Proprietary AI.
Deterministic math.
You always know which is which.

Two systems. One product. One clear line between them.

SpendCraft runs two kinds of logic in parallel. Each has a specific job.

Proprietary AI is used for:

• Vendor and item normalization: recognizing aliases, clustering near-duplicates
• Classification: mapping transactions into your taxonomy
• Narrative generation: composing plain-language summaries in Ask Crafter and scan outputs
• Pattern recognition: identifying structural signals in tail spend and supplier behavior

Deterministic computation is used for:

• Savings figures and Min–Max ranges
• Confidence scores
• Transaction counts, supplier counts, spend shares
• Benchmark comparisons
• Concentration ratios and variance calculations

The AI narrates. The math counts. The content-type visual contract makes the difference visible on every screen: AI-generated narrative and deterministic data are treated differently in the product, by design.

People review. People refine. People decide.

Every AI output in SpendCraft is reviewable. Every classification can be changed. Every suggested rule can be accepted, edited, or rejected. Classifier Agents run the work; your team sets the standard.

In practice:

• Classification results are inspectable at the transaction level.
• Normalization clusters can be merged, split, or corrected.
• Rules layered on top of AI classification are authored by your team.
• Taxonomies can be switched, extended, or rebuilt without re-ingesting data.

The product is built so business users can refine the output without waiting on data science.

Every signal traces back to the data that produced it.

Savings Opportunity Scan findings cite the transactions, vendor patterns, and pricing variances behind them. Tail Spend Scan opportunities carry their supporting facts with them. Ask Crafter responses show the dataset, period, and taxonomy the answer was drawn from. Scout records keep the baseline evidence attached as the opportunity is worked.

If the data changes, the signal changes. If the taxonomy changes, the classification changes. Nothing in SpendCraft is a black box that the user has to trust on faith.

Ranges, not point estimates. Directional, not guaranteed.

Every savings figure in SpendCraft is displayed as a Min–Max range. Coverage percentages appear where relevant: the share of transactions the finding is based on. “Directional, not guaranteed” is standard language on scan outputs.

This is a deliberate posture. Procurement and finance teams have been burned by inflated savings claims for decades. SpendCraft's job is to surface only what the data supports.

Your data stays yours.

Raw transaction data sent to SpendCraft is used to run classification and analysis for your workspace. It is not used to train SpendCraft's AI models. It is not shared across tenants. Tenant isolation is architectural, not a policy overlay. See Security & Data Handling for the full posture.

How the AI behaves today is how it behaves tomorrow, unless we tell you otherwise.

SpendCraft's AI models are versioned. Material changes to classification logic, normalization behavior, or narrative generation are communicated in release notes. For enterprise customers, change windows and review periods can be scoped into the deployment.

This is the posture that lets procurement and finance teams certify SpendCraft outputs for audit, board reporting, and executive decisions. The AI doesn't shift underneath the data.