Ten decisions African businesses are overpaying a frontier model for

Greg Isenberg's advice for System One models is blunt: find an expensive queue and put the model at the front of it. Here are ten queues we see in Nigerian and Ghanaian businesses, with the typed questions we'd ask and what plain code does with the answers.

1. The WhatsApp order line

• Choice: order, booking, price question, complaint, or a person.
• Score: urgency.

Code checks stock and the calendar, then sends a payment link. The concierge never quotes a price that isn't on the list.

2. Support tickets at a fintech

• Choice: intent.
• Score: complexity and frustration.
• Noul: refund requested?

Order status goes to a database lookup with no model at all. Refunds go to a person.

3. AML alerts

• Score: risk.
• Choice: close or escalate.
• Noul: needs an STR?

Code computes amounts and velocities. An analyst approves every filing.

4. KYC files

• Choice: document type.
• Noul: legible, and signs of obvious tampering?

Name and date comparisons happen in code, never in a model.

5. Inbound leads

• Choice: which offer fits.
• Score: urgency and budget.
• Noul: decision maker, and spam?

Code sets the priority: book a call, send the form, nurture, or discard. (This is literally our intake form.)

6. Grant applications

• Score: one per rubric criterion.
• Noul: eligible, unsupported claims, and ready to submit?

Code weights the rubric and names the three weakest sections.

7. Content before it ships

• Score: hook, specificity and voice.
• Noul: unsourced numbers?

Unsourced numbers block publishing.

8. Study-group placement

• Choice: track and goal.
• Score: level and hours per week.
• Noul: wants live sessions?

Code assigns the pod.

9. Approvals inside agents

• Choice: the effect of this tool call (read, draft, write, send, destroy, move money).

Reads must be proven; everything else pauses for a person. The only dangerous mistake is calling a write a read, so the gate is deliberately asymmetric.

10. Agent payments

• Noul: does this payment match what the user asked for, and is this endpoint what it claims to be?

Budgets and allowlists are enforced in code first. A model can block a payment; it can never raise a limit.

The pattern

Look at the list again. In every case:

• The model answers typed questions. It never writes the reply, moves the money or files the report.
• Code does the exact parts: stock, dates, amounts, limits.
• A person owns anything irreversible.

That's a service business that runs like software.

Start with one

Pick the queue that costs you the most human hours. Write five typed questions about one item in it. Run them on 50 old examples and compare against what your team actually decided. If the answers agree, move that decision. Then do the next one.

If you'd rather we do it, the Agent Readiness Audit maps every decision in your business this way, and the first five are free.