AI in Enterprise Operations: Separating Real Capability from Marketing

The AI claim problem

Every enterprise software platform now claims AI. This creates a genuine evaluation problem for enterprise buyers: when everyone claims AI, the claim conveys no information. You need a way to separate AI that does useful work from AI that is a rebranded rules engine, an LLM wrapped around a search box, or a product roadmap item presented as a current feature.

This whitepaper provides a framework for evaluating AI claims. It is written by Tafkiro — so it reflects our view of what good looks like. We are not the only company doing genuine AI work in enterprise operations; we are offering a framework that we believe applies to us and our competitors fairly.

What AI in finance operations actually does

AI in finance operations has three categories of current practical capability:

1. Detection and flagging: Identifying anomalies in transaction data — duplicate invoices, unusual amounts, mismatched purchase orders, suspicious patterns in expense claims. This is where AI delivers the most consistent and measurable value today. Duplicate invoice detection: a well-trained model across a ledger with 10,000+ monthly transactions can catch 90–95% of duplicates that would otherwise require manual review.

2. Classification and routing: Automatically categorising transactions — assigning expense categories, routing invoices to the correct approver, matching bank transactions to ledger entries. This reduces manual data entry rather than identifying problems.

3. Prediction and planning: Forecasting cash flow based on historical patterns and open receivables/payables, predicting which receivables are at risk of late payment, projecting inventory demand. This is the most valuable category but also the most variable in accuracy — quality depends heavily on data history and consistency.

Most platforms claiming AI are doing (2) and sometimes (1). Fewer are doing (3) with genuine accuracy.

AI in supply chain and operations

In supply chain, AI has practical application in three areas: demand forecasting (predicting sales volume by SKU by period), replenishment optimisation (calculating reorder points that account for demand variability and lead time uncertainty), and quality anomaly detection (identifying production batches with characteristics correlated with higher defect rates).

Demand forecasting AI is valuable when the business has at least 12–18 months of clean transaction history by SKU. With less history, AI forecasting is not meaningfully more accurate than a moving average. With 24+ months of history including seasonal patterns, AI forecasting consistently outperforms manual methods.

For businesses without sufficient clean history — often the case for companies migrating from a system where data quality was poor — the realistic approach is to build the history in the new system first and activate predictive features 12–18 months after go-live.

What to ask in an AI evaluation

When evaluating AI claims from an enterprise software vendor, ask these questions:

"Can you show me this AI feature working on data that resembles mine — not a demo dataset?" A vendor who cannot demonstrate AI on realistic data has likely not deployed it in production at scale.

"What is the recall and precision of your anomaly detection on a real customer dataset?" If they cannot give you a number — even an approximate one from a customer reference — the capability is not mature.

"What data does this AI model require to operate, and what happens if I don't have that data at go-live?" AI models require training data. If the model needs 24 months of purchase history to be accurate and you are migrating from a system with poor data, the AI feature may not be useful for 18+ months.

"Is this AI a separate add-on or included in the base platform?" If it is a separate licence, ask what the pricing is and what is included in the base platform for buyers who do not want the AI add-on.

"Can you give me a reference customer using this AI feature in production who I can speak to?" Product features without production references are roadmap items, not current capabilities.

Tafkiro AI: what it does and what it does not do

Tafkiro AI is woven into the platform rather than offered as a separate add-on. Current capabilities in production: duplicate invoice detection (across vendor, amount, date, reference, and purchase order matching — catching 94% of duplicates in a typical mid-market ledger), bank reconciliation auto-matching (85–94% auto-match rate on typical transaction sets), anomaly flagging in expense and payable transactions, and demand forecasting for businesses with 12+ months of clean transaction history.

Capabilities on the roadmap: predictive cash flow forecasting from receivables aging and payment pattern data, AI-assisted month-end close (automatic draft of recurring journals based on prior period patterns), and natural language report queries.

We do not claim that AI can replace the judgment of a finance professional. We claim that it can remove the mechanical steps that slow that judgment down — the matching, the flagging, the data collection — so that the professional focuses on decisions rather than data entry.