Accuracy & audit

Prove every number. Audit every decision.

Matching in Brisc is deterministic, rule-based and explainable; AI reads the documents and argues the borderline cases, never the match itself. Every record carries what the AI proposed, what a human chose, and the evidence for both.

REC-2087 GBP 18,240.00
AI proposed Marmalade Broking · alias “MRLD BRKNG LDN” · 94%
Human chose Marmalade Broking · confirmed
Reviewer J.O. · 12 May, 14:32
bank statement broker statement p.3 June BDX line 18
Deterministic matching

Is the matching AI or rules?

Rules. The arithmetic never belongs to a model, which is why the numbers are never a guess.

Rules decide

The same inputs produce the same answer every run, and every match names the rule that made it.

AI reads and argues

The model extracts from messy documents and argues borderline matches; it proposes, never decides.

Uncertainty queues

Confidence tiers route anything unclear to a person, reasoning and source documents attached.

The honest posture

How accurate is it?

Two numbers, deliberately kept apart: what posts is 97%+ accurate, because a human clears everything uncertain first. The match rate underneath is a ramp, not a promise.

80% 92–95%

The honest ramp. 80% of records match on day one; 92 to 95% by about 90 days. At steady state, the review queue is roughly one row in twenty.

Evidence

Every value names its source.

Not a confidence score and a shrug. Every extracted value is stamped with the document it came from and a written reason it was read that way, one click from the source file.

Source stamping, per value. Each field points to the exact document, page and line behind it.
Reasoning, written down. Why this value, why this match, in words a reviewer or an auditor can read.
The rule or alias named. Which rule fired, which confirmed alias resolved the payer, what tolerance applied.
Exportable in full. Year-end, client-money review, internal audit: hand over the trail, not a reconstruction.
Human-in-the-loop

Who is accountable for what the AI does?

A named person, on every record. The loop is audit-grade, not a checkbox.

Lineage on every record

What the AI proposed and what a human chose, stored side by side, inspectable forever.

Reviewer identity

Who approved it, when, and what they saw, with a timestamp.

Per-record exception queues

Exceptions carry their own reasoning and sources, worked record by record, never a bulk write-off.

Workflow states

Draft, reviewed, approved, export: nothing leaves while it is still an opinion.

The fail-open principle

When unsure, it stops and asks.

Nothing is guessed, nothing is silently corrected. A record the engine cannot settle gets a place in the queue, and autonomy is earned in stages your team controls, the way a new hire earns it.

Stage one. Every match waits for approval. Your team reviews all of it.
Stage two. Routine volume clears; exceptions and new patterns still queue.
Stage three. Finished work posts to your systems. Exceptions still route to a person.
Model posture

Does your data train a model?

No. That is a design decision, not a settings toggle.

Provider-agnostic catalog

Models from multiple providers, the right one per reading job, no single-vendor lock.

Pinned versions

Model versions are pinned per tenant; the model reading your bordereaux changes only when you change it.

Your data stays yours

Your data never trains shared models; what it learns from your corrections stays in your tenant. In full: security and tenancy.

This is the trust story under every Analyst: reconciliation, bordereaux, submissions, claims, cessions. The engine they share: the platform · integrations.

Common questions

Questions auditors ask us

Is the matching AI or rules?

Rules. Matching is deterministic, rule-based and explainable: the same inputs produce the same answer every run, and every match names the rule that made it. AI reads the documents and argues the borderline cases; it never decides a match.

How accurate is it, honestly?

Output accuracy is 97%+, because a human reviews everything the engine isn’t sure about before it posts. The match rate itself is 80% on day one and climbs to 92 to 95% by about 90 days, as the rule base absorbs what the model surfaces. At steady state your team reviews roughly one row in twenty.

What happens when it cannot match something?

It fails open: the record stops and asks. Unmatched rows queue per record with the reasoning and source documents attached. Nothing is guessed, nothing is silently corrected.

What does the audit trail actually contain?

On every record: what the AI proposed, what a human chose, the reviewer’s identity, the timestamp, the rule or alias that fired, and a click-through to the exact source document behind each value. The full trail is exportable.

Does our data train your models?

Your data never trains shared models. What the platform learns from your corrections, rules, prompt tuning, confirmed aliases, stays inside your tenant and belongs to your book. Brisc runs a provider-agnostic model catalog with pinned model versions per tenant, so the model reading your documents doesn’t change under you and nothing you send us leaks into anyone else’s system.

Bring your own file. Audit the run.

30 minutes, no slide deck. We run your data and you inspect the lineage on every record: what the AI proposed, what a human chose, and the evidence behind both.

Book the walkthrough