Case B · AI Workflow · Process Optimisation

AI-Driven Menu Data
Automation Pipeline

UK · FintechClaudeChatGPT100 Items / 5 MinReplicable SOP

Merchant menus arrived in chaos — scanned PDFs, WhatsApp screenshots, mixed-language documents. I designed a structured semi-automated AI workflow that compressed 100-item menu processing to 5 minutes, and built a replicable SOP for the whole team.

5min
Per 100 Menu Items
0
Extra Headcount Needed
High
Accuracy Post-Review
SOP
Team-Replicable System
Background

Every menu was a format disaster

NeroPay's merchant customers — mostly Middle Eastern and Turkish restaurants — needed to submit menu data for system import when joining the platform. But these menus arrived in every format imaginable: PDF scans, WhatsApp screenshots, mixed-language Word documents, with inconsistent field names, price formats, and category structures throughout.

The traditional approach was fully manual data entry. As merchant volume increased, that cost would scale linearly. My judgement: this is pattern-based work with fixed logic. AI can intervene.

Workflow

From raw image to system import — the logic behind each step

Input A
Download Sample CSV
ePost Editor
Export the standard field template from NeroPay's internal system. This step defines "what correct output looks like" — without this baseline, AI has nothing to align to.
+
Input B
Merchant Menu Upload
PDF · Image · Screenshot
Merchant-provided raw format — scanned photos, PDF files, WhatsApp screenshots all accepted. No format restrictions — AI handles recognition and conversion.
Both inputs uploaded to Claude simultaneously
Step 01 · AI Processing
Prompt Design
Claude
Instruction to Claude: "Using the sample CSV's field structure as the standard, convert all items in the menu image to the same format." Prompt design is the core — describing the standard clearly matters more than which tool is used.
Step 02 · AI Output
Structured CSV Generated
Claude · Structured Output
Claude identifies item names, prices, and categories from the image, aligns to the sample field structure, and outputs a standardised CSV. 100 items typically complete in seconds.
Output CSV enters human QA stage
Step 03 · Human QA
Manual Spot-Check
Manual Review
Quick scan of output to confirm no field misalignment or recognition errors. AI accuracy is high; this step takes only a few minutes. Role shifts from operator to reviewer.
Step 04 · Final
Direct System Import
ePost Editor · Import
Confirmed CSV imported directly. 100 items live on the platform. Total time: approximately 5 minutes. No manual line-by-line entry required.
Tool Selection Rationale

Not the most popular tool — the most appropriate one

I use multiple AI tools daily, but each has a different strength. These choices were deliberate, not default.

Claude
Structured Data · CSV Processing
Claude performs most stably on structured output, field alignment, and format conversion. It understands "field logic" rather than just "text recognition" — critical for reliable CSV generation.
ChatGPT
Conversational Drafts · Quick Copy
Better suited for rapid natural language generation — marketing copy drafts, customer communication templates. Complementary to Claude, not a replacement.
Higgs · Perplexity
Image Generation · Live Research
Higgs used for marketing visuals — far cheaper than design resources. Perplexity for queries requiring real-time web information.
Standard Operating Procedure

Designed so the process doesn't depend on one person

This pipeline was built by me and currently run by me, but it was designed to be fully handoff-ready — any team member can follow the SOP without re-asking or re-inventing anything.

01

Download sample CSV from ePost Editor

Ensure you have the latest field template version. Save to local working folder.

~1 min
02

Collect merchant's raw menu files

Accepted formats: PDF, image screenshots, WhatsApp photos. If multi-page, combine before uploading.

~2 min
03

Upload both to Claude with standard prompt

Upload sample CSV and menu image simultaneously. Use the established prompt template requesting output in the sample's field structure.

~1 min
04

Download output CSV — manual spot-check

Key checks: item name recognition, price fields, category accuracy. Any anomalies: fix manually for that item.

~2–3 min
05

Import into ePost Editor — confirm live

Use system import function to load CSV directly. Confirm item count matches expected. Done.

~1 min
Outcome

From manual entry to five-minute completion

This pipeline reduced the marginal cost of merchant menu onboarding to near zero. As merchant volume grows, the system scales without adding headcount. The entire SOP is designed for full handoff — my role shifted from executor to process designer.

5 min
100 Items Processing Time
High
Accuracy After QA
0
Extra Headcount Required

My Rationale

Choosing Claude over other tools was a deliberate decision based on the nature of the problem. This is a "structural conversion" task, not a "text generation" task. Claude is more stable on field-logic alignment and formatted output.

"AI's value isn't replacing thinking — it's replacing repetition. Menu processing has fixed logic. Once you can clearly describe 'what is the standard format,' AI can execute it reliably."

Designing the prompt is essentially writing rules, not just using a tool. That framing — starting from the problem structure rather than from "what can this tool do" — is how I approach AI integration across every workflow.

← Case A — Fraud & Legal Ops
Next Case
Case C — HR Systems from Zero →