payments. The Association for Financial Professionals puts duplicate payments at 0.1% to 0.5% of organisational disbursements, while APQC benchmarking finds 0.8% to 2% of total disbursements are duplicate or erroneous payments. At high transaction volumes, that leak recurs quietly.
• Off-contract and maverick spend. The Hackett Group finds roughly 29% of indirect spend is off-contract, and its User Experience and Maverick Spend Study reports that organisations lose up to 16% of their negotiated savings due to maverick purchasing.
• Occupational fraud. The ACFE’s Occupational Fraud 2024: A Report to the Nations, based on 1,921 cases across 138 countries, estimates that a typical organisation loses 5% of its revenue to fraud each year, with a median loss of 145,000 US dollars per case. The report names billing schemes (phantom vendors, inflated invoices) among the highest-risk asset-misappropriation sub-schemes, noting that “check and payment tampering, billing, and theft of noncash assets present the greatest overall risk to organizations.”

Anomaly detection does not prove wrongdoing. It prioritises a human review queue, which is exactly the right division of labour.

Predicting what comes next

Beyond describing the past, machine learning supports forward-looking analysis:

• Demand and spend forecasting. Time-series models learn seasonality and trend to project category spend, giving procurement early warning of budget overruns.
• Price and inflation forecasting. Regression and ensemble models correlate spend with external indices to anticipate category inflation before it compounds.
• Contract renewal and expiry risk. Models flag upcoming renewals and predict renewal likelihood, so teams renegotiate before an auto-renewal locks in legacy pricing. World Commerce and Contracting research (August 2025) found that poor contracting practices erode value equivalent to almost 9% of annual revenue on average, and 15% or more in more complex industries. Missed renewals and unmanaged terms are a big part of that leakage.
• Savings opportunity identification. Classified, resolved data lets models surface consolidation targets and off-contract leakage ranked by potential value.

Forecasts are probabilistic. A useful forecast comes with a confidence interval and an explanation, not a single number presented as certainty.

Discovering opportunities with clustering

Not every insight comes from a label you already have. Unsupervised techniques such as clustering group similar transactions or suppliers without being told the answer in advance. In procurement this is powerful for opportunity discovery: clustering can reveal that the same category is being bought from dozens of fragmented suppliers, that several departments are each using separate vendors for the same service, or that a swathe of tail spend could be consolidated. Because clustering reads patterns at line-item level, it surfaces overlaps that a simple spend ranking would miss.

Where large language models fit, and where they do not

The obvious 2025 to 2026 question is what generative AI and large language models (LLMs) add. The honest answer is: real value in specific places, and real risk in others.

Where they help: LLMs are strong at interpreting unstructured text, summarising contracts, drafting category descriptions, enriching supplier context from public information, and letting users query spend in natural language. In the enrichment stage described in Part 1, a lightweight LLM can read a supplier website or a contract and summarise what is being bought.

Where they are unreliable: LLMs can produce confident, plausible falsehoods, commonly called hallucinations. Research by Kalai and colleagues, published in Nature (“Evaluating large language models for accuracy incentivizes hallucinations“), argues that “metrics such as accuracy systematically reward guessing over admitting uncertainty in large language models”. That is precisely the failure mode you cannot afford when assigning a category or a supplier. It is why classification at scale still often relies on traditional supervised models: they are more predictable, cheaper to run over millions of lines, produce calibrated confidence scores, and are easier to audit. The pragmatic architecture uses classical machine learning for high-volume classification and reserves generative AI for interpretation, enrichment and natural-language interaction, with human validation on anything material.

Manual versus AI-assisted enrichment, expanded

Part 1 introduced the choice between manual and AI-assisted enrichment. Here is the deeper comparison, because this is where most teams either save enormous time or accidentally corrupt their data.

Throughput and cost. The gap in raw throughput is large. Ardent Partners benchmarking illustrates the cost gulf in accounts payable: best-in-class teams process an invoice for about 2.78 US dollars in 3.1 days, against roughly 12.88 US dollars and 17.4 days for other groups. A 2026 multi-agent document-processing study put manual handling at around 4,500 invoices per full-time employee per year, against about 25,000 for fully automated processing and 15,000 for a human-in-the-loop hybrid. The direction of travel is not in doubt: automation is dramatically faster and cheaper per transaction.

Accuracy is not automatic. Pure automation is fast but not flawless. That same study reported roughly 85% accuracy for fully automated classification, rising to around 98.5% when a human reviewed only the flagged minority (about 15%) of documents, while manual-only sat near 95% but at a fraction of the throughput. The lesson: full automation trades accuracy for speed, and full manual trades speed for cost, but a well-designed hybrid beats both.

Why human-in-the-loop is the realistic model. The strongest approach routes only low-confidence and high-value items to people. AI catches the errors humans make from fatigue and inconsistency; humans catch the errors AI makes on ambiguity, novel products and edge cases. This is the model we advocate: automation for the clear patterns, human judgement for the ambiguous, stakeholder input where documents are not enough.

Feedback loops and retraining. Every human correction should feed back into the training data so the model handles similar transactions correctly next time. Without this loop, accuracy stalls; with it, accuracy compounds.

Taxonomy drift. Categories are not static. New suppliers, new products, acquisitions and reorganisations all erode a taxonomy’s relevance over time. A classification approach that is not maintained will degrade. Treating classification as a continuous capability rather than a one-off project is what keeps accuracy from sliding.

When manual still wins. Manual review remains the right call for high-value strategic spend, genuinely ambiguous line items, complex global suppliers that span many categories, and anything that must be defended to auditors or the board.

Why data quality is still the precondition

None of this works on poor data. Garbage in, garbage out is the most common failure point for machine learning in procurement. If descriptions are too vague, supplier records are duplicated or context is missing, the model has nothing reliable to learn from and will infer or guess. Worse, a confidently wrong classification poisons everything built on top of it: category strategies, savings baselines and any downstream AI. This is the direct link back to Part 1. The quality work is not a nice-to-have before analysis; it is what makes the analysis trustworthy. If you have not addressed it, our spend data cleansing and enrichment approach is the place to start.

Making it work in practice

A few implementation realities separate projects that deliver from those that stall:

• Historical data. Pull at least six to twelve months of transaction history across ERP, accounts payable, purchase orders, card and contract data before expecting reliable classification.
• Measure the right metric. Track coverage (the share of spend classified to a useful depth) as well as accuracy on classified spend. High accuracy on a third of spend is not success.
• Keep it explainable. Procurement has to defend its numbers to finance. Favour interpretable outputs, surface confidence scores, and keep an audit trail of how classifications were made. Regulatory direction of travel, including phased obligations under the EU AI Act, is pushing hard towards traceable, explainable AI in financial decisions.
• Govern it continuously. Assign ownership, review flagged items on a cadence, and retrain on corrections.

Common pitfalls and realistic expectations

Expect strong coverage, not perfection. Chasing the last few percent of accuracy on fragmented, low-value tail spend rarely pays for itself. Do not accept a black-box output you cannot explain. Do not treat go-live as the finish line, because models drift. And do not expect machine learning to rescue data that was never good enough to begin with.

The bottom line

Machine learning does the heavy lifting of modern spend analysis: classifying, resolving suppliers, flagging anomalies, forecasting and surfacing opportunities at a scale humans cannot match. But it is a force multiplier for good data and good judgement, not a substitute for either. The teams that win pair automation with human oversight, close the feedback loop, and keep every output explainable.

See how machine learning classifies, scores and explains your own spend. Book a demo. Put clean data and ML-driven analysis to work with our spend analysis services for procurement teams.

---

Sources

1. Deloitte, 2023 Global Chief Procurement Officer Survey. Procurement risk exposure and predictive capability figures.
2. Classification of procurement text to UNSPSC, peer-reviewed research. Support vector machine accuracy of 93% under train-test split.
3. Association for Financial Professionals. Duplicate payments as a share of organisational disbursements.
4. APQC Open Standards Benchmarking, Accounts Payable. Duplicate and erroneous payments as a share of total disbursements.
5. The Hackett Group, User Experience and Maverick Spend Study. Off-contract indirect spend and negotiated savings leakage.
6. Association of Certified Fraud Examiners, Occupational Fraud 2024: A Report to the Nations. Revenue lost to occupational fraud, median loss per case, and billing scheme risk.
7. World Commerce and Contracting, August 2025 research. Value erosion from poor contracting practices.
8. Kalai et al., Evaluating large language models for accuracy incentivizes hallucinations, Nature.
9. Ardent Partners, Accounts Payable Metrics That Matter. Invoice processing cost and cycle time benchmarks.
10. MADP: A Multi-Agent Pipeline for Sustainable Document Processing with Human-in-the-Loop, arXiv, 2026. Throughput and accuracy comparison across manual, automated and hybrid processing.
11. EU AI Act implementation timeline. Phased obligations for AI influencing financial outcomes.