How Quantixr produces a Business Risk Score — and what it deliberately won't do.

Every assessment runs through the same deterministic pipeline. A reviewer submits structured business information and a bank statement. Quantixr extracts and calculates financial metrics inside the application — never via a language model — and then combines those metrics with location, operational, and digital signals into four scored dimensions: Financial Health, Location Context, Business Stability, and Digital Footprint.

The four dimensions are weighted and summed into a single Business Risk Score between 0 and 100, mapped to a tier band. The same inputs always produce the same score. Reviewers see the component breakdown and the underlying figures, so the result can be inspected and challenged rather than taken on trust.

The score is built from structured, auditable inputs in four categories:

• Trading data — the business's own bank statement, parsed deterministically for revenue, expenses, cash flow, and recurring obligations.
• Official statistics — Statistics South Africa (Stats SA) economic and demographic datasets at suburb and municipality level. NCR and SARS statistics on industry sector.
• Reported crime data — South African Police Service (SAPS) crime statistics published at police-station and suburb level.
• Municipal & open data — City of Cape Town and comparable open-data portals covering economic activity and business density.
• Digital signals — verified Google Business listings, the business's website, and its public social profiles.

Every assessment produces one overall score (0–100) from four weighted components. Sub-factor weights below match the scoring engine exactly.

Public datasets in South Africa are not updated in real time. Stats SA releases are periodic, SAPS crime data is reported in fixed cycles, and municipal open-data feeds vary in freshness. A score therefore reflects the most recent reliably-published figures, not yesterday's reality.

Where an exact suburb match is not available, Quantixr falls back to a city-level average for that sub-factor and labels it as approximate in the reviewer's report; if the address cannot be resolved at all, Location Context is dropped from the score entirely rather than guessed, and the remaining weightings are renormalised. Financial Health is treated differently: it is never estimated or reweighted, so a bank statement that cannot be parsed or reconciled blocks the assessment instead of producing a score. We would rather show less than show something we cannot defend.

The score is a decision-support signal, not a verdict. It compresses imperfect information into a single number and should be read alongside the underlying figures, not in place of them.

South African data carries the imprint of the country's history. Reported crime statistics are influenced by where policing is concentrated and where incidents are reported in the first place. Economic activity indices reflect formal economic measurement and systematically under-represent informal trade, which is the lived reality for a large share of SMEs.

Suburbs that have historically been under-served by infrastructure, banking, and municipal investment will look "thinner" in the data than they are in practice. We are explicit about this: Location Context describes the measured environment around a business; it is not a judgement about the business itself or the people who operate it.

We deliberately cap the influence of any single environmental signal, never let location dominate the score, and surface the underlying components separately so reviewers can apply their own judgement rather than defer entirely to the headline number.

Artificial intelligence is used in one narrow place: writing the plain-English narrative summary of an already-calculated profile, and a one-line explanatory note for individual sub-factor scores.

The AI never sees raw bank statement text and it never performs arithmetic or decides the score. It receives only the outputs the deterministic engine has already produced — the overall score and tier, the four component scores, and every sub-factor score by name — together with the business name, industry, years trading, and the trading period observed on the statement.

The risk flags and the improvement recommendations shown on the report are not written by the AI. They are generated by a deterministic, threshold-based rule set that runs after scoring, so the same scores always produce the same flags and the same recommendations. The strongest flag and recommendation are passed into the summary prompt only so the narrative stays consistent with what is listed beneath it.

The summary itself is constrained: three to five sentences, opening with the business, naming the tier and overall score, referencing the strongest positive and the strongest negative signal, and avoiding lending language entirely. Where the bank statement failed to reconcile, the summary must lead with that and may not describe revenue or cash flow favourably. Any output that breaks these rules is rejected and replaced with a deterministic fallback summary. The numbers on the report are not generated by a model — they are calculated, then described.

• Quantixr does not make decisions for or against any business. It produces a risk signal; decisions remain with the reviewing organisation.
• Quantixr is not a registered financial services provider and does not issue affordability rulings or scoring products regulated under the NCA.
• Quantixr does not interact with the small business being assessed. SMEs do not log in, and no marketing or outreach is sent to them.
• Quantixr does not use AI to compute, adjust, or override financial figures, and does not generate narrative claims that go beyond the structured inputs.
• Quantixr does not sell, share, or repurpose submitted business data for any use outside the reviewer's own assessment workflow.