Document processing
Sheet 10/12 · Medium · 3–10 weeksInvoices, delivery notes, IDs and forms arrive as photos and are typed into systems by hand.
- Supplier invoices, delivery notes and claim forms arrive as WhatsApp photos, emails and paper — and are retyped.
- KYC packs (ID, proof of address, company registration) are checked by eye and filed by hand.
- The same document is handled by three people before it reaches the system.
- Errors surface weeks later as wrong payments, wrong stock or a rejected claim.
- Nobody can find the original when an auditor or ZIMRA asks.
- Classify incoming documents by type and route them (invoice → accounts, delivery note → stores, ID → onboarding).
- Extract the fields you name — supplier, date, amount, currency, TIN, ID number — into your system with a confidence per field.
- Check documents against rules (does the TIN match the supplier record; is the ID expiry in the future; do the totals add up).
- Keep the original image linked to every record for audit.
- Read what the camera did not capture; blurry, cropped or shadowed photos set the ceiling. Capture guidance is part of the project.
- Issue fiscal invoices or replace your fiscal device — it reads and prepares, the device issues.
- Reach 100%; the question is which fields, at what confidence, and who checks the rest.
- Handle a new document layout it has never seen without a person the first few times.
| Effort band | Medium · 3–10 weeks |
|---|---|
| Data you need | 200+ real documents per type; the fields you need; the target system's import format |
| Integration points | Email/WhatsApp intake; accounting/inventory/KYC systems; document storage |
| People | A reviewer for the exception queue; a data owner who signs off accuracy |
| Measure | Field accuracy; % straight-through; minutes per document; errors found downstream |
| Zimbabwe constraint | Phone-camera capture quality; fiscal-invoice handling under FDMS; personal data in KYC packs; storage location |
| Code | Shape | Fit | Why |
|---|---|---|---|
| S-02 | Pilot | Primary | Two hundred of your real documents through extraction; measure field-level accuracy. Sign nothing before this number. |
| S-04 | Integration | Primary | The extracted fields must land in accounting, inventory or the KYC system — that connection is the delivery. |
| S-03 | Implementation | Secondary | Classification rules, validation checks, the review queue and exception handling built for your document mix. |
| Tool | Listed platform with invoice and expense capture | Secondary | Small traders with one document type may be served by a subscription. |
| S-06 | Managed support | Later | Supplier layouts change; accuracy drifts; someone retrains and reports. |
- On 200 of our own documents — including WhatsApp photos — what is the field-level accuracy per field, and how did you measure it?
- What confidence threshold routes a document to a person, and what does the review queue look like?
- Where are our documents processed and stored, in which country, and can we keep originals in Zimbabwe?
- Which validation rules run (TIN format, totals, duplicates), and can we add our own?
- How does the system learn a new supplier's layout, and does that cost extra?
- How does extracted invoice data reach our fiscal process without bypassing the fiscal device?
The full bank of 60 questions, grouped by risk, is at /vendor-questions/.
The accuracy arithmetic
Suppose a pilot on 200 supplier invoices gives 96% field accuracy on amount and 88% on supplier name, and each invoice has six fields you need. Per document, the chance every field is right is roughly the product of the field accuracies — well under 96%. At 1,000 invoices a month that means a few hundred documents with at least one wrong field. That is not a failure; it is the specification for the review queue. Ask the vendor to show straight-through rate (documents needing no human touch) as well as field accuracy, and to route anything below the threshold to a person with the image alongside.
Capture guidance is half the accuracy
Most errors on Zimbabwean document flows come from the photo, not the model: a supplier invoice photographed at an angle on a truck dashboard, a delivery note with a thumb over the total. A one-page capture guide for field staff (flat, full page, daylight, one document per photo) and an intake bot that rejects unreadable images at source raise accuracy more than any model change. Include it in the scope.
Fiscal documents
Where the document is a fiscal tax invoice you received, extraction is straightforward and the QR code can be validated against ZIMRA’s FDMS portal. Where the document is an invoice you issue, the AI’s job ends at preparing buyer details and lines for the fiscal device — see the finance sheet (P-06) and the fiscalisation guide.