Earlier this week, a supplier invoice came in priced 14% above the last purchase order — the difference would have cleared payment unnoticed. It was caught at 6:42 on Tuesday morning, before anyone was at their desk, by a quiet layer sitting over the ledger.

An attentive reader.

A patient editor.

A second pair of eyes.

A quiet hand in the margins.

AI/ eɪ aɪ /noun

An exception layer that reads what the platform records, then quietly marks what would otherwise pass unnoticed.

From

the recognition that people see best what they choose to look at — and that running a business asks them to look at everything.

Four verbs. Each one quieter than the last.

Reads.

Every transaction, every message, every line of the ledger — without taking a coffee break, without losing focus on Friday afternoon.

Invoices42
Bank feed18
Payroll6
Tickets27
Inventory91

Surfaces.

What looks unlike the pattern. What duplicates. What sits too long. What costs more than it should. Lifts the few that matter out of the many that don’t.

PO-2418 · Bidvest
PO-2419 · Cipla SA
PO-2420 · Sasol+14%
PO-2421 · Macsteel

Drafts.

The chase email. The variance note. The supplier follow-up. Composed in your voice, queued for review, never sent on its own.

DRAFT · awaiting review

Dear Mr Daniels, our records show invoice INV-1042 is

18 days overdue. As discussed last quarter, we’d appreciate confirmation of the payment date…

Models.

What happens if the largest account pays thirty days late. What the next two months of cash look like. What hiring two more would do.

Three things, before the day begins.

A real morning brief, from a fictional but representative Friday. Each item arrives with proposed wording — nothing is acted on without you.

Morning brieffri · oct 11
3 flagged · 112 read

06:14

INV-2417 · Sasol

Supplier price up 14% vs last purchase order. Same SKU, same delivery point. Last six months show a steady RPI-tracking range — this sits outside it.

draft sent to procurement queue · pending sign-off

06:21

AR-Aging · Daniels & Co

Largest customer is now 32 days overdue, the longest since the relationship began in 2022. Two emails sent, no reply. Pattern unusual.

draft chase prepared in your tone · review and send

06:28

JE-1842 · Office rent

Posted to Travel & Entertainment rather than Occupancy. Amount and counterparty match the recurring lease pattern. Likely a mis-code.

reclassification suggested · one click to apply

Sources, listed plainly.

Nothing outside the platform’s perimeter. Nothing inferred from elsewhere. Every observation traces back to one of these.

01

General ledger

every journal entry, posted and unposted, with their counterparties.

02

Bank feed

all daily movements across operating, savings and project accounts.

03

Sales invoices

issued, paid, partially paid, and ageing buckets.

04

Supplier bills

matched, unmatched, duplicate risk, price drift.

05

Payroll register

current and prior runs, variance against expectation.

06

Stock movement

receipts, issues, levels, reorder triggers.

07

Helpdesk tickets

open, ageing, SLA position, sentiment cues.

08

Documented rules

your covenants, approval thresholds, retention rules.

Says what it knows. Says what it doesn’t.

Every observation arrives with a position on this line. A match is not the same as a guess. A guess is not the same as a hunch.

Bank line matches invoice INV-1042

Bill is a duplicate of BILL-887

Supplier price is unusually high

Customer is likely to delay payment

Q4 cash cover holds at current pace

CertainLikelyUncertain

A claim near the left is a match, a duplicate, an arithmetic certainty. A claim near the right is a possibility — useful, but never grounds for action without you.

Four things it will not do.

01

Doesn’t send anything you haven’t seen.

Drafts queue; humans send. The signature on the letter is yours.

02

Doesn’t move money.

Payments are released by people with banking authority. Always.

03

Doesn’t decide who’s at fault.

It surfaces the variance; you decide what it means.

04

Doesn’t pretend to know.

Where the data is thin, it says so. Silence is allowed.

AI at each tier of the platform.

The same platform, the same reader. What changes is how much it’s asked to do.

01Workflow Fix

A weekly digest of duplicate-bill candidates and ageing items.

R 22 500

once-off

02Essential

Daily exception brief in your inbox by 07:00. Flags, drafts queued for review.

R 12 000

per month

03Growth
Most chosen

Daily brief + scenario modelling on demand. Cash, AR, hiring. Pre-built and bespoke.

R 22 000

per month

04Full

All of the above + covenant compliance reviews, AI-drafted board notes, anomaly tuning.

R 38 000

per month

What people ask before they trust it.

Every weekday morning, AI reads the previous day’s transactions via the platform’s API, compares them to historical patterns and the client’s documented business rules, and produces a partner brief. Anomalies are flagged with proposed actions. The partner starts the day knowing what needs attention.

Anomalies (unusual transactions, duplicate bills, mis-coded entries), watch items (concentration risks, ageing patterns), recommendations (proposed reclassifications, supplier follow-ups), and scenarios on demand. The brief is structured so the partner can act in 15 minutes rather than reconstruct context for two hours.

No. AI proposes; humans approve. The bookkeeper or partner reviews every AI-flagged item before it goes into the ledger or out to a customer. The only AI outputs that ship without a human sign-off are technical match suggestions inside automation workflows (e.g. matching a payment to an invoice) — and even those have audit trail.

Not without human review. AI-drafted dunning letters, contract renewal notes and similar communications are queued for partner sign-off before being sent. WhatsApp confirmations on invoice / payment events run on deterministic templates, not generative AI.

Common scenarios pre-built: revenue drop / lift, AR delays, AP acceleration, major capex investment, hiring waves, financing rounds, customer concentration shifts. Bespoke scenarios can be built per client on Full.

AI runs covenant compliance checks after each month-end, flags POPIA retention obligations, surfaces likely VAT-classification issues before submission, and reviews payroll outliers before the run is committed. The partner reviews everything before it’s lodged.

All AI processing happens inside the platform’s perimeter. Anthropic Claude is accessed via API for inference but never retains your data for training. POPIA consent is enforced at the data layer. Audit trail captures every AI decision and every human override.

In closing

The first morning is the strangest.

After that, you wonder how you ran a business without it.